Authors: Rongzheng Wang, Yihong Huang, Muquan Li, Jiakai Li, Di Liang, Bob Simons, Pei Ke, Shuang Liang, Ke Qin
Abstract: Large Language Models (LLMs) have advanced the field of Combinatorial Optimization through automated heuristic generation. Instead of relying on manual design, this LLM-Driven Heuristic Design (LHD) process leverages LLMs to iteratively generate and refine solvers to achieve high performance. However, existing LHD frameworks face two critical limitations: (1) Endpoint-only evaluation, which ranks solvers solely by final quality, ignoring the convergence process and runtime efficiency; (2) High adaptation costs, where distribution shifts necessitate re-adaptation to generate specialized solvers for new instance groups. To address these issues, we propose Dynamics-Aware Solver Heuristics (DASH), a framework that co-optimizes solver search mechanisms and runtime schedules guided by a convergence-aware metric, thereby identifying efficient and high-performance solvers. Furthermore, to mitigate expensive re-adaptation, DASH incorporates Profiled Library Retrieval (PLR). PLR efficiently archives specialized solvers concurrently with the evolutionary process, enabling cost-effective warm-starts for heterogeneous distributions. Experiments on four combinatorial optimization problems demonstrate that DASH improves runtime efficiency by over 3 times, while surpassing the solution quality of state-of-the-art baselines across diverse problem scales. Furthermore, by enabling profile-based warm starts, DASH maintains superior accuracy under different distributions while cutting LLM adaptation costs by over 90%.
Authors: Atik Faysal, Mohammad Rostami, Reihaneh Gh. Roshan, Nikhil Muralidhar, Huaxia Wang
Abstract: Multimodal pretraining is effective for building general-purpose representations, but in many practical deployments, only one modality is heavily used during downstream fine-tuning. Standard pretraining strategies treat all modalities uniformly, which can lead to under-optimized representations for the modality that actually matters. We propose Finetune-Informed Pretraining (FIP), a model-agnostic method that biases representation learning toward a designated target modality needed at fine-tuning time. FIP combines higher masking difficulty, stronger loss weighting, and increased decoder capacity for the target modality, without modifying the shared encoder or requiring additional supervision. When applied to masked modeling on constellation diagrams for wireless signals, FIP consistently improves downstream fine-tuned performance with no extra data or compute. FIP is simple to implement, architecture-compatible, and broadly applicable across multimodal masked modeling pipelines.
Authors: Xiyuan Liu, Christian Hacker, Shengnian Wang, Yuhua Duan
Abstract: Developing new metal hydrides is a critical step toward efficient hydrogen storage in carbon-neutral energy systems. However, existing materials databases, such as the Materials Project, contain a limited number of well-characterized hydrides, which constrains the discovery of optimal candidates. This work presents a framework that integrates causal discovery with a lightweight generative machine learning model to generate novel metal hydride candidates that may not exist in current databases. Using a dataset of 450 samples (270 training, 90 validation, and 90 testing), the model generates 1,000 candidates. After ranking and filtering, six previously unreported chemical formulas and crystal structures are identified, four of which are validated by density functional theory simulations and show strong potential for future experimental investigation. Overall, the proposed framework provides a scalable and time-efficient approach for expanding hydrogen storage datasets and accelerating materials discovery.
Authors: Jiangyang Li, Chenhao Ding, Songlin Dong, Qiang Wang, Jianchao Zhao, Yuhang He, Yihong Gong
Abstract: Prompt-based continual learning methods effectively mitigate catastrophic forgetting. However, most existing methods assign a fixed set of prompts to each task, completely isolating knowledge across tasks and resulting in suboptimal parameter utilization. To address this, we consider the practical needs of continual learning and propose a prompt-sharing framework. This framework constructs a global prompt pool and introduces a task-aware gated routing mechanism that sparsely activates a subset of prompts to achieve dynamic decoupling and collaborative optimization of task-specific feature representations. Furthermore, we introduce a history-aware modulator that leverages cumulative prompt activation statistics to protect frequently used prompts from excessive updates, thereby mitigating inefficient parameter usage and knowledge forgetting. Extensive analysis and empirical results demonstrate that our approach consistently outperforms existing static allocation strategies in effectiveness and efficiency.
Authors: Luca Pinchetti, Simon Frieder, Thomas Lukasiewicz, Tommaso Salvatori
Abstract: Research aimed at scaling up neuroscience inspired learning algorithms for neural networks is accelerating. Recently, a key research area has been the study of energy-based learning algorithms such as predictive coding, due to their versatility and mathematical grounding. However, the applicability of such methods is held back by the large computational requirements caused by their iterative nature. In this work, we address this problem by showing that the choice of initialization of the neurons in a predictive coding network matters significantly and can notably reduce the required training times. Consequently, we propose a new initialization technique for predictive coding networks that aims to preserve the iterative progress made on previous training samples. Our approach suggests a promising path toward reconciling the disparities between predictive coding and backpropagation in terms of computational efficiency and final performance. In fact, our experiments demonstrate substantial improvements in convergence speed and final test loss in both supervised and unsupervised settings.
Authors: Nikita Makarov, Maria Bordukova, Lena Voith von Voithenberg, Estrella Pivel-Villanueva, Sabrina Mielke, Jonathan Wickes, Hanchen Wang, Mingyu Derek Ma, Keunwoo Choi, Kyunghyun Cho, Stephen Ra, Raul Rodriguez-Esteban, Fabian Schmich, Michael Menden
Abstract: Precision oncology requires forecasting clinical events and trajectories, yet modeling sparse, multi-modal clinical time series remains a critical challenge. We introduce TwinWeaver, an open-source framework that serializes longitudinal patient histories into text, enabling unified event prediction as well as forecasting with large language models, and use it to build Genie Digital Twin (GDT) on 93,054 patients across 20 cancer types. In benchmarks, GDT significantly reduces forecasting error, achieving a median Mean Absolute Scaled Error (MASE) of 0.87 compared to 0.97 for the strongest time-series baseline (p<0.001). Furthermore, GDT improves risk stratification, achieving an average concordance index (C-index) of 0.703 across survival, progression, and therapy switching tasks, surpassing the best baseline of 0.662. GDT also generalizes to out-of-distribution clinical trials, matching trained baselines at zero-shot and surpassing them with fine-tuning, achieving a median MASE of 0.75-0.88 and outperforming the strongest baseline in event prediction with an average C-index of 0.672 versus 0.648. Finally, TwinWeaver enables an interpretable clinical reasoning extension, providing a scalable and transparent foundation for longitudinal clinical modeling.
Authors: Chen Feng, Minghe Shen, Ananth Balashankar, Carsten Gerner-Beuerle, Miguel R. D. Rodrigues
Abstract: Reliable certification of Large Language Models (LLMs)-verifying that failure rates are below a safety threshold-is critical yet challenging. While "LLM-as-a-Judge" offers scalability, judge imperfections, noise, and bias can invalidate statistical guarantees. We introduce a "Noisy but Valid" hypothesis testing framework to address this. By leveraging a small human-labelled calibration set to estimate the judge's True Positive and False Positive Rates (TPR/FPR), we derive a variance-corrected critical threshold applied to a large judge-labelled dataset. Crucially, our framework theoretically guarantees finite-sample Type-I error control (validity) despite calibration uncertainty. This distinguishes our work from Prediction-Powered Inference (PPI), positioning our method as a diagnostic tool that explicitly models judge behavior rather than a black-box estimator. Our contributions include: (1) Theoretical Guarantees: We derive the exact conditions under which noisy testing yields higher statistical power than direct evaluation; (2) Empirical Validation: Experiments on Jigsaw Comment, Hate Speech and SafeRLHF confirm our theory; (3) The Oracle Gap: We reveal a significant performance gap between practical methods and the theoretical "Oracle" (perfectly known judge parameters), quantifying the cost of estimation. Specifically, we provide the first systematic treatment of the imperfect-judge setting, yielding interpretable diagnostics of judge reliability and clarifying how evaluation power depends on judge quality, dataset size, and certification levels. Together, these results sharpen understanding of statistical evaluation with LLM judges, and highlight trade-offs among competing inferential tools.
Authors: Anni Zhao, Ayca Ermis, Jeffrey Robert Vitt, Sergio Brasil, Wellingson Paiva, Magdalena Kasprowicz, Malgorzata Burzynska, Robert Hamilton, Runze Yan, Ofer Sadan, J. Claude Hemphill, Lieven Vandenberghe, Xiao Hu
Abstract: Objective: Accurate noninvasive estimation of intracranial pressure (ICP) remains a major challenge in critical care. We developed a bespoke machine learning algorithm that integrates system identification and ranking-constrained optimization to estimate mean ICP from noninvasive signals. Methods: A machine learning framework was proposed to obtain accurate mean ICP values using arbitrary noninvasive signals. The subspace system identification algorithm is employed to identify cerebral hemodynamics models for ICP simulation using arterial blood pressure (ABP), cerebral blood velocity (CBv), and R-wave to R-wave interval (R-R interval) signals in a comprehensive database. A mapping function to describe the relationship between the features of noninvasive signals and the estimation errors is learned using innovative ranking constraints through convex optimization. Patients across multiple clinical settings were randomly split into testing and training datasets for performance evaluation of the mapping function. Results: The results indicate that about 31.88% of testing entries achieved estimation errors within 2 mmHg and 34.07% of testing entries between 2 mmHg to 6 mmHg from the nonlinear mapping with constraints. Conclusion: Our results demonstrate the feasibility of the proposed noninvasive ICP estimation approach. Significance: Further validation and technical refinement are required before clinical deployment, but this work lays the foundation for safe and broadly accessible ICP monitoring in patients with acute brain injury and related conditions.
Authors: Steve Hanneke, Shay Moran
Abstract: We provide a complete theory of optimal universal rates for binary classification in the agnostic setting. This extends the realizable-case theory of Bousquet, Hanneke, Moran, van Handel, and Yehudayoff (2021) by removing the realizability assumption on the distribution. We identify a fundamental tetrachotomy of optimal rates: for every concept class, the optimal universal rate of convergence of the excess error rate is one of $e^{-n}$, $e^{-o(n)}$, $o(n^{-1/2})$, or arbitrarily slow. We further identify simple combinatorial structures which determine which of these categories any given concept class falls into.
Authors: Ahmed Rashwan, Keith Briggs, Chris Budd, Lisa Kreusser
Abstract: Monotone optimisation problems admit specialised global solvers such as the Polyblock Outer Approximation (POA) algorithm, but these methods typically require explicit objective and constraint functions. In many applications, these functions are only available through data, making POA difficult to apply directly. We introduce an algorithm-aware learning approach that integrates learned models into POA by directly predicting its projection primitive via the radial inverse, avoiding the costly bisection procedure used in standard POA. We propose Homogeneous-Monotone Radial Inverse (HM-RI) networks, structured neural architectures that enforce key monotonicity and homogeneity properties, enabling fast projection estimation. We provide a theoretical characterisation of radial inverse functions and show that, under mild structural conditions, a HM-RI predictor corresponds to the radial inverse of a valid set of monotone constraints. To reduce training overhead, we further develop relaxed monotonicity conditions that remain compatible with POA. Across multiple monotone optimisation benchmarks (indefinite quadratic programming, multiplicative programming, and transmit power optimisation), our approach yields substantial speed-ups in comparison to direct function estimation while maintaining strong solution quality, outperforming baselines that do not exploit monotonic structure.
Authors: Abdullah Akg\"ul, Gulcin Baykal, Manuel Hau{\ss}mann, Mustafa Mert \c{C}elikok, Melih Kandemir
Abstract: Optimal control of complex environments with robotic systems faces two complementary and intertwined challenges: efficient organization of sensory state information and far-sighted action planning. Because the reinforcement learning framework addresses only the latter, it tends to deliver sample-inefficient solutions. Active inference is the state-of-the-art process theory that explains how biological brains handle this dual problem. However, its applications to artificial intelligence have thus far been limited to extensions of existing model-based approaches. We present a formal abstraction of reinforcement learning algorithms that spans model-based, distributional, and model-free approaches. This abstraction seamlessly integrates active inference into the distributional reinforcement learning framework, making its performance advantages accessible without transition dynamics modeling.
Authors: Md Muhtasim Munif Fahim, Md Rezaul Karim
Abstract: A large number of children experience preventable developmental delays each year, yet the deployment of machine learning in new countries has been stymied by a data bottleneck: reliable models require thousands of samples, while new programs begin with fewer than 100. We introduce the first pre-trained encoder for global child development, trained on 357,709 children across 44 countries using UNICEF survey data. With only 50 training samples, the pre-trained encoder achieves an average AUC of 0.65 (95% CI: 0.56-0.72), outperforming cold-start gradient boosting at 0.61 by 8-12% across regions. At N=500, the encoder achieves an AUC of 0.73. Zero-shot deployment to unseen countries achieves AUCs up to 0.84. We apply a transfer learning bound to explain why pre-training diversity enables few-shot generalization. These results establish that pre-trained encoders can transform the feasibility of ML for SDG 4.2.1 monitoring in resource-constrained settings.
Authors: Lutz Oettershagen
Abstract: Identifying the top-$k$ items is fundamental but often prohibitive when exact valuations are expensive. We study a two-oracle setting with a fast, noisy weak oracle and a scarce, high-fidelity strong oracle (e.g., human expert verification or expensive simulation). We first analyze a simple screen-then-certify baseline (STC) and prove it makes at most $m(4\varepsilon_{\max})$ strong calls given jointly valid weak confidence intervals with maximum radius $\varepsilon_{\max}$, where $m(\cdot)$ denotes the near-tie mass around the top-$k$ threshold. We establish a conditional lower bound of $\Omega(m(\varepsilon_{\max}))$ for any algorithm given the same weak uncertainty. Our main contribution is ACE, an adaptive certification algorithm that focuses strong queries on critical boundary items, achieving the same $O(m(4\varepsilon_{\max}))$ bound while reducing strong calls in practice. We then introduce ACE-W, a fully adaptive two-phase method that allocates weak budget adaptively before running ACE, further reducing strong costs.
Authors: Md Muhtasim Munif Fahim, Md Rezaul Karim
Abstract: Neural scaling laws describe how language model loss decreases with parameters and data, but treat architecture as interchangeable--a billion parameters could arise from a shallow-wide model (10 layers & 8,192 hidden dimension) or a deep-narrow one (80 layers & 2,048 hidden dimension). We propose architecture-conditioned scaling laws decomposing this dependence, finding that optimal depth scales as D* ~ C^0.12 while optimal width scales as W* ~ C^0.34, meaning width should grow 2.8x faster than depth. We discover a critical depth phenomenon: beyond D_crit ~ W^0.44 (sublinear in W), adding layers increases loss despite adding parameters--the Depth Delusion. Empirically, we validate these findings across 30 transformer architectures spanning 17M to 7B parameters, each trained on representative high-compute samples, achieving R^2 = 0.922. Our central finding: at 7B scale, a 64-layer model (6.38B params) underperforms a 32-layer model (6.86B params) by 0.12 nats, despite being significantly deeper. This demonstrates that optimal depth-width tradeoffs persist at the production scale.
Authors: Shreshth A Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J. Roberts, Aron Walsh, Yarin Gal
Abstract: Existing benchmarks for computational materials discovery primarily evaluate static predictive tasks or isolated computational sub-tasks. While valuable, these evaluations neglect the inherently iterative and adaptive nature of scientific discovery. We introduce MAterials Discovery Environments (MADE), a novel framework for benchmarking end-to-end autonomous materials discovery pipelines. MADE simulates closed-loop discovery campaigns in which an agent or algorithm proposes, evaluates, and refines candidate materials under a constrained oracle budget, capturing the sequential and resource-limited nature of real discovery workflows. We formalize discovery as a search for thermodynamically stable compounds relative to a given convex hull, and evaluate efficacy and efficiency via comparison to baseline algorithms. The framework is flexible; users can compose discovery agents from interchangeable components such as generative models, filters, and planners, enabling the study of arbitrary workflows ranging from fixed pipelines to fully agentic systems with tool use and adaptive decision making. We demonstrate this by conducting systematic experiments across a family of systems, enabling ablation of components in discovery pipelines, and comparison of how methods scale with system complexity.
Authors: Ruicheng Ao, David Simchi-Levi, Xinshang Wang
Abstract: Operations Research practitioners routinely debug infeasible models through an iterative process: analyzing Irreducible Infeasible Subsystems (\IIS{}), identifying constraint conflicts, and systematically repairing formulations until feasibility is achieved. Yet existing LLM benchmarks evaluate OR as one-shot translation -- given a problem description, generate solver code -- ignoring this diagnostic loop entirely. We introduce two benchmarks that place the \textbf{solver in the evaluation loop}. \textbf{\ORDebug{}} evaluates iterative self-correction through 5,000+ problems spanning 9 error types; each repair action triggers solver re-execution and \IIS{} recomputation, providing deterministic, verifiable feedback. \textbf{\ORBias{}} evaluates behavioral rationality through 2,000 newsvendor instances (1,000 ID + 1,000 OOD), measuring systematic deviations from closed-form optimal policies. Across 26 models and 12,000+ samples, we find that domain-specific RLVR training enables an 8B model to surpass frontier APIs: 95.3\% vs 86.2\% recovery rate (+9.1\%), 62.4\% vs 47.8\% diagnostic accuracy (+14.6\%), and 2.25 vs 3.78 steps to resolution (1.7$\times$ faster). On \ORBias{}, curriculum training achieves the only negative ID$\rightarrow$OOD bias drift among models evaluated (-9.6\%), reducing systematic bias by 48\% (from 20.0\% to 10.4\%). These results demonstrate that process-level evaluation with verifiable oracles enables targeted training that outperforms scale.
Authors: Young Kyung Kim, Oded Schlesinger, Qiangqiang Wu, J. Mat\'ias Di Martino, Guillermo Sapiro
Abstract: Test-Time Adaptation (TTA) enables pre-trained models to adjust to distribution shift by learning from unlabeled test-time streams. However, existing methods typically treat these streams as independent samples, overlooking the supervisory signal inherent in temporal dynamics. To address this, we introduce Order-Aware Test-Time Adaptation (OATTA). We formulate test-time adaptation as a gradient-free recursive Bayesian estimation task, using a learned dynamic transition matrix as a temporal prior to refine the base model's predictions. To ensure safety in weakly structured streams, we introduce a likelihood-ratio gate (LLR) that reverts to the base predictor when temporal evidence is absent. OATTA is a lightweight, model-agnostic module that incurs negligible computational overhead. Extensive experiments across image classification, wearable and physiological signal analysis, and language sentiment analysis demonstrate its universality; OATTA consistently boosts established baselines, improving accuracy by up to 6.35%. Our findings establish that modeling temporal dynamics provides a critical, orthogonal signal beyond standard order-agnostic TTA approaches.
Authors: Jos\'e Afonso, Pedro Viegas, Rodrigo Ventura, Vasco Guerra
Abstract: Surrogate modeling for complex physical systems typically faces a trade-off between data-fitting accuracy and physical consistency. Physics-consistent approaches typically treat physical laws as soft constraints within the loss function, a strategy that frequently fails to guarantee strict adherence to the governing equations, or rely on post-processing corrections that do not intrinsically learn the underlying solution geometry. To address these limitations, we introduce the {Conditional Denoising Model (CDM)}, a generative model designed to learn the geometry of the physical manifold itself. By training the network to restore clean states from noisy ones, the model learns a vector field that points continuously towards the valid solution subspace. We introduce a time-independent formulation that transforms inference into a deterministic fixed-point iteration, effectively projecting noisy approximations onto the equilibrium manifold. Validated on a low-temperature plasma physics and chemistry benchmark, the CDM achieves higher parameter and data efficiency than physics-consistent baselines. Crucially, we demonstrate that the denoising objective acts as a powerful implicit regularizer: despite never seeing the governing equations during training, the model adheres to physical constraints more strictly than baselines trained with explicit physics losses.
Authors: Zongheng Guo, Tao Chen, Yang Jiao, Yi Pan, Xiao Hu, Manuela Ferrario
Abstract: Current foundation model for photoplethysmography (PPG) signals is challenged by the intrinsic redundancy and noise of the signal. Standard masked modeling often yields trivial solutions while contrastive methods lack morphological precision. To address these limitations, we propose a Statistical-prior Informed Generative Masking Architecture (SIGMA-PPG), a generative foundation model featuring a Prior-Guided Adversarial Masking mechanism, where a reinforcement learning-driven teacher leverages statistical priors to create challenging learning paths that prevent overfitting to noise. We also incorporate a semantic consistency constraint via vector quantization to ensure that physiologically identical waveforms (even those altered by recording artifacts or minor perturbations) map to shared indices. This enhances codebook semantic density and eliminates redundant feature structures. Pre-trained on over 120,000 hours of data, SIGMA-PPG achieves superior average performance compared to five state-of-the-art baselines across 12 diverse downstream tasks. The code is available at https://github.com/ZonghengGuo/SigmaPPG.
Authors: Omer Rochman-Sharabi, Gilles Louppe
Abstract: Neural emulators based on diffusion models show promise for scientific applications, but vanilla models cannot guarantee physical accuracy or constraint satisfaction. We address this by introducing a constrained sampling framework that enforces hard constraints, such as physical laws or observational consistency, at generation time. Our approach defines a constrained forward process that diffuses only over the feasible set of constraint-satisfying samples, inducing constrained marginal distributions. To reverse this, we propose Predict-Project-Renoise (PPR), an iterative algorithm that samples from the constrained marginals by alternating between denoising predictions, projecting onto the feasible set, and renoising. Experiments on 2D distributions, PDEs, and global weather forecasting demonstrate that PPR reduces constraint violations by over an order of magnitude while improving sample consistency and better matching the true constrained distribution compared to baselines.
Authors: Chengzu Li, Zanyi Wang, Jiaang Li, Yi Xu, Han Zhou, Huanyu Zhang, Ruichuan An, Dengyang Jiang, Zhaochong An, Ivan Vuli\'c, Serge Belongie, Anna Korhonen
Abstract: Vision-Language Models have excelled at textual reasoning, but they often struggle with fine-grained spatial understanding and continuous action planning, failing to simulate the dynamics required for complex visual reasoning. In this work, we formulate visual reasoning by means of video generation models, positing that generated frames can act as intermediate reasoning steps between initial states and solutions. We evaluate their capacity in two distinct regimes: Maze Navigation for sequential discrete planning with low visual change and Tangram Puzzle for continuous manipulation with high visual change. Our experiments reveal three critical insights: (1) Robust Zero-Shot Generalization: In both tasks, the model demonstrates strong performance on unseen data distributions without specific finetuning. (2) Visual Context: The model effectively uses visual context as explicit control, such as agent icons and tangram shapes, enabling it to maintain high visual consistency and adapt its planning capability robustly to unseen patterns. (3) Visual Test-Time Scaling: We observe a test-time scaling law in sequential planning; increasing the generated video length (visual inference budget) empowers better zero-shot generalization to spatially and temporally complex paths. These findings suggest that video generation is not merely a media tool, but a scalable, generalizable paradigm for visual reasoning.
Authors: Yiqiao Liao, Farinaz Koushanfar, Parinaz Naghizadeh
Abstract: Unsupervised neural combinatorial optimization (NCO) enables learning powerful solvers without access to ground-truth solutions. Existing approaches fall into two disjoint paradigms: models trained for generalization across instances, and instance-specific models optimized independently at test time. While the former are efficient during inference, they lack effective instance-wise adaptability; the latter are flexible but fail to exploit learned inductive structure and are prone to poor local optima. This motivates the central question of our work: how can we leverage the inductive bias learned through generalization while unlocking the flexibility required for effective instance-wise adaptation? We first identify a challenge in bridging these two paradigms: generalization-focused models often constitute poor warm starts for instance-wise optimization, potentially underperforming even randomly initialized models when fine-tuned at test time. To resolve this incompatibility, we propose TACO, a model-agnostic test-time adaptation framework that unifies and extends the two existing paradigms for unsupervised NCO. TACO applies strategic warm-starting to partially relax trained parameters while preserving inductive bias, enabling rapid and effective unsupervised adaptation. Crucially, compared to naively fine-tuning a trained generalizable model or optimizing an instance-specific model from scratch, TACO achieves better solution quality while incurring negligible additional computational cost. Experiments on canonical CO problems, Minimum Vertex Cover and Maximum Clique, demonstrate the effectiveness and robustness of TACO across static, distribution-shifted, and dynamic combinatorial optimization problems, establishing it as a practical bridge between generalizable and instance-specific unsupervised NCO.
Authors: Haokun Zhou
Abstract: Conventional anomaly detection in multivariate time series relies on the assumption that the set of observed variables remains static. In operational environments, however, monitoring systems frequently experience sensor churn. Signals may appear, disappear, or be renamed, creating data windows where the cardinality varies and may include values unseen during training. To address this challenge, we propose SMKC, a framework that decouples the dynamic input structure from the anomaly detector. We first employ permutation-invariant feature hashing to sketch raw inputs into a fixed size state sequence. We then construct a hybrid kernel image to capture global temporal structure through pairwise comparisons of the sequence and its derivatives. The model learns normal patterns using masked reconstruction and a teacher-student prediction objective. Our evaluation reveals that robust log-distance channels provide the primary discriminative signal, whereas cosine representations often fail to capture sufficient contrast. Notably, we find that a detector using random projections and nearest neighbors on the SMKC representation performs competitively with fully trained baselines without requiring gradient updates. This highlights the effectiveness of the representation itself and offers a practical cold-start solution for resource-constrained deployments.
Authors: Seungki Hong, Kyeongwon Jeong, Taekwang Jang
Abstract: Ising machines have emerged as accelerators for combinatorial optimization. To enable practical deployment, this work aims to reduce time-to-solution by addressing three challenges: (1) hardware topology, (2) spin selection and update algorithms, and (3) scalable coupling-coefficient precision. Restricted topologies require minor embedding; naive parallel updates can oscillate or stall; and limited precision can preclude feasible mappings or degrade solution quality. This work presents Snowball, a digital, scalable, all-to-all coupled Ising machine that integrates dual-mode Markov chain Monte Carlo spin selection with asynchronous spin updates to promote convergence and reduce time-to-solution. The digital architecture supports wide, configurable coupling precision, unlike many analog realizations at high bit widths. A prototype on an AMD Alveo U250 accelerator card achieves an 8$\times$ reduction in time-to-solution relative to a state-of-the-art Ising machine on the same benchmark instance.
Authors: Zhuoyan Li, Aditya Bansal, Jinzhao Li, Shishuang He, Zhuoran Lu, Mutian Zhang, Qin Liu, Yiwei Yang, Swati Jain, Ming Yin, Yunyao Li
Abstract: Large language models (LLMs) are increasingly used to automate feature engineering in tabular learning. Given task-specific information, LLMs can propose diverse feature transformation operations to enhance downstream model performance. However, current approaches typically assign the LLM as a black-box optimizer, responsible for both proposing and selecting operations based solely on its internal heuristics, which often lack calibrated estimations of operation utility and consequently lead to repeated exploration of low-yield operations without a principled strategy for prioritizing promising directions. In this paper, we propose a human-LLM collaborative feature engineering framework for tabular learning. We begin by decoupling the transformation operation proposal and selection processes, where LLMs are used solely to generate operation candidates, while the selection is guided by explicitly modeling the utility and uncertainty of each proposed operation. Since accurate utility estimation can be difficult especially in the early rounds of feature engineering, we design a mechanism within the framework that selectively elicits and incorporates human expert preference feedback, comparing which operations are more promising, into the selection process to help identify more effective operations. Our evaluations on both the synthetic study and the real user study demonstrate that the proposed framework improves feature engineering performance across a variety of tabular datasets and reduces users' cognitive load during the feature engineering process.
Authors: Alexandre Larouche, Audrey Durand
Abstract: Generative Flow Networks (GFlowNets; GFNs) are a class of generative models that learn to sample compositional objects proportionally to their a priori unknown value, their reward. We focus on the case where the reward has a specified, actionable structure, namely that it is submodular. We show submodularity can be harnessed to retrieve upper bounds on the reward of compositional objects that have not yet been observed. We provide in-depth analyses of the probability of such bounds occurring, as well as how many unobserved compositional objects can be covered by a bound. Following the Optimism in the Face of Uncertainty principle, we then introduce SUBo-GFN, which uses the submodular upper bounds to train a GFN. We show that SUBo-GFN generates orders of magnitude more training data than classical GFNs for the same number of queries to the reward function. We demonstrate the effectiveness of SUBo-GFN in terms of distribution matching and high-quality candidate generation on synthetic and real-world submodular tasks.
Authors: Minghui Chen, Wenlong Deng, James Zou, Han Yu, Xiaoxiao Li
Abstract: Large language models (LLMs) are increasingly deployed as part of compound AI systems that coordinate multiple modules (e.g., retrievers, tools, verifiers) over long-horizon workflows. Recent approaches that propagate textual feedback globally (e.g., TextGrad) make it feasible to optimize such pipelines, but we find that performance degrades as system depth grows. In particular, long-horizon agentic workflows exhibit two depth-scaling failure modes: 1) exploding textual gradient, where textual feedback grows exponentially with depth, leading to prohibitively long message and amplifies evaluation biases; and 2) vanishing textual gradient, where limited long-context ability causes models overemphasize partial feedback and compression of lengthy feedback causes downstream messages to lose specificity gradually as they propagate many hops upstream. To mitigate these issues, we introduce Textual Equilibrium Propagation (TEP), a local learning principle inspired by Equilibrium Propagation in energy-based models. TEP includes two phases: 1) a free phase where a local LLM critics iteratively refine prompts until reaching equilibrium (no further improvements are suggested); and 2) a nudged phase which applies proximal prompt edits with bounded modification intensity, using task-level objectives that propagate via forward signaling rather than backward feedback chains. This design supports local prompt optimization followed by controlled adaptation toward global goals without the computational burden and signal degradation of global textual backpropagation. Across long-horizon QA benchmarks and multi-agent tool-use dataset, TEP consistently improves accuracy and efficiency over global propagation methods such as TextGrad. The gains grows with depth, while preserving the practicality of black-box LLM components in deep compound AI system.
Authors: Haoyang Li, Haibo Chen, Xin Wang, Wenwu Zhu
Abstract: Graphs are a fundamental data structure for representing relational information in domains such as social networks, molecular systems, and knowledge graphs. However, graph learning models often suffer from limited generalization when applied beyond their training distributions. In practice, distribution shifts may arise from changes in graph structure, domain semantics, available modalities, or task formulations. To address these challenges, graph foundation models (GFMs) have recently emerged, aiming to learn general-purpose representations through large-scale pretraining across diverse graphs and tasks. In this survey, we review recent progress on GFMs from the perspective of out-of-distribution (OOD) generalization. We first discuss the main challenges posed by distribution shifts in graph learning and outline a unified problem setting. We then organize existing approaches based on whether they are designed to operate under a fixed task specification or to support generalization across heterogeneous task formulations, and summarize the corresponding OOD handling strategies and pretraining objectives. Finally, we review common evaluation protocols and discuss open directions for future research. To the best of our knowledge, this paper is the first survey for OOD generalization in GFMs.
Authors: Shivam Patel, William Cocke, Gauri Joshi
Abstract: The rapidly growing ecosystem of Large Language Models (LLMs) makes it increasingly challenging to manage and utilize the vast and dynamic pool of models effectively. We propose LOCUS, a method that produces low-dimensional vector embeddings that compactly represent a language model's capabilities across queries. LOCUS is an attention-based approach that generates embeddings by a deterministic forward pass over query encodings and evaluation scores via an encoder model, enabling seamless incorporation of new models to the pool and refinement of existing model embeddings without having to perform any retraining. We additionally train a correctness predictor that uses model embeddings and query encodings to achieve state-of-the-art routing accuracy on unseen queries. Experiments show that LOCUS needs up to 4.8x fewer query evaluation samples than baselines to produce informative and robust embeddings. Moreover, the learned embedding space is geometrically meaningful: proximity reflects model similarity, enabling a range of downstream applications including model comparison and clustering, model portfolio selection, and resilient proxies of unavailable models.
Authors: Marvin Sextro, Weronika K{\l}os, Gabriel Dernbach
Abstract: Planning effective interventions in biological systems requires treatment-effect models that adapt to unseen biological contexts by identifying their specific underlying mechanisms. Yet single-cell perturbation datasets span only a handful of biological contexts, and existing methods cannot leverage new interventional evidence at inference time to adapt beyond their training data. To meta-learn a perturbation effect estimator, we present MapPFN, a prior-data fitted network (PFN) pretrained on synthetic data generated from a prior over causal perturbations. Given a set of experiments, MapPFN uses in-context learning to predict post-perturbation distributions, without gradient-based optimization. Despite being pretrained on in silico gene knockouts alone, MapPFN identifies differentially expressed genes, matching the performance of models trained on real single-cell data. Our code and data are available at https://github.com/marvinsxtr/MapPFN.
Authors: Minjae Kwon, Josephine Lamp, Lu Feng
Abstract: Safe Reinforcement Learning (RL) algorithms are typically evaluated under fixed training conditions. We investigate whether training-time safety guarantees transfer to deployment under distribution shift, using diabetes management as a safety-critical testbed. We benchmark safe RL algorithms on a unified clinical simulator and reveal a safety generalization gap: policies satisfying constraints during training frequently violate safety requirements on unseen patients. We demonstrate that test-time shielding, which filters unsafe actions using learned dynamics models, effectively restores safety across algorithms and patient populations. Across eight safe RL algorithms, three diabetes types, and three age groups, shielding achieves Time-in-Range gains of 13--14\% for strong baselines such as PPO-Lag and CPO while reducing clinical risk index and glucose variability. Our simulator and benchmark provide a platform for studying safety under distribution shift in safety-critical control domains. Code is available at https://github.com/safe-autonomy-lab/GlucoSim and https://github.com/safe-autonomy-lab/GlucoAlg.
URLs: https://github.com/safe-autonomy-lab/GlucoSim, https://github.com/safe-autonomy-lab/GlucoAlg.
Authors: Shicheng Fan, Kun Zhang, Lu Cheng
Abstract: Temporal causal representation learning methods assume that causal mechanisms switch instantaneously between discrete domains, yet real-world systems often exhibit continuous mechanism transitions. For example, a vehicle's dynamics evolve gradually through a turning maneuver, and human gait shifts smoothly from walking to running. We formalize this setting by modeling transitional mechanisms as convex combinations of finitely many atomic mechanisms, governed by time-varying mixing coefficients. Our theoretical contributions establish that both the latent causal variables and the continuous mixing trajectory are jointly identifiable. We further propose TRACE, a Mixture-of-Experts framework where each expert learns one atomic mechanism during training, enabling recovery of mechanism trajectories at test time. This formulation generalizes to intermediate mechanism states never observed during training. Experiments on synthetic and real-world data demonstrate that TRACE recovers mixing trajectories with up to 0.99 correlation, substantially outperforming discrete-switching baselines.
Authors: Kevin Han, Haolin Cong, Bowen Deng, Amir Barati Farimani
Abstract: Machine learning interatomic potentials (MLIPs) have proven to be wildly useful for molecular dynamics simulations, powering countless drug and materials discovery applications. However, MLIPs face two primary bottlenecks preventing them from reaching realistic simulation scales: inference time and memory consumption. In this work, we address both issues by challenging the long-held belief that the cutoff radius for the MLIP must be held to a fixed, constant value. For the first time, we introduce a dynamic cutoff formulation that still leads to stable, long timescale molecular dynamics simulation. In introducing the dynamic cutoff, we are able to induce sparsity onto the underlying atom graph by targeting a specific number of neighbors per atom, significantly reducing both memory consumption and inference time. We show the effectiveness of a dynamic cutoff by implementing it onto 4 state of the art MLIPs: MACE, Nequip, Orbv3, and TensorNet, leading to 2.26x less memory consumption and 2.04x faster inference time, depending on the model and atomic system. We also perform an extensive error analysis and find that the dynamic cutoff models exhibit minimal accuracy dropoff compared to their fixed cutoff counterparts on both materials and molecular datasets. All model implementations and training code will be fully open sourced.
Authors: Maria Despoina Siampou, Shushman Choudhury, Shang-Ling Hsu, Neha Arora, Cyrus Shahabi
Abstract: Recent progress in geospatial foundation models highlights the importance of learning general-purpose representations for real-world locations, particularly points-of-interest (POIs) where human activity concentrates. Existing approaches, however, focus primarily on place identity derived from static textual metadata, or learn representations tied to trajectory context, which capture movement regularities rather than how places are actually used (i.e., POI's function). We argue that POI function is a missing but essential signal for general POI representations. We introduce Mobility-Embedded POIs (ME-POIs), a framework that augments POI embeddings derived, from language models with large-scale human mobility data to learn POI-centric, context-independent representations grounded in real-world usage. ME-POIs encodes individual visits as temporally contextualized embeddings and aligns them with learnable POI representations via contrastive learning to capture usage patterns across users and time. To address long-tail sparsity, we propose a novel mechanism that propagates temporal visit patterns from nearby, frequently visited POIs across multiple spatial scales. We evaluate ME-POIs on five newly proposed map enrichment tasks, testing its ability to capture both the identity and function of POIs. Across all tasks, augmenting text-based embeddings with ME-POIs consistently outperforms both text-only and mobility-only baselines. Notably, ME-POIs trained on mobility data alone can surpass text-only models on certain tasks, highlighting that POI function is a critical component of accurate and generalizable POI representations.
Authors: Henry Kvinge, Andrew Aguilar, Nayda Farnsworth, Grace O'Brien, Robert Jasper, Sarah Scullen, Helen Jenne
Abstract: While a real-world research program in mathematics may be guided by a motivating question, the process of mathematical discovery is typically open-ended. Ideally, exploration needed to answer the original question will reveal new structures, patterns, and insights that are valuable in their own right. This contrasts with the exam-style paradigm in which the machine learning community typically applies AI to math. To maximize progress in mathematics using AI, we will need to go beyond simple question answering. With this in mind, we explore the extent to which narrow models trained to solve a fixed mathematical task learn broader mathematical structure that can be extracted by a researcher or other AI system. As a basic test case for this, we use the task of training a neural network to predict a group operation (for example, performing modular arithmetic or composition of permutations). We describe a suite of tests designed to assess whether the model captures significant group-theoretic notions such as the identity element, commutativity, or subgroups. Through extensive experimentation we find evidence that models learn representations capable of capturing abstract algebraic properties. For example, we find hints that models capture the commutativity of modular arithmetic. We are also able to train linear classifiers that reliably distinguish between elements of certain subgroups (even though no labels for these subgroups are included in the data). On the other hand, we are unable to extract notions such as the concept of the identity element. Together, our results suggest that in some cases the representations of even small neural networks can be used to distill interesting abstract structure from new mathematical objects.
Authors: Carlos A. Pereira, St\'ephane Gaudreault, Valentin Dallerit, Christopher Subich, Shoyon Panday, Siqi Wei, Sasa Zhang, Siddharth Rout, Eldad Haber, Raymond J. Spiteri, David Millard, Emilia Diaconescu
Abstract: Recent machine-learning approaches to weather forecasting often employ a monolithic architecture, where distinct physical mechanisms (advection, transport), diffusion-like mixing, thermodynamic processes, and forcing are represented implicitly within a single large network. This representation is particularly problematic for advection, where long-range transport must be treated with expensive global interaction mechanisms or through deep, stacked convolutional layers. To mitigate this, we present PARADIS, a physics-inspired global weather prediction model that imposes inductive biases on network behavior through a functional decomposition into advection, diffusion, and reaction blocks acting on latent variables. We implement advection through a Neural Semi-Lagrangian operator that performs trajectory-based transport via differentiable interpolation on the sphere, enabling end-to-end learning of both the latent modes to be transported and their characteristic trajectories. Diffusion-like processes are modeled through depthwise-separable spatial mixing, while local source terms and vertical interactions are modeled via pointwise channel interactions, enabling operator-level physical structure. PARADIS provides state-of-the-art forecast skill at a fraction of the training cost. On ERA5-based benchmarks, the 1 degree PARADIS model, with a total training cost of less than a GPU month, meets or exceeds the performance of 0.25 degree traditional and machine-learning baselines, including the ECMWF HRES forecast and DeepMind's GraphCast.
Authors: Michael Ibrahim, Nagi Gebraeel, Weijun Xie
Abstract: We study the problem of federated clustering when the total number of clusters $K$ across clients is unknown, and the clients have heterogeneous but potentially overlapping cluster sets in their local data. To that end, we develop FedGEM: a federated generalized expectation-maximization algorithm for the training of mixture models with an unknown number of components. Our proposed algorithm relies on each of the clients performing EM steps locally, and constructing an uncertainty set around the maximizer associated with each local component. The central server utilizes the uncertainty sets to learn potential cluster overlaps between clients, and infer the global number of clusters via closed-form computations. We perform a thorough theoretical study of our algorithm, presenting probabilistic convergence guarantees under common assumptions. Subsequently, we study the specific setting of isotropic GMMs, providing tractable, low-complexity computations to be performed by each client during each iteration of the algorithm, as well as rigorously verifying assumptions required for algorithm convergence. We perform various numerical experiments, where we empirically demonstrate that our proposed method achieves comparable performance to centralized EM, and that it outperforms various existing federated clustering methods.
Authors: Shuai Liu, Alireza Bakhtiari, Alex Ayoub, Botao Hao, Csaba Szepesv\'ari
Abstract: We study stochastic contextual logistic bandits under the simple regret objective. While simple regret guarantees have been established for the linear case, no such results were previously known for the logistic setting. Building on ideas from contextual linear bandits and self-concordant analysis, we propose the first algorithm that achieves simple regret $\tilde{\mathcal{O}}(d/\sqrt{T})$. Notably, the leading term of our regret bound is free of the constant $\kappa = \mathcal O(\exp(S))$, where $S$ is a bound on the magnitude of the unknown parameter vector. The algorithm is shown to be fully tractable when the action set is finite. We also introduce a new variant of Thompson Sampling tailored to the simple-regret setting. This yields the first simple regret guarantee for randomized algorithms in stochastic contextual linear bandits, with regret $\tilde{\mathcal{O}}(d^{3/2}/\sqrt{T})$. Extending this method to the logistic case, we obtain a similarly structured Thompson Sampling algorithm that achieves the same regret bound -- $\tilde{\mathcal{O}}(d^{3/2}/\sqrt{T})$ -- again with no dependence on $\kappa$ in the leading term. The randomized algorithms, as expected, are cheaper to run than their deterministic counterparts. Finally, we conducted a series of experiments to empirically validate these theoretical guarantees.
Authors: Augusto Santos, Teresa Santos, Catarina Rodrigues, Jos\'e M. F. Moura
Abstract: Emergent phenomena -- onset of epileptic seizures, sudden customer churn, or pandemic outbreaks -- often arise from hidden causal interactions in complex systems. We propose a machine learning method for their early detection that addresses a core challenge: unveiling and harnessing a system's latent causal structure despite the data-generating process being unknown and partially observed. The method learns an optimal feature representation from a one-parameter family of estimators -- powers of the empirical covariance or precision matrix -- offering a principled way to tune in to the underlying structure driving the emergence of critical events. A supervised learning module then classifies the learned representation. We prove structural consistency of the family and demonstrate the empirical soundness of our approach on seizure detection and churn prediction, attaining competitive results in both. Beyond prediction, and toward explainability, we ascertain that the optimal covariance power exhibits evidence of good identifiability while capturing structural signatures, thus reconciling predictive performance with interpretable statistical structure.
Authors: Kamal Berahmand, Saman Forouzandeh, Mehrnoush Mohammadi, Parham Moradi, Mahdi Jalili
Abstract: Graph anomaly detection aims to identify abnormal patterns in networks, but faces significant challenges from label scarcity and extreme class imbalance. While graph contrastive learning offers a promising unsupervised solution, existing methods suffer from two critical limitations: random augmentations break semantic consistency in positive pairs, while naive negative sampling produces trivial, uninformative contrasts. We propose AC2L-GAD, an Active Counterfactual Contrastive Learning framework that addresses both limitations through principled counterfactual reasoning. By combining information-theoretic active selection with counterfactual generation, our approach identifies structurally complex nodes and generates anomaly-preserving positive augmentations alongside normal negative counterparts that provide hard contrasts, while restricting expensive counterfactual generation to a strategically selected subset. This design reduces computational overhead by approximately 65% compared to full-graph counterfactual generation while maintaining detection quality. Experiments on nine benchmark datasets, including real-world financial transaction graphs from GADBench, show that AC2L-GAD achieves competitive or superior performance compared to state-of-the-art baselines, with notable gains in datasets where anomalies exhibit complex attribute-structure interactions.
Authors: Yuanning Cui, Zequn Sun, Wei Hu, Kexuan Xin, Zhangjie Fu
Abstract: Entity alignment (EA) is critical for knowledge graph (KG) fusion. Existing EA models lack transferability and are incapable of aligning unseen KGs without retraining. While using graph foundation models (GFMs) offer a solution, we find that directly adapting GFMs to EA remains largely ineffective. This stems from a critical "reasoning horizon gap": unlike link prediction in GFMs, EA necessitates capturing long-range dependencies across sparse and heterogeneous KG structuresTo address this challenge, we propose a EA foundation model driven by a parallel encoding strategy. We utilize seed EA pairs as local anchors to guide the information flow, initializing and encoding two parallel streams simultaneously. This facilitates anchor-conditioned message passing and significantly shortens the inference trajectory by leveraging local structural proximity instead of global search. Additionally, we incorporate a merged relation graph to model global dependencies and a learnable interaction module for precise matching. Extensive experiments verify the effectiveness of our framework, highlighting its strong generalizability to unseen KGs.
Authors: Xin Peng, Ang Gao
Abstract: The scalability of continuous normalizing flows (CNFs) for unbiased Boltzmann sampling remains limited in high-dimensional systems due to the cost of Jacobian-determinant evaluation, which requires $D$ backpropagation passes through the flow layers. Existing stochastic Jacobian estimators such as the Hutchinson trace estimator reduce computation but introduce bias, while the recently proposed Flow Perturbation method is unbiased yet suffers from high variance. We present \textbf{Flow Perturbation++}, a variance-reduced extension of Flow Perturbation that discretizes the probability-flow ODE and performs unbiased stepwise Jacobian estimation at each integration step. This multi-step construction retains the unbiasedness of Flow Perturbation while achieves substantially lower estimator variance. Integrated into a Sequential Monte Carlo framework, Flow Perturbation++ achieves significantly improved equilibrium sampling on a 1000D Gaussian Mixture Model and the all-atom Chignolin protein compared with Hutchinson-based and single-step Flow Perturbation baselines.
Authors: Xin Peng, Ang Gao
Abstract: Generative models, including diffusion and flow-based models, often exhibit systematic biases that degrade sample quality, particularly in high-dimensional settings. We revisit refinement methods and show that effective bias correction can be achieved as a post-hoc procedure, without noise injection or multi-step resampling of the sampling process. We propose a flow-matching-based \textbf{Bi-stage Flow Refinement (BFR)} framework with two refinement strategies operating at different stages: latent space alignment for approximately invertible generators and data space refinement trained with lightweight augmentations. Unlike previous refiners that perturb sampling dynamics, BFR preserves the original ODE trajectory and applies deterministic corrections to generated samples. Experiments on MNIST, CIFAR-10, and FFHQ at 256x256 resolution demonstrate consistent improvements in fidelity and coverage; notably, starting from base samples with FID 3.95, latent space refinement achieves a \textbf{state-of-the-art} FID of \textbf{1.46} on MNIST using only a single additional function evaluation (1-NFE), while maintaining sample diversity.
Authors: Weiguang Wang, Yong Liu, Yingjie Gao, Guangyuan Xu
Abstract: Steady-state visually evoked potentials (SSVEP)-based brain-computer interfaces (BCIs) are widely used due to their high signal-to-noise ratio and user-friendliness. Accurate decoding of SSVEP signals is crucial for interpreting user intentions in BCI applications. However, signal variability across subjects and the costly user-specific annotation limit recognition performance. Therefore, we propose a novel cross-subject domain adaptation method built upon the self-training paradigm. Specifically, a Filter-Bank Euclidean Alignment (FBEA) strategy is designed to exploit frequency information from SSVEP filter banks. Then, we propose a Cross-Subject Self-Training (CSST) framework consisting of two stages: Pre-Training with Adversarial Learning (PTAL), which aligns the source and target distributions, and Dual-Ensemble Self-Training (DEST), which refines pseudo-label quality. Moreover, we introduce a Time-Frequency Augmented Contrastive Learning (TFA-CL) module to enhance feature discriminability across multiple augmented views. Extensive experiments on the Benchmark and BETA datasets demonstrate that our approach achieves state-of-the-art performance across varying signal lengths, highlighting its superiority.
Authors: Chuan-Shen Hu
Abstract: Combinatorial and topological structures, such as graphs, simplicial complexes, and cell complexes, form the foundation of geometric and topological deep learning (GDL and TDL) architectures. These models aggregate signals over such domains, integrate local features, and generate representations for diverse real-world applications. However, the distribution and diffusion behavior of GDL and TDL features during training remains an open and underexplored problem. Motivated by this gap, we introduce a cellular sheaf theoretic framework for modeling and analyzing the local consistency and harmonicity of node features and edge weights in graph-based architectures. By tracking local feature alignments and agreements through sheaf structures, the framework offers a topological perspective on feature diffusion and aggregation. Furthermore, a multiscale extension inspired by topological data analysis (TDA) is proposed to capture hierarchical feature interactions in graph models. This approach enables a joint characterization of GDL and TDL architectures based on their underlying geometric and topological structures and the learned signals defined on them, providing insights for future studies on conventional tasks such as node classification, substructure detection, and community detection.
Authors: Mehmet Ergezer
Abstract: We study how model architecture and temporal context interact in naturalistic EEG decoding. Using the HBN movie-watching dataset, we benchmark five architectures, CNN, LSTM, a stabilized Transformer (EEGXF), S4, and S5, on a 4-class task across segment lengths from 8s to 128s. Accuracy improves with longer context: at 64s, S5 reaches 98.7%+/-0.6 and CNN 98.3%+/-0.3, while S5 uses ~20x fewer parameters than CNN. To probe real-world robustness, we evaluate zero-shot cross-frequency shifts, cross-task OOD inputs, and leave-one-subject-out generalization. S5 achieves stronger cross-subject accuracy but makes over-confident errors on OOD tasks; EEGXF is more conservative and stable under frequency shifts, though less calibrated in-distribution. These results reveal a practical efficiency-robustness trade-off: S5 for parameter-efficient peak accuracy; EEGXF when robustness and conservative uncertainty are critical.
Authors: Daniel T. Bernstein, Luca Di Carlo, David Schwab
Abstract: We show that introducing short-range attractive couplings between the weights of a neural network during training provides a novel avenue for model quantization. These couplings rapidly induce the discretization of a model's weight distribution, and they do so in a mixed-precision manner despite only relying on two additional hyperparameters. We demonstrate that, within an appropriate range of hyperparameters, our "soft quantization'' scheme outperforms histogram-equalized post-training quantization on ResNet-20/CIFAR-10. Soft quantization provides both a new pipeline for the flexible compression of machine learning models and a new tool for investigating the trade-off between compression and generalization in high-dimensional loss landscapes.
Authors: Kaiyuan Tan, Kendra Givens, Peilun Li, Thomas Beckers
Abstract: Diffusion models provide expressive priors for forecasting trajectories of dynamical systems, but are typically unreliable in the sparse data regime. Physics-informed machine learning (PIML) improves reliability in such settings; however, most methods require \emph{explicit governing equations} during training, which are often only partially known due to complex and nonlinear dynamics. We introduce \textbf{PHDME}, a port-Hamiltonian diffusion framework designed for \emph{sparse observations} and \emph{incomplete physics}. PHDME leverages port-Hamiltonian structural prior but does not require full knowledge of the closed-form governing equations. Our approach first trains a Gaussian process distributed Port-Hamiltonian system (GP-dPHS) on limited observations to capture an energy-based representation of the dynamics. The GP-dPHS is then used to generate a physically consistent artificial dataset for diffusion training, and to inform the diffusion model with a structured physics residual loss. After training, the diffusion model acts as an amortized sampler and forecaster for fast trajectory generation. Finally, we apply split conformal calibration to provide uncertainty statements for the generated predictions. Experiments on PDE benchmarks and a real-world spring system show improved accuracy and physical consistency under data scarcity.
Authors: Luwei Sun, Dongrui Shen, Jianfe Li, Yulong Zhao, Han Feng
Abstract: Motivated by challenges in conditional generative modeling, where the target conditional density takes the form of a ratio f1 over f2, this paper develops a theoretical framework for approximating such ratio-type functionals. Here, f1 and f2 are kernel-based marginal densities that capture structured interactions, a setting central to diffusion-based generative models. We provide a concise proof for approximating these ratio-type functionals using deep neural networks with the SignReLU activation function, leveraging the activation's piecewise structure. Under standard regularity assumptions, we establish L^p(Omega) approximation bounds and convergence rates. Specializing to Denoising Diffusion Probabilistic Models (DDPMs), we construct a SignReLU-based neural estimator for the reverse process and derive bounds on the excess Kullback-Leibler (KL) risk between the generated and true data distributions. Our analysis decomposes this excess risk into approximation and estimation error components. These results provide generalization guarantees for finite-sample training of diffusion-based generative models.
Authors: Yiju Guo, Tianyi Hu, Zexu Sun, Yankai Lin
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has advanced LLM reasoning, but remains constrained by inefficient exploration under limited rollout budgets, leading to low sampling success and unstable training in complex tasks. We find that many exploration failures arise not from problem difficulty, but from a small number of prompt tokens that introduce interference. Building on this insight, we propose the Less Noise Sampling Framework (LENS), which first prompts by identifying and removing interference tokens. then transfers successful rollouts from the purification process to supervise policy optimization on the original noisy prompts, enabling the model to learn to ignore interference in the real-world, noisy prompting settings. Experimental results show that LENS significantly outperforms GRPO, delivering higher performance and faster convergence, with a 3.88% average gain and over 1.6$\times$ speedup. Our work highlights the critical role of pruning interference tokens in improving rollout efficiency, offering a new perspective for RLVR research.
Authors: Namkyung Yoon, Sanghong Kim, Hwangnam Kim
Abstract: Gas chromatography-mass spectrometry (GC-MS) is a widely used analytical method for chemical substance detection, but measurement reliability tends to deteriorate in the presence of interfering substances. In particular, interfering substances cause nonspecific peaks, residence time shifts, and increased background noise, resulting in reduced sensitivity and false alarms. To overcome these challenges, in this paper, we propose an artificial intelligence discrimination framework based on a peak-aware conditional generative model to improve the reliability of GC-MS measurements under interference conditions. The framework is learned with a novel peak-aware mechanism that highlights the characteristic peaks of GC-MS data, allowing it to generate important spectral features more faithfully. In addition, chemical and solvent information is encoded in a latent vector embedded with it, allowing a conditional generative adversarial neural network (CGAN) to generate a synthetic GC-MS signal consistent with the experimental conditions. This generates an experimental dataset that assumes indirect substance situations in chemical substance data, where acquisition is limited without conducting real experiments. These data are used for the learning of AI-based GC-MS discrimination models to help in accurate chemical substance discrimination. We conduct various quantitative and qualitative evaluations of the generated simulated data to verify the validity of the proposed framework. We also verify how the generative model improves the performance of the AI discrimination framework. Representatively, the proposed method is shown to consistently achieve cosine similarity and Pearson correlation coefficient values above 0.9 while preserving peak number diversity and reducing false alarms in the discrimination model.
Authors: Zhuochen Liu, Hans Walker, Rahul Jain
Abstract: Neural network models are increasingly used for state estimation in control and decision-making problems, yet it remains unclear to what extent they behave as principled filters in nonlinear dynamical systems. Unlike classical filters, which rely on explicit knowledge of system dynamics and noise models, neural estimators can be trained purely from data without access to the underlying system equations. In this work, we present a systematic empirical comparison between such model-free neural network models and classical filtering methods across multiple nonlinear scenarios. Our study evaluates Transformer-based models, state-space neural networks, and recurrent architectures alongside particle filters and nonlinear Kalman filters. The results show that neural models (in particular, state-space models (SSMs)) achieve state estimation performance that approaches strong nonlinear Kalman filters in nonlinear scenarios and outperform weaker classical baselines despite lacking access to system models, while also attaining substantially higher inference throughput.
Authors: Licheng Wang, Yuzi Yan, Mingtao Huang, Yuan Shen
Abstract: Neural combinatorial optimization (NCO) solvers, implemented with graph neural networks (GNNs), have introduced new approaches for solving routing problems. Trained with reinforcement learning (RL), the state-of-the-art graph attention model (GAM) achieves near-optimal solutions without requiring expert knowledge or labeled data. In this work, we generalize the existing graph attention mechanism and propose the extended graph attention model (EGAM). Our model utilizes multi-head dot-product attention to update both node and edge embeddings, addressing the limitations of the conventional GAM, which considers only node features. We employ an autoregressive encoder-decoder architecture and train it with policy gradient algorithms that incorporate a specially designed baseline. Experiments show that EGAM matches or outperforms existing methods across various routing problems. Notably, the proposed model demonstrates exceptional performance on highly constrained problems, highlighting its efficiency in handling complex graph structures.
Authors: Yisheng Zhong, Zhengbang Yang, Zhuangdi Zhu
Abstract: LLM unlearning is a technique to remove the impacts of undesirable knowledge from the model without retraining from scratch, which is indispensable towards trustworthy AI. Existing unlearning methods face significant limitations: conventional tuning-based unlearning is computationally heavy and prone to catastrophic forgetting. In contrast, in-contextualized unlearning is lightweight for precise unlearning but vulnerable to prompt removal or reverse engineering attacks. In response, we propose Distilled Unlearning from an Efficient Teacher (DUET), a novel distillation-based unlearning method that combines the merits of these two lines of work. It learns a student model to imitate the behavior of a prompt-steered teacher that effectively refuses undesirable knowledge generation while preserving general domain knowledge. Extensive evaluations on existing benchmarks with our enriched evaluation protocols demonstrate that DUET achieves higher performance in both forgetting and utility preservation, while being orders of magnitude more data-efficient than state-of-the-art unlearning methods.
Authors: Tianyi Zeng, Tianyi Wang, Jiaru Zhang, Zimo Zeng, Feiyang Zhang, Yiming Xu, Sikai Chen, Yajie Zou, Yangyang Wang, Junfeng Jiao, Christian Claudel, Xinbo Chen
Abstract: Diffusion models have emerged as powerful generative tools for modeling complex data distributions, yet their purely data-driven nature limits applicability in practical engineering and scientific problems where physical laws need to be followed. This paper proposes Physics-Informed Learning via Diffusion (PILD), a framework that unifies diffusion modeling and first-principles physical constraints by introducing a virtual residual observation sampled from a Laplace distribution to supervise generation during training. To further integrate physical laws, a conditional embedding module is incorporated to inject physical information into the denoising network at multiple layers, ensuring consistent guidance throughout the diffusion process. The proposed PILD framework is concise, modular, and broadly applicable to problems governed by ordinary differential equations, partial differential equations, as well as algebraic equations or inequality constraints. Extensive experiments across engineering and scientific tasks including estimating vehicle trajectories, tire forces, Darcy flow and plasma dynamics, demonstrate that our PILD substantially improves accuracy, stability, and generalization over existing physics-informed and diffusion-based baselines.
Authors: Ruifeng Zhang, Zexi Huang, Zikai Wang, Ke Sun, Bohang Zheng, Zhen Ouyang, Huimin Xie, Phil Shen, Junlin Zhang, Wentao Guo, Qinglei Wang
Abstract: Accurately capturing feature interactions is essential in recommender systems, and recent trends show that scaling up model capacity could be a key driver for next-level predictive performance. While prior work has explored various model architectures to capture multi-granularity feature interactions, relatively little attention has been paid to efficient feature handling and scaling model capacity without incurring excessive inference latency. In this paper, we address this by presenting Zenith, a scalable and efficient ranking architecture that learns complex feature interactions with minimal runtime overhead. Zenith is designed to handle a few high-dimensional Prime Tokens with Token Fusion and Token Boost modules, which exhibits superior scaling laws compared to other state-of-the-art ranking methods, thanks to its improved token heterogeneity. Its real-world effectiveness is demonstrated by deploying the architecture to TikTok Live, a leading online livestreaming platform that attracts billions of users globally. Our A/B test shows that Zenith achieves +1.05%/-1.10% in online CTR AUC and Logloss, and realizes +9.93% gains in Quality Watch Session / User and +8.11% in Quality Watch Duration / User.
Authors: Akash Pandey, Payal Mohapatra, Wei Chen, Qi Zhu, Sinan Keten
Abstract: Identifying the extent to which every temporal segment influences a model's predictions is essential for explaining model decisions and increasing transparency. While post-hoc explainable methods based on gradients and feature-based attributions have been popular, they suffer from reference state sensitivity and struggle to generalize across time-series datasets, as they treat time points independently and ignore sequential dependencies. Another perspective on explainable time-series classification is through interpretable components of the model, for instance, leveraging self-attention mechanisms to estimate temporal attribution; however, recent findings indicate that these attention weights often fail to provide faithful measures of temporal importance. In this work, we advance this perspective and present a novel explainability-driven deep learning framework, TimeSliver, which jointly utilizes raw time-series data and its symbolic abstraction to construct a representation that maintains the original temporal structure. Each element in this representation linearly encodes the contribution of each temporal segment to the final prediction, allowing us to assign a meaningful importance score to every time point. For time-series classification, TimeSliver outperforms other temporal attribution methods by 11% on 7 distinct synthetic and real-world multivariate time-series datasets. TimeSliver also achieves predictive performance within 2% of state-of-the-art baselines across 26 UEA benchmark datasets, positioning it as a strong and explainable framework for general time-series classification.
Authors: Changyu Li, Dingcheng Huang, Kexuan Yao, Xiaoya Ni, Lijuan Shen, Fei Luo
Abstract: Reliability-centered prognostics for rotating machinery requires early warning signals that remain accurate under nonstationary operating conditions, domain shifts across speed/load/sensors, and severe class imbalance, while keeping the false-alarm rate small and predictable. We propose the Physics-Guided Tiny-Mamba Transformer (PG-TMT), a compact tri-branch encoder tailored for online condition monitoring. A depthwise-separable convolutional stem captures micro-transients, a Tiny-Mamba state-space branch models near-linear long-range dynamics, and a lightweight local Transformer encodes cross-channel resonances. We derive an analytic temporal-to-spectral mapping that ties the model's attention spectrum to classical bearing fault-order bands, yielding a band-alignment score that quantifies physical plausibility and provides physics-grounded explanations. To ensure decision reliability, healthy-score exceedances are modeled with extreme-value theory (EVT), which yields an on-threshold achieving a target false-alarm intensity (events/hour); a dual-threshold hysteresis with a minimum hold time further suppresses chatter. Under a leakage-free streaming protocol with right-censoring of missed detections on CWRU, Paderborn, XJTU-SY, and an industrial pilot, PG-TMT attains higher precision-recall AUC (primary under imbalance), competitive or better ROC AUC, and shorter mean time-to-detect at matched false-alarm intensity, together with strong cross-domain transfer. By coupling physics-aligned representations with EVT-calibrated decision rules, PG-TMT delivers calibrated, interpretable, and deployment-ready early warnings for reliability-centric prognostics and health management.
Authors: Anders Gj{\o}lbye, Ida Kargaard, Emma Kargaard, Lars Kai Hansen
Abstract: Partial Least Squares (PLS) learns shared structure from paired data via the top singular vectors of the empirical cross-covariance (PLS-SVD), but multimodal datasets often have missing entries in both views. We study PLS-SVD under independent entry-wise missing-completely-at-random masking in a proportional high-dimensional spiked model. After appropriate normalization, the masked cross-covariance behaves like a spiked rectangular random matrix whose effective signal strength is attenuated by $\sqrt{\rho}$, where $\rho$ is the joint entry retention probability. As a result, PLS-SVD exhibits a sharp BBP-type phase transition: below a critical signal-to-noise threshold the leading singular vectors are asymptotically uninformative, while above it they achieve nontrivial alignment with the latent shared directions, with closed-form asymptotic overlap formulas. Simulations and semi-synthetic multimodal experiments corroborate the predicted phase diagram and recovery curves across aspect ratios, signal strengths, and missingness levels.
Authors: Shaobo Wang, Yantai Yang, Guo Chen, Peiru Li, Kaixin Li, Yufa Zhou, Zhaorun Chen, Linfeng Zhang
Abstract: Dataset Distillation (DD) seeks to create a compact dataset from a large, real-world dataset. While recent methods often rely on heuristic approaches to balance efficiency and quality, the fundamental relationship between original and synthetic data remains underexplored. This paper revisits knowledge distillation-based dataset distillation within a solid theoretical framework. We introduce the concepts of Informativeness and Utility, capturing crucial information within a sample and essential samples in the training set, respectively. Building on these principles, we define optimal dataset distillation mathematically. We then present InfoUtil, a framework that balances informativeness and utility in synthesizing the distilled dataset. InfoUtil incorporates two key components: (1) game-theoretic informativeness maximization using Shapley Value attribution to extract key information from samples, and (2) principled utility maximization by selecting globally influential samples based on Gradient Norm. These components ensure that the distilled dataset is both informative and utility-optimized. Experiments demonstrate that our method achieves a 6.1\% performance improvement over the previous state-of-the-art approach on ImageNet-1K dataset using ResNet-18.
Authors: Zijun Chen, Zaiwei Chen, Nian Si, Shengbo Wang
Abstract: We present the convergence rates of synchronous and asynchronous Q-learning for average-reward Markov decision processes, where the absence of contraction poses a fundamental challenge. Existing non-asymptotic results overcome this challenge by either imposing strong assumptions to enforce seminorm contraction or relying on discounted or episodic Markov decision processes as successive approximations, which either require unknown parameters or result in suboptimal sample complexity. In this work, under a reachability assumption, we establish optimal $\widetilde{O}(\varepsilon^{-2})$ sample complexity guarantees (up to logarithmic factors) for a simple variant of synchronous and asynchronous Q-learning that samples from the lazified dynamics, where the system remains in the current state with some fixed probability. At the core of our analysis is the construction of an instance-dependent seminorm and showing that, after a lazy transformation of the Markov decision process, the Bellman operator becomes one-step contractive under this seminorm.
Authors: Wei-Di Chang, Mikael Henaff, Brandon Amos, Gregory Dudek, Scott Fujimoto
Abstract: This paper investigates search in model-based reinforcement learning (RL). Conventional wisdom holds that long-term predictions and compounding errors are the primary obstacles for model-based RL. We challenge this view, showing that search is not a plug-and-play replacement for a learned policy. Surprisingly, we find that search can harm performance even when the model is highly accurate. Instead, we show that mitigating distribution shift matters more than improving model or value function accuracy. Building on this insight, we identify key techniques for enabling effective search, achieving state-of-the-art performance across multiple popular benchmark domains.
Authors: Huaming Du, Yijie Huang, Su Yao, Yiying Wang, Yueyang Zhou, Jingwen Yang, Jinshi Zhang, Han Ji, Yu Zhao, Guisong Liu, Hegui Zhang, Carl Yang, Gang Kou
Abstract: The increasing scale of graph datasets has significantly improved the performance of graph representation learning methods, but it has also introduced substantial training challenges. Graph dataset condensation techniques have emerged to compress large datasets into smaller yet information-rich datasets, while maintaining similar test performance. However, these methods strictly require downstream applications to match the original dataset and task, which often fails in cross-task and cross-domain scenarios. To address these challenges, we propose a novel causal-invariance-based and transferable graph dataset condensation method, named \textbf{TGCC}, providing effective and transferable condensed datasets. Specifically, to preserve domain-invariant knowledge, we first extract domain causal-invariant features from the spatial domain of the graph using causal interventions. Then, to fully capture the structural and feature information of the original graph, we perform enhanced condensation operations. Finally, through spectral-domain enhanced contrastive learning, we inject the causal-invariant features into the condensed graph, ensuring that the compressed graph retains the causal information of the original graph. Experimental results on five public datasets and our novel \textbf{FinReport} dataset demonstrate that TGCC achieves up to a 13.41\% improvement in cross-task and cross-domain complex scenarios compared to existing methods, and achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.
Authors: Xiaozhuang Li, Xindi Tang, Fang He
Abstract: With the rapid expansion of electric vehicles (EVs) and charging infrastructure, the effective management of Autonomous Electric Taxi (AET) fleets faces a critical challenge in environments with dynamic and uncertain charging availability. While most existing research assumes a static charging network, this simplification creates a significant gap between theoretical models and real-world operations. To bridge this gap, we propose GAT-PEARL, a novel meta-reinforcement learning framework that learns an adaptive operational policy. Our approach integrates a graph attention network (GAT) to effectively extract robust spatial representations under infrastructure layouts and model the complex spatiotemporal relationships of the urban environment, and employs probabilistic embeddings for actor-critic reinforcement learning (PEARL) to enable rapid, inference-based adaptation to changes in charging network layouts without retraining. Through extensive simulations on real-world data in Chengdu, China, we demonstrate that GAT-PEARL significantly outperforms conventional reinforcement learning baselines, showing superior generalization to unseen infrastructure layouts and achieving higher overall operational efficiency in dynamic settings.
Authors: Seonghwi Kim, Sung Ho Jo, Wooseok Ha, Minwoo Chae
Abstract: Unsupervised domain adaptation (UDA) is a statistical learning problem when the distribution of training (source) data is different from that of test (target) data. In this setting, one has access to labeled data only from the source domain and unlabeled data from the target domain. The central objective is to leverage the source data and the unlabeled target data to build models that generalize to the target domain. Despite its potential, existing UDA approaches often struggle in practice, particularly in scenarios where the target domain offers only limited unlabeled data or spurious correlations dominate the source domain. To address these challenges, we propose a novel distributionally robust learning framework that models uncertainty in both the covariate distribution and the conditional label distribution. Our approach is motivated by the multi-source domain adaptation setting but is also directly applicable to the single-source scenario, making it versatile in practice. We develop an efficient learning algorithm that can be seamlessly integrated with existing UDA methods. Extensive experiments under various distribution shift scenarios show that our method consistently outperforms strong baselines, especially when target data are extremely scarce.
Authors: Aoyu Pang, Maonan Wang, Zifan Sha, Wenwei Yue, Changle Li, Chung Shue Chen, Man-On Pun
Abstract: Urban Air Mobility (UAM) has emerged as a transformative solution to alleviate urban congestion by utilizing low-altitude airspace, thereby reducing pressure on ground transportation networks. To enable truly efficient and seamless door-to-door travel experiences, UAM requires close integration with existing ground transportation infrastructure. However, current research on optimal integrated routing strategies for passengers in air-ground mobility systems remains limited, with a lack of systematic exploration.To address this gap, we first propose a unified optimization model that integrates strategy selection for both air and ground transportation. This model captures the dynamic characteristics of multimodal transport networks and incorporates real-time traffic conditions alongside passenger decision-making behavior. Building on this model, we propose a Unified Air-Ground Mobility Coordination (UAGMC) framework, which leverages deep reinforcement learning (RL) and Vehicle-to-Everything (V2X) communication to optimize vertiport selection and dynamically plan air taxi routes. Experimental results demonstrate that UAGMC achieves a 34\% reduction in average travel time compared to conventional proportional allocation methods, enhancing overall travel efficiency and providing novel insights into the integration and optimization of multimodal transportation systems. This work lays a solid foundation for advancing intelligent urban mobility solutions through the coordination of air and ground transportation modes. The related code can be found at https://github.com/Traffic-Alpha/UAGMC.
Authors: Achraf Hsain, Ahmed Abdelkader, Emmanuel Baldwin Mbaya, Hamoud Aljamaan
Abstract: Deep neural networks are vulnerable to adversarial examples--inputs with imperceptible perturbations causing misclassification. While adversarial transfer within neural networks is well-documented, whether classical ML pipelines using handcrafted features inherit this vulnerability when attacked via neural surrogates remains unexplored. Feature engineering creates information bottlenecks through gradient quantization and spatial binning, potentially filtering high-frequency adversarial signals. We evaluate this hypothesis through the first comprehensive study of adversarial transfer from DNNs to HOG-based classifiers. Using VGG16 as a surrogate, we generate FGSM and PGD adversarial examples and test transfer to four classical classifiers (KNN, Decision Tree, Linear SVM, Kernel SVM) and a shallow neural network across eight HOG configurations on CIFAR-10. Our results strongly refute the protective hypothesis: all classifiers suffer 16.6%-59.1% relative accuracy drops, comparable to neural-to-neural transfer. More surprisingly, we discover attack hierarchy reversal--contrary to patterns where iterative PGD dominates FGSM within neural networks, FGSM causes greater degradation than PGD in 100% of classical ML cases, suggesting iterative attacks overfit to surrogate-specific features that don't survive feature extraction. Block normalization provides partial but insufficient mitigation. These findings demonstrate that adversarial vulnerability is not an artifact of end-to-end differentiability but a fundamental property of image classification systems, with implications for security-critical deployments across computational paradigms.
Authors: Filippo Portera
Abstract: We propose a new convex loss for SVMs, both for the binary classification and for the regression models. Therefore, we show the mathematical derivation of the dual problems and we experiment them with several small data-sets. The minimal dimension of those data-sets is due to the difficult scalability of the SVM method to bigger instances. This preliminary study should prove that using pattern correlations inside the loss function could enhance the generalisation performances. Coherently, results show that generalisation measures are never worse than the standard losses and several times they are better. In our opinion, it should be considered a careful study of this loss, coupled with shallow and deep neural networks. In fact, we present some novel results obtained with those architectures.
Authors: Thuy Phuong Vu, Mai Viet Hoang Do, Minhhuy Le, Dinh-Cuong Hoang, Phan Xuan Tan
Abstract: Controlling memorization in diffusion models is critical for applications that require generated data to closely match the training distribution. Existing approaches mainly focus on data centric or model centric modifications, treating the diffusion model as an isolated predictor. In this paper, we study memorization in diffusion models from a denoising centric perspective. We show that uniform timestep sampling leads to unequal learning contributions across denoising steps due to differences in signal to noise ratio, which biases training toward memorization. To address this, we propose a timestep sampling strategy that explicitly controls where learning occurs along the denoising trajectory. By adjusting the width of the confidence interval, our method provides direct control over the memorization generalization trade off. Experiments on image and 1D signal generation tasks demonstrate that shifting learning emphasis toward later denoising steps consistently reduces memorization and improves distributional alignment with training data, validating the generality and effectiveness of our approach.
Authors: Minghao Yang, Ren Togo, Guang Li, Takahiro Ogawa, Miki Haseyama
Abstract: Mixture-of-Experts (MoE) models scale neural networks by conditionally activating a small subset of experts, where the router plays a central role in determining expert specialization and overall model performance. However, many modern MoE systems still adopt linear routers in raw high-dimensional representation spaces, where representation mismatch, angular concentration, and scale-sensitive scoring can jointly undermine routing discriminability and stable expert specialization. In this work, we propose Low-rank \& Lipschitz-controlled Routing (L2R), a unified routing framework that reshapes both the routing space and scoring geometry. L2R performs expert assignment in a shared low-rank latent routing space and introduces Saturated Inner-Product Scoring (SIPS) to explicitly control the Lipschitz behavior of routing functions, yielding smoother and more stable routing geometry. In addition, L2R incorporates a parameter-efficient multi-anchor routing mechanism to enhance expert expressiveness. Extensive experiments on a large-scale language MoE model and a vision MoE setting on ImageNet demonstrate that L2R consistently improves routing stability, expert specialization, and overall model performance.
Authors: Yupei Yang, Lin Yang, Wanxi Deng, Lin Qu, Fan Feng, Biwei Huang, Shikui Tu, Lei Xu
Abstract: A reliable reward model is essential for aligning large language models with human preferences through reinforcement learning from human feedback. However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to reward hacking, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model's contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.
Authors: Chendong Song, Meixuan Wang, Hang Zhou, Hong Liang, Yuan Lyu, Zixi Chen, Yuwei Fan, Zijie Zhou
Abstract: Attention-FFN disaggregation (AFD) is an emerging architecture for LLM decoding that separates state-heavy, KV-cache-dominated Attention computation from stateless, compute-intensive FFN computation, connected by per-step communication. While AFD enables independent scaling of memory and compute resources, its performance is highly sensitive to the Attention/FFN provisioning ratio: mis-sizing induces step-level blocking and costly device idle time. We develop a tractable analytical framework for sizing AFD bundles in an $r$A-$1$F topology, where the key difficulty is that Attention-side work is nonstationary-token context grows and requests are continuously replenished with random lengths-while FFN work is stable given the aggregated batch. Using a probabilistic workload model, we derive closed-form rules for the optimal A/F ratio that maximize average throughput per instance across the system. A trace-calibrated AFD simulator validates the theory: across workloads, the theoretical optimal A/F ratio matches the simulation-optimal within 10%, and consistently reduces idle time.
Authors: Joshua Hang Sai Ip, Georgios Makrygiorgos, Ali Mesbah
Abstract: Bayesian Optimization (BO) is a principled approach for optimizing expensive black-box functions, with Expected Improvement (EI) being one of the most widely used acquisition functions. Despite its empirical success, EI is known to be overly exploitative and can converge to suboptimal stationary points. We propose Expected Improvement via Gradient Norms (EI-GN), a novel acquisition function that applies the improvement principle to a gradient-aware auxiliary objective, thereby promoting sampling in regions that are both high-performing and approaching first-order stationarity. EI-GN relies on gradient observations used to learn gradient-enhanced surrogate models that enable principled gradient inference from function evaluations. We derive a tractable closed-form expression for EI-GN that allows efficient optimization and show that the proposed acquisition is consistent with the improvement-based acquisition framework. Empirical evaluations on standard BO benchmarks demonstrate that EI-GN yields consistent improvements against standard baselines. We further demonstrate applicability of EI-GN to control policy learning problems.
Authors: Luan Pham
Abstract: Failures in complex systems demand rapid Root Cause Analysis (RCA) to prevent cascading damage. Existing RCA methods that operate without dependency graph typically assume that the root cause having the highest anomaly score. This assumption fails when faults propagate, as a small delay at the root cause can accumulate into a much larger anomaly downstream. In this paper, we propose PRISM, a simple and efficient framework for RCA when the dependency graph is absent. We formulate a class of component-based systems under which PRISM performs RCA with theoretical guarantees. On 735 failures across 9 real-world datasets, PRISM achieves 68% Top-1 accuracy, a 258% improvement over the best baseline, while requiring only 8ms per diagnosis.
Authors: Antonio \'Alvarez-L\'opez, Borjan Geshkovski, Dom\`enec Ruiz-Balet
Abstract: The forward pass of a Transformer can be seen as an interacting particle system on the unit sphere: time plays the role of layers, particles that of token embeddings, and the unit sphere idealizes layer normalization. In some weight settings the system can even be seen as a gradient flow for an explicit energy, and one can make sense of the infinite context length (mean-field) limit thanks to Wasserstein gradient flows. In this paper we study the effect of the perceptron block in this setting, and show that critical points are generically atomic and localized on subsets of the sphere.
Authors: Yinlin Zhu, Di Wu, Xianzhi Zhang, Yuming Ai, Xunkai Li, Miao Hu, Guocong Quan
Abstract: Recent studies of federated graph foundational models (FedGFMs) break the idealized and untenable assumption of having centralized data storage to train graph foundation models, and accommodate the reality of distributed, privacy-restricted data silos. Despite their simplicity and intuition, existing studies that project aligned generalizable knowledge onto a discrete token space via vector-quantized backbones suffer from irreversible knowledge loss during the quantization process. In this context, we argue that reconciling the semantic-structural orthogonality and integrity between pre-trained language models (PLMs) and graph neural networks (GNNs) is paramount for developing effective FedGFMs while simultaneously mitigating the severe data heterogeneity and communication constraints inherent in distributed, resource-limited environments. To address these issues, we propose FedGALA (Federated Graph And Language Alignment), a framework that resolves graph-based semantic-structural orthogonality and integrity in federated settings by employing unsupervised contrastive learning to align GNNs and frozen PLMs within a continuous embedding space, thereby capturing robust, transferable general knowledge. Subsequently, FedGALA leverages a communication-efficient prompt tuning mechanism to steer these pre-aligned encoders and frozen PLMs, facilitating effective adaptation to diverse downstream tasks while circumventing the prohibitive overhead of full-parameter fine-tuning. The comprehensive experiments validate that FedGALA outperforms all competitive baselines across multi-domain datasets on multiple tasks with up to 14.37% performance improvement.
Authors: Tianhao Zhang, Shusen Ma, Yu Kang, Yun-Bo Zhao
Abstract: Multivariate time-series forecasting, as a typical problem in the field of time series prediction, has a wide range of applications in weather forecasting, traffic flow prediction, and other scenarios. However, existing works do not effectively consider the impact of extraneous variables on the prediction of the target variable. On the other hand, they fail to fully extract complex sequence information based on various time patterns of the sequences. To address these drawbacks, we propose a DA-SPS model, which adopts different modules for feature extraction based on the information characteristics of different variables. DA-SPS mainly consists of two stages: the target variable processing stage (TVPS) and the extraneous variables processing stage (EVPS). In TVPS, the model first uses Singular Spectrum Analysis (SSA) to process the target variable sequence and then uses Long Short-Term Memory (LSTM) and P-Conv-LSTM which deploys a patching strategy to extract features from trend and seasonality components, respectively. In EVPS, the model filters extraneous variables that have a strong correlation with the target variate by using Spearman correlation analysis and further analyses them using the L-Attention module which consists of LSTM and attention mechanism. Finally, the results obtained by TVPS and EVPS are combined through weighted summation and linear mapping to produce the final prediction. The results on four public datasets demonstrate that the DA-SPS model outperforms existing state-of-the-art methods. Additionally, its performance in real-world scenarios is further validated using a private dataset collected by ourselves, which contains the test items' information on laptop motherboards.
Authors: Hui Ma, Qingzhong Li, Jin Wang, Jie Wu, Shaoyu Dou, Li Feng, Xinjun Pei
Abstract: Network traffic forecasting plays a crucial role in intelligent network operations, but existing techniques often perform poorly when faced with limited data. Additionally, multi-task learning methods struggle with task imbalance and negative transfer, especially when modeling various service types. To overcome these challenges, we propose Sim-MSTNet, a multi-task spatiotemporal network traffic forecasting model based on the sim2real approach. Our method leverages a simulator to generate synthetic data, effectively addressing the issue of poor generalization caused by data scarcity. By employing a domain randomization technique, we reduce the distributional gap between synthetic and real data through bi-level optimization of both sample weighting and model training. Moreover, Sim-MSTNet incorporates attention-based mechanisms to selectively share knowledge between tasks and applies dynamic loss weighting to balance task objectives. Extensive experiments on two open-source datasets show that Sim-MSTNet consistently outperforms state-of-the-art baselines, achieving enhanced accuracy and generalization.
Authors: Rui Zhang, Jianwei Niu, Xuefeng Liu, Shaojie Tang, Jing Yuan
Abstract: The Job-Shop Scheduling Problem (JSSP), under various forms of manufacturing uncertainty, has recently attracted considerable research attention. Most existing studies focus on parameter uncertainty, such as variable processing times, and typically adopt the actor-critic framework. In this paper, we explore a different but prevalent form of uncertainty in JSSP: structural uncertainty. Structural uncertainty arises when a job may follow one of several routing paths, and the selection is determined not by policy, but by situational factors (e.g., the quality of intermediate products) that cannot be known in advance. Existing methods struggle to address this challenge due to incorrect credit assignment: a high-quality action may be unfairly penalized if it is followed by a time-consuming path. To address this problem, we propose a novel method named UP-AAC. In contrast to conventional actor-critic methods, UP-AAC employs an asymmetric architecture. While its actor receives a standard stochastic state, the critic is crucially provided with a deterministic state reconstructed in hindsight. This design allows the critic to learn a more accurate value function, which in turn provides a lower-variance policy gradient to the actor, leading to more stable learning. In addition, we design an attention-based Uncertainty Perception Model (UPM) to enhance the actor's scheduling decisions. Extensive experiments demonstrate that our method outperforms existing approaches in reducing makespan on benchmark instances.
Authors: Minjae Cho, Huy Trong Tran
Abstract: Exploration is essential in reinforcement learning as an agent relies on trial and error to learn an optimal policy. However, when rewards are sparse, naive exploration strategies, like noise injection, are often insufficient. Intrinsic rewards can also provide principled guidance for exploration by, for example, combining them with extrinsic rewards to optimize a policy or using them to train subpolicies for hierarchical learning. However, the former approach suffers from unstable credit assignment, while the latter exhibits sample inefficiency and sub-optimality. We propose a policy optimization framework that leverages multiple intrinsic rewards to directly optimize a policy for an extrinsic reward without pretraining subpolicies. Our algorithm -- intrinsic reward policy optimization (IRPO) -- achieves this by using a surrogate policy gradient that provides a more informative learning signal than the true gradient in sparse-reward environments. We demonstrate that IRPO improves performance and sample efficiency relative to baselines in discrete and continuous environments, and formally analyze the optimization problem solved by IRPO. Our code is available at https://github.com/Mgineer117/IRPO.
Authors: Qian Wan, Ziao Xu, Luona Wei, Xiaoxuan Shen, Jianwen Sun
Abstract: Large Reasoning Models (LRMs) achieve explicit chain-of-thought expansion by imitating deep thinking behaviors of humans, demonstrating excellent performance in complex task scenarios. However, the deep-thinking mode often leads to unnecessarily lengthy reasoning and resource inefficiency when handling simple tasks. This overthinking phenomenon may arise from the generation preference triggered by the reward function during post-training. Existing research attempts to mitigate overthinking from the perspective of prompt design or model training, but generally underestimates the importance of task difficulty awareness, which makes it difficult for LRMs to effectively allocate reasoning resources. In this paper, we propose Difficulty-aware Policy Optimization (DiPO), a reinforcement learning-based LRM training framework. DiPO encourages LRM to spontaneously model task complexity, and integrates them into reinforcement learning framework to adjust the generation preferences introduced by post-training. A difficulty modeling method based on model self-reasoning is proposed, which significantly reduces the dependence on manual annotation and formalize task complexity. We further develop a difficulty-signal-enhanced reward function that incorporates a penalty for lengthy reasoning while considering reasoning performance and output format. Experimental results indicate that DiPO enables the model to spontaneously adjust inference overhead, significantly reducing redundant tokens without losing performance due to thought compression.
Authors: Qing Jin, Chaoyang Wang
Abstract: Recent advances in diffusion and flow matching models have highlighted a shift in the preferred prediction target -- moving from noise ($\varepsilon$) and velocity (v) to direct data (x) prediction -- particularly in high-dimensional settings. However, a formal explanation of why the optimal target depends on the specific properties of the data remains elusive. In this work, we provide a theoretical framework based on a generalized prediction formulation that accommodates arbitrary output targets, of which $\varepsilon$-, v-, and x-prediction are special cases. We derive the analytical relationship between data's geometry and the optimal prediction target, offering a rigorous justification for why x-prediction becomes superior when the ambient dimension significantly exceeds the data's intrinsic dimension. Furthermore, while our theory identifies dimensionality as the governing factor for the optimal prediction target, the intrinsic dimension of manifold-bound data is typically intractable to estimate in practice. To bridge this gap, we propose k-Diff, a framework that employs a data-driven approach to learn the optimal prediction parameter k directly from data, bypassing the need for explicit dimension estimation. Extensive experiments in both latent-space and pixel-space image generation demonstrate that k-Diff consistently outperforms fixed-target baselines across varying architectures and data scales, providing a principled and automated approach to enhancing generative performance.
Authors: Zihao Huang, Jundong Zhou, Xingwei Qu, Qiyang Min, Ge Zhang
Abstract: Large language models allocate uniform computation across all tokens, ignoring that some sequences are trivially predictable while others require deep reasoning. We introduce ConceptMoE, which dynamically merges semantically similar tokens into concept representations, performing implicit token-level compute allocation. A learnable chunk module identifies optimal boundaries by measuring inter-token similarity, compressing sequences by a target ratio $R$ before they enter the compute-intensive concept model. Crucially, the MoE architecture enables controlled evaluation: we reallocate saved computation to match baseline activated FLOPs (excluding attention map computation) and total parameters, isolating genuine architectural benefits. Under these conditions, ConceptMoE consistently outperforms standard MoE across language and vision-language tasks, achieving +0.9 points on language pretraining, +2.3 points on long context understanding, and +0.6 points on multimodal benchmarks. When converting pretrained MoE during continual training with layer looping, gains reach +5.5 points, demonstrating practical applicability. Beyond performance, ConceptMoE reduces attention computation by up to $R^2\times$ and KV cache by $R\times$. At $R=2$, empirical measurements show prefill speedups reaching 175\% and decoding speedups up to 117\% on long sequences. The minimal architectural modifications enable straightforward integration into existing MoE, demonstrating that adaptive concept-level processing fundamentally improves both effectiveness and efficiency of large language models.
Authors: Anderson de Andrade, Alon Harell, Ivan V. Baji\'c
Abstract: Many computer vision tasks share substantial overlapping information, yet conventional codecs tend to ignore this, leading to redundant and inefficient representations. The Gray-Wyner network, a classical concept from information theory, offers a principled framework for separating common and task-specific information. Inspired by this idea, we develop a learnable three-channel codec that disentangles shared information from task-specific details across multiple vision tasks. We characterize the limits of this approach through the notion of lossy common information, and propose an optimization objective that balances inherent tradeoffs in learning such representations. Through comparisons of three codec architectures on two-task scenarios spanning six vision benchmarks, we demonstrate that our approach substantially reduces redundancy and consistently outperforms independent coding. These results highlight the practical value of revisiting Gray-Wyner theory in modern machine learning contexts, bridging classic information theory with task-driven representation learning.
Authors: Hang Ni, Weijia Zhang, Fei Wang, Zezhi Shao, Hao Liu
Abstract: Advances in multi-modal large language models (MLLMs) have inspired time series understanding and reasoning tasks, that enable natural language querying over time series, producing textual analyses of complex temporal dynamics. Recent attempts hybridize numerical time series with their visualized plots, facilitating precise value reasoning and visual structure comprehension for comprehensive time series understanding of MLLMs. However, effective cross-modal integration remains challenging due to fine-grained temporal misalignment across modalities and severe entanglement between shared and modality-specific semantics, which hinder localized interpretation and complementary reasoning. To address these issues, we propose MADI, a multi-modal LLM enhanced with fine-grained alignment and disentangled interaction, featuring (1) Patch-level Alignment, which enforces physically grounded fine-grained correspondence across heterogeneous modalities, (2) Discrete Disentangled Interaction, which separates modality-common semantics into compact discrete latents and adaptively synergizes the purified modality-unique information, and (3) Critical-token Highlighting, which emphasizes informative, query-relevant signals for robust reasoning. Experiments on synthetic and real-world benchmarks show that MADI consistently outperforms general-purpose LLMs and time-series-specialized MLLMs.
Authors: Yu Sun, Yaqiong Liu, Nan Cheng, Jiayuan Li, Zihan Jia, Xialin Du, Mugen Peng
Abstract: Driven by the evolution toward 6G and AI-native edge intelligence, network operations increasingly require predictive and risk-aware adaptation under stringent computation and latency constraints. Network Traffic Matrix (TM), which characterizes flow volumes between nodes, is a fundamental signal for proactive traffic engineering. However, accurate TM forecasting remains challenging due to the stochastic, non-linear, and bursty nature of network dynamics. Existing discriminative models often suffer from over-smoothing and provide limited uncertainty awareness, leading to poor fidelity under extreme bursts. To address these limitations, we propose LEAD, a Large Language Model (LLM)-Enhanced Adapter-based conditional Diffusion model. First, LEAD adopts a "Traffic-to-Image" paradigm to transform traffic matrices into RGB images, enabling global dependency modeling via vision backbones. Then, we design a "Frozen LLM with Trainable Adapter" model, which efficiently captures temporal semantics with limited computational cost. Moreover, we propose a Dual-Conditioning Strategy to precisely guide a diffusion model to generate complex, dynamic network traffic matrices. Experiments on the Abilene and GEANT datasets demonstrate that LEAD outperforms all baselines. On the Abilene dataset, LEAD attains a remarkable 45.2% reduction in RMSE against the best baseline, with the error margin rising only marginally from 0.1098 at one-step to 0.1134 at 20-step predictions. Meanwhile, on the GEANT dataset, LEAD achieves a 0.0258 RMSE at 20-step prediction horizon which is 27.3% lower than the best baseline.
Authors: Francesco Zola, Lucia Mu\~noz, Andrea Venturi, Amaia Gil
Abstract: Illicit financial activities such as money laundering often manifest through recurrent topological patterns in transaction networks. Detecting these patterns automatically remains challenging due to the scarcity of labeled real-world data and strict privacy constraints. To address this, we investigate whether Graph Autoencoders (GAEs) can effectively learn and distinguish topological patterns that mimic money laundering operations when trained on synthetic data. The analysis consists of two phases: (i) data generation, where synthetic samples are created for seven well-known illicit activity patterns using parametrized generators that preserve structural consistency while introducing realistic variability; and (ii) model training and validation, where separate GAEs are trained on each pattern without explicit labels, relying solely on reconstruction error as an indicator of learned structure. We compare three GAE implementations based on three distinct convolutional layers: Graph Convolutional (GAE-GCN), GraphSAGE (GAE-SAGE), and Graph Attention Network (GAE-GAT). Experimental results show that GAE-GCN achieves the most consistent reconstruction performance across patterns, while GAE-SAGE and GAE-GAT exhibit competitive results only in few specific patterns. These findings suggest that graph-based representation learning on synthetic data provides a viable path toward developing AI-driven tools for detecting illicit behaviors, overcoming the limitations of financial datasets.
Authors: Yu Xie, Xing Kai Ren, Ying Qi, Hu Yao
Abstract: While works such as OneRec have validated the scaling laws of Large Language Models (LLMs) in recommender systems, they rely on a cumbersome separate vocabulary. This dependency prevents the model architecture from reusing native LLM vocabularies, resulting in high maintenance costs and poor scalability. In response, we aim to efficiently reuse open-source LLM architectures without constructing a separate tokenization vocabulary. Furthermore, we identify that the optimization strategy of OneRec Gradient Bounded Policy Optimization (GBPO),suffers from a "Symmetric Conservatism" problem: its static gradient boundaries structurally suppress the update momentum required for cold-start items and fail to prevent diversity collapse in high-noise environments.To address this issue, we propose SAGE (Sequence-level Adaptive Gradient Evolution), a unified optimization framework tailored for list-wise generative recommendation. SAGE introduces two key innovations:(1) Sequence-level Signal Decoupling: By combining a geometric mean importance ratio with decoupled multi-objective advantages, we eliminate token-level variance and resolve the "Reward Collapse" problem. (2) Asymmetric Adaptive Dynamics: We construct a dynamic gradient manifold that applies a "Boost Factor" to high-potential cold start items to achieve super-linear updates and employs an "Entropy Aware Penalty" to break information cocoons. Theoretical analysis and empirical results demonstrate that SAGE effectively unblocks cold-start traffic and sustains recommendation diversity, all while retaining the numerical stability of GBPO.
Authors: Chengyu Du, Xintao Wang, Aili Chen, Weiyuan Li, Rui Xu, Junteng Liu, Zishan Huang, Rong Tian, Zijun Sun, Yuhao Li, Liheng Feng, Deming Ding, Pengyu Zhao, Yanghua Xiao
Abstract: LLM role-playing, i.e., using LLMs to simulate specific personas, has emerged as a key capability in various applications, such as companionship, content creation, and digital games. While current models effectively capture character tones and knowledge, simulating the inner thoughts behind their behaviors remains a challenge. Towards cognitive simulation in LLM role-play, previous efforts mainly suffer from two deficiencies: data with high-quality reasoning traces, and reliable reward signals aligned with human preferences. In this paper, we propose HER, a unified framework for cognitive-level persona simulation. HER introduces dual-layer thinking, which distinguishes characters' first-person thinking from LLMs' third-person thinking. To bridge these gaps, we curate reasoning-augmented role-playing data via reverse engineering and construct human-aligned principles and reward models. Leveraging these resources, we train \method models based on Qwen3-32B via supervised and reinforcement learning. Extensive experiments validate the effectiveness of our approach. Notably, our models significantly outperform the Qwen3-32B baseline, achieving a 30.26 improvement on the CoSER benchmark and a 14.97 gain on the Minimax Role-Play Bench. Our datasets, principles, and models will be released to facilitate future research.
Authors: Albert Tseng, Christopher De Sa
Abstract: Modern sparse language models typically achieve sparsity through Mixture-of-Experts (MoE) layers, which dynamically route tokens to dense MLP "experts." However, dynamic hard routing has a number of drawbacks, such as potentially poor hardware efficiency and needing auxiliary losses for stable training. In contrast, the tokenizer embedding table, which is natively sparse, largely avoids these issues by selecting a single embedding per token at the cost of not having contextual information. In this work, we introduce the Large Lookup Layer (L$^3$), which unlocks a new axis of sparsity by generalizing embedding tables to model decoder layers. L$^3$ layers use static token-based routing to aggregate a set of learned embeddings per token in a context-dependent way, allowing the model to efficiently balance memory and compute by caching information in embeddings. L$^3$ has two main components: (1) a systems-friendly architecture that allows for fast training and CPU-offloaded inference with no overhead, and (2) an information-theoretic embedding allocation algorithm that effectively balances speed and quality. We empirically test L$^3$ by training transformers with up to 2.6B active parameters and find that L$^3$ strongly outperforms both dense models and iso-sparse MoEs in both language modeling and downstream tasks.
Authors: Shihao Shao, Cong Fang, Zhouchen Lin, Dacheng Tao
Abstract: We study partial-feedback online learning, where each instance admits a set of correct labels, but the learner only observes one correct label per round; any prediction within the correct set is counted as correct. This model captures settings such as language generation, where multiple responses may be valid but data provide only a single reference. We give a near-complete characterization of minimax regret for both deterministic and randomized learners in the set-realizable regime, i.e., in the regime where sublinear regret is generally attainable. For deterministic learners, we introduce the Partial-Feedback Littlestone dimension (PFLdim) and show it precisely governs learnability and minimax regret; technically, PFLdim cannot be defined via the standard version space, requiring a new collection version space viewpoint and an auxiliary dimension used only in the proof. We further develop the Partial-Feedback Measure Shattering dimension (PMSdim) to obtain tight bounds for randomized learners. We identify broad conditions ensuring inseparability between deterministic and randomized learnability (e.g., finite Helly number or nested-inclusion label structure), and extend the argument to set-valued online learning, resolving an open question of Raman et al. [2024b]. Finally, we show a sharp separation from weaker realistic and agnostic variants: outside set realizability, the problem can become information-theoretically intractable, with linear regret possible even for $|H|=2$. This highlights the need for fundamentally new, noise-sensitive complexity measures to meaningfully characterize learnability beyond set realizability.
Authors: Guillaume Lauga (LJAD)
Abstract: Block-coordinate descent (BCD) is the method of choice to solve numerous large scale optimization problems, however their theoretical study for non-convex optimization, has received less attention. In this paper, we present a new block-coordinate descent (BCD) framework to tackle non-convex composite optimization problems, ensuring decrease of the objective function and convergence to a solution. This framework is general enough to include variable metric proximal gradient updates, proximal Newton updates, and alternated minimization updates. This generality allows to encompass three versions of the most used solvers in the sparse precision matrix estimation problem, deemed Graphical Lasso: graphical ISTA, Primal GLasso, and QUIC. We demonstrate the value of this new framework on non-convex sparse precision matrix estimation problems, providing convergence guarantees and up to a $100$-fold reduction in the number of iterations required to reach state-of-the-art estimation quality.
Authors: Ruicheng Ao, Hongyu Chen, Haoyang Liu, David Simchi-Levi, Will Wei Sun
Abstract: We study semi-supervised stochastic optimization when labeled data is scarce but predictions from pre-trained models are available. PPI and SVRG both reduce variance through control variates -- PPI uses predictions, SVRG uses reference gradients. We show they are mathematically equivalent and develop PPI-SVRG, which combines both. Our convergence bound decomposes into the standard SVRG rate plus an error floor from prediction uncertainty. The rate depends only on loss geometry; predictions affect only the neighborhood size. When predictions are perfect, we recover SVRG exactly. When predictions degrade, convergence remains stable but reaches a larger neighborhood. Experiments confirm the theory: PPI-SVRG reduces MSE by 43--52\% under label scarcity on mean estimation benchmarks and improves test accuracy by 2.7--2.9 percentage points on MNIST with only 10\% labeled data.
Authors: Ruicheng Ao, Hongyu Chen, Siyang Gao, Hanwei Li, David Simchi-Levi
Abstract: We study fixed-confidence best-arm identification (BAI) where a cheap but potentially biased proxy (e.g., LLM judge) is available for every sample, while an expensive ground-truth label can only be acquired selectively when using a human for auditing. Unlike classical multi-fidelity BAI, the proxy is biased (arm- and context-dependent) and ground truth is selectively observed. Consequently, standard multi-fidelity methods can mis-select the best arm, and uniform auditing, though accurate, wastes scarce resources and is inefficient. We prove that without bias correction and propensity adjustment, mis-selection probability may not vanish (even with unlimited proxy data). We then develop an estimator for the mean of each arm that combines proxy scores with inverse-propensity-weighted residuals and form anytime-valid confidence sequences for that estimator. Based on the estimator and confidence sequence, we propose an algorithm that adaptively selects and audits arms. The algorithm concentrates audits on unreliable contexts and close arms and we prove that a plug-in Neyman rule achieves near-oracle audit efficiency. Numerical experiments confirm the theoretical guarantees and demonstrate the superior empirical performance of the proposed algorithm.
Authors: Xiuyu Li, Jinkai Zhang, Mingyang Yi, Yu Li, Longqiang Wang, Yue Wang, Ju Fan
Abstract: Reinforcement Learning (RL) post-training alignment for language models is effective, but also costly and unstable in practice, owing to its complicated training process. To address this, we propose a training-free inference method to sample directly from the optimal RL policy. The transition probability applied to Masked Language Modeling (MLM) consists of a reference policy model and an energy term. Based on this, our algorithm, Energy-Guided Test-Time Scaling (ETS), estimates the key energy term via online Monte Carlo, with a provable convergence rate. Moreover, to ensure practical efficiency, ETS leverages modern acceleration frameworks alongside tailored importance sampling estimators, substantially reducing inference latency while provably preserving sampling quality. Experiments on MLM (including autoregressive models and diffusion language models) across reasoning, coding, and science benchmarks show that our ETS consistently improves generation quality, validating its effectiveness and design.
Authors: Tim Tomov, Dominik Fuchsgruber, Stephan G\"unnemann
Abstract: In many applications of LLMs, natural language responses often have an underlying structure such as representing discrete labels, numerical values, or graphs. Yet, existing decoding and uncertainty estimation methods operate only in language space and largely disregard structural information. We address this by modeling LLM outputs directly in a task-dependent latent structure. By equipping this structure with a dissimilarity measure, we can compute Bayes-optimal responses. These are not selected from sampled generations but are newly synthesized by combining individual responses in the latent space. Across different tasks, Bayes-optimal responses consistently outperform standard decoding methods like beam search. Moreover, quantifying uncertainty via the induced Bayesian risk captures variations in terms of the latent structure and improves alignment with output quality and correctness. Our decision-theoretic framework is applicable to any problem that admits a latent response structure and enables reliable task-aware LLM predictions.
Authors: Eloi Campagne (CB), Yvenn Amara-Ouali (LMO), Yannig Goude (LMO), Mathilde Mougeot (CB, ENSIIE, ENS Paris Saclay), Argyris Kalogeratos (CB, ENS Paris Saclay)
Abstract: Many-Task Learning refers to the setting where a large number of related tasks need to be learned, the exact relationships between tasks are not known. We introduce the Cascaded Transfer Learning, a novel many-task transfer learning paradigm where information (e.g. model parameters) cascades hierarchically through tasks that are learned by individual models of the same class, while respecting given budget constraints. The cascade is organized as a rooted tree that specifies the order in which tasks are learned and refined. We design a cascaded transfer mechanism deployed over a minimum spanning tree structure that connects the tasks according to a suitable distance measure, and allocates the available training budget along its branches. Experiments on synthetic and real many-task settings show that the resulting method enables more accurate and cost effective adaptation across large task collections compared to alternative approaches.
Authors: Chi-Sheng Chen, En-Jui Kuo, Guan-Ying Chen, Xinyu Zhang, Fan Zhang
Abstract: Spatial covariance matrices of EEG signals are Symmetric Positive Definite (SPD) and lie on a Riemannian manifold, yet the theoretical connection between embedding geometry and optimization dynamics remains unexplored. We provide a formal analysis linking embedding choice to gradient conditioning and numerical stability for SPD manifolds, establishing three theoretical results: (1) BWSPD's $\sqrt{\kappa}$ gradient conditioning (vs $\kappa$ for Log-Euclidean) via Daleckii-Kre\u{\i}n matrices provides better gradient conditioning on high-dimensional inputs ($d \geq 22$), with this advantage reducing on low-dimensional inputs ($d \leq 8$) where eigendecomposition overhead dominates; (2) Embedding-Space Batch Normalization (BN-Embed) approximates Riemannian normalization up to $O(\varepsilon^2)$ error, yielding $+26\%$ accuracy on 56-channel ERP data but negligible effect on 8-channel SSVEP data, matching the channel-count-dependent prediction; (3) bi-Lipschitz bounds prove BWSPD tokens preserve manifold distances with distortion governed solely by the condition ratio $\kappa$. We validate these predictions via a unified Transformer framework comparing BWSPD, Log-Euclidean, and Euclidean embeddings within identical architecture across 1,500+ runs on three EEG paradigms (motor imagery, ERP, SSVEP; 36 subjects). Our Log-Euclidean Transformer achieves state-of-the-art performance on all datasets, substantially outperforming classical Riemannian classifiers and recent SPD baselines, while BWSPD offers competitive accuracy with similar training time.
Authors: Sagi Meir, Tommer D. Keidar, Noam Levi, Shlomi Reuveni, Barak Hirshberg
Abstract: The performance of large language models (LLMs) on verifiable tasks is usually measured by pass@k, the probability of answering a question correctly at least once in k trials. At a fixed budget, a more suitable metric is coverage@cost, the average number of unique questions answered as a function of the total number of attempts. We connect the two metrics and show that the empirically-observed power-law behavior in pass@k leads to a sublinear growth of the coverage@cost (diminishing returns). To solve this problem, we propose Reset-and-Discard (ReD), a query method of LLMs that increases coverage@cost for any given budget, regardless of the pass@k form. Moreover, given a pass@k, we can quantitatively predict the savings in the total number of attempts using ReD. If pass@k is not available for the model, ReD can infer its power-law exponent. Experiments on three LLMs using HumanEval demonstrate that ReD substantially reduces the required attempts, tokens, and USD cost to reach a desired coverage, while also offering an efficient way to measure inference power-laws.
Authors: Bang Giang Le, Viet Cuong Ta
Abstract: To promote cooperation in Multi-Agent Reinforcement Learning, the reward signals of all agents can be aggregated together, forming global rewards that are commonly known as the fully cooperative setting. However, global rewards are usually noisy because they contain the contributions of all agents, which have to be resolved in the credit assignment process. On the other hand, using local reward benefits from faster learning due to the separation of agents' contributions, but can be suboptimal as agents myopically optimize their own reward while disregarding the global optimality. In this work, we propose a method that combines the merits of both approaches. By using a graph of interaction between agents, our method discerns the individual agent contribution in a more fine-grained manner than a global reward, while alleviating the cooperation problem with agents' local reward. We also introduce a practical approach for approximating such a graph. Our experiments demonstrate the flexibility of the approach, enabling improvements over the traditional local and global reward settings.
Authors: Robert van der Klis, Ricardo Ch\'avez Torres, Max van Spengler, Yuhui Ding, Thomas Hofmann, Pascal Mettes
Abstract: Hyperbolic space is quickly gaining traction as a promising geometry for hierarchical and robust representation learning. A core open challenge is the development of a mathematical formulation of hyperbolic neural networks that is both efficient and captures the key properties of hyperbolic space. The Lorentz model of hyperbolic space has been shown to enable both fast forward and backward propagation. However, we prove that, with the current formulation of Lorentz linear layers, the hyperbolic norms of the outputs scale logarithmically with the number of gradient descent steps, nullifying the key advantage of hyperbolic geometry. We propose a new Lorentz linear layer grounded in the well-known ``distance-to-hyperplane" formulation. We prove that our formulation results in the usual linear scaling of output hyperbolic norms with respect to the number of gradient descent steps. Our new formulation, together with further algorithmic efficiencies through Lorentzian activation functions and a new caching strategy results in neural networks fully abiding by hyperbolic geometry while simultaneously bridging the computation gap to Euclidean neural networks. Code available at: https://github.com/robertdvdk/hyperbolic-fully-connected.
URLs: https://github.com/robertdvdk/hyperbolic-fully-connected.
Authors: Lige Zhang, Ali Maatouk, Jialin Chen, Leandros Tassiulas, Rex Ying
Abstract: Real-world time series exhibit complex and evolving dynamics, making accurate forecasting extremely challenging. Recent multi-modal forecasting methods leverage textual information such as news reports to improve prediction, but most rely on token-level fusion that mixes temporal patches with language tokens in a shared embedding space. However, such fusion can be ill-suited when high-quality time-text pairs are scarce and when time series exhibit substantial variation in scale and characteristics, thus complicating cross-modal alignment. In parallel, Mixture-of-Experts (MoE) architectures have proven effective for both time series modeling and multi-modal learning, yet many existing MoE-based modality integration methods still depend on token-level fusion. To address this, we propose Expert Modulation, a new paradigm for multi-modal time series prediction that conditions both routing and expert computation on textual signals, enabling direct and efficient cross-modal control over expert behavior. Through comprehensive theoretical analysis and experiments, our proposed method demonstrates substantial improvements in multi-modal time series prediction. The current code is available at https://github.com/BruceZhangReve/MoME
Authors: Susu Hu, Qinghe Zeng, Nithya Bhasker, Jakob Nicolas Kather, Stefanie Speidel
Abstract: Predicting spatial gene expression from H&E histology offers a scalable and clinically accessible alternative to sequencing, but realizing clinical impact requires models that generalize across cancer types and capture biologically coherent signals. Prior work is often limited to per-cancer settings and variance-based evaluation, leaving functional relevance underexplored. We introduce HistoPrism, an efficient transformer-based architecture for pan-cancer prediction of gene expression from histology. To evaluate biological meaning, we introduce a pathway-level benchmark, shifting assessment from isolated gene-level variance to coherent functional pathways. HistoPrism not only surpasses prior state-of-the-art models on highly variable genes , but also more importantly, achieves substantial gains on pathway-level prediction, demonstrating its ability to recover biologically coherent transcriptomic patterns. With strong pan-cancer generalization and improved efficiency, HistoPrism establishes a new standard for clinically relevant transcriptomic modeling from routinely available histology.
Authors: Fanping Liu, Hua Yang, Jiasi Zou
Abstract: Standard deep learning relies on Backpropagation (BP), which is constrained by biologically implausible weight symmetry and suffers from significant gradient interference within dense representations. To mitigate these bottlenecks, we propose Selective Adaptive Learning (SAL), a training method that combines selective parameter activation with adaptive area partitioning. Specifically, SAL decomposes the parameter space into mutually exclusive, sample-dependent regions. This decoupling mitigates gradient interference across divergent semantic patterns and addresses explicit weight symmetry requirements through our refined feedback alignment. Empirically, SAL demonstrates competitive convergence rates, leading to improved classification performance across 10 standard benchmarks. Additionally, SAL achieves numerical consistency and competitive accuracy even in deep regimes (up to 128 layers) and large-scale models (up to 1B parameters). Our approach is loosely inspired by biological learning mechanisms, offering a plausible alternative that contributes to the study of scalable neural network training.
Authors: Antonio Almud\'evar, Alfonso Ortega
Abstract: Machine unlearning seeks to remove the influence of specific training data from a model, a need driven by privacy regulations and robustness concerns. Existing approaches typically modify model parameters, but such updates can be unstable, computationally costly, and limited by local approximations. We introduce Representation Unlearning, a framework that performs unlearning directly in the model's representation space. Instead of modifying model parameters, we learn a transformation over representations that imposes an information bottleneck: maximizing mutual information with retained data while suppressing information about data to be forgotten. We derive variational surrogates that make this objective tractable and show how they can be instantiated in two practical regimes: when both retain and forget data are available, and in a zero-shot setting where only forget data can be accessed. Experiments across several benchmarks demonstrate that Representation Unlearning achieves more reliable forgetting, better utility retention, and greater computational efficiency than parameter-centric baselines.
Authors: Yutao Jin, Yuang Tao, Junyong Zhai
Abstract: Causal Disentangled Representation Learning(CDRL) aims to learn and disentangle low dimensional representations and their underlying causal structure from observations. However, existing disentanglement methods rely on a standard mean-field approximation with a diagonal posterior covariance, which decorrelates all latent dimensions. Additionally, these methods often assume isotropic Gaussian priors for exogenous noise, failing to capture the complex, non-Gaussian statistical properties prevalent in real-world causal factors. Therefore, we propose FlexCausal, a novel CDRL framework based on a block-diagonal covariance VAE. FlexCausal utilizes a Factorized Flow-based Prior to realistically model the complex densities of exogenous noise, effectively decoupling the learning of causal mechanisms from distributional statistics. By integrating supervised alignment objectives with counterfactual consistency constraints, our framework ensures a precise structural correspondence between the learned latent subspaces and the ground-truth causal relations. Finally, we introduce a manifold-aware relative intervention strategy to ensure high-fidelity generation. Experimental results on both synthetic and real-world datasets demonstrate that FlexCausal significantly outperforms other methods.
Authors: Antonio Almud\'evar, Alfonso Ortega
Abstract: We present a unified framework for quantifying the similarity between representations through the lens of \textit{usable information}, offering a rigorous theoretical and empirical synthesis across three key dimensions. First, addressing functional similarity, we establish a formal link between stitching performance and conditional mutual information. We further reveal that stitching is inherently asymmetric, demonstrating that robust functional comparison necessitates a bidirectional analysis rather than a unidirectional mapping. Second, concerning representational similarity, we prove that reconstruction-based metrics and standard tools (e.g., CKA, RSA) act as estimators of usable information under specific constraints. Crucially, we show that similarity is relative to the capacity of the predictive family: representations that appear distinct to a rigid observer may be identical to a more expressive one. Third, we demonstrate that representational similarity is sufficient but not necessary for functional similarity. We unify these concepts through a task-granularity hierarchy: similarity on a complex task guarantees similarity on any coarser derivative, establishing representational similarity as the limit of maximum granularity: input reconstruction.
Authors: Neil Rathi, Alec Radford
Abstract: Current approaches to reducing undesired capabilities in language models are largely post hoc, and can thus be easily bypassed by adversaries. A natural alternative is to shape capabilities during pretraining itself. On the proxy task of removing medical capabilities, we show that the simple intervention of filtering pretraining data is highly effective, robust, and inexpensive at scale. Inspired by work on data attribution, we show that filtering tokens is more effective than filtering documents, achieving the same hit to undesired capabilities at a lower cost to benign ones. Training models spanning two orders of magnitude, we then demonstrate that filtering gets more effective with scale: for our largest models, token filtering leads to a 7000x compute slowdown on the forget domain. We also show that models trained with token filtering can still be aligned on the forget domain. Along the way, we introduce a methodology for labeling tokens with sparse autoencoders and distilling cheap, high-quality classifiers. We also demonstrate that filtering can be robust to noisy labels with sufficient pretraining compute.
Authors: Jinhao Li, Yuhao Sun, Zhiyuan Ma, Hao He, Xinche Zhang, Xing Chen, Jin Li, Sen Song
Abstract: Recurrent spiking neural networks (RSNNs) are a promising substrate for energy-efficient control policies, but training them for high-dimensional, long-horizon reinforcement learning remains challenging. Population-based, gradient-free optimization circumvents backpropagation through non-differentiable spike dynamics by estimating gradients. However, with finite populations, high variance of these estimates can induce harmful and overly aggressive update steps. Inspired by trust-region methods in reinforcement learning that constrain policy updates in distribution space, we propose \textbf{Signal-Adaptive Trust Regions (SATR)}, a distributional update rule that constrains relative change by bounding KL divergence normalized by an estimated signal energy. SATR automatically expands the trust region under strong signals and contracts it when updates are noise-dominated. We instantiate SATR for Bernoulli connectivity distributions, which have shown strong empirical performance for RSNN optimization. Across a suite of high-dimensional continuous-control benchmarks, SATR improves stability under limited populations and reaches competitive returns against strong baselines including PPO-LSTM. In addition, to make SATR practical at scale, we introduce a bitset implementation for binary spiking and binary weights, substantially reducing wall-clock training time and enabling fast RSNN policy search.
Authors: Mutian Yang, Zisen Zhan, Yutong Chen, Haolin Li, Kaiwen Wang, Kaili Zheng, Yuguang Wang, Qi Wang, Jiandong Gao, Ji Wu
Abstract: Catastrophic forgetting during knowledge injection severely undermines the continual learning capability of large language models (LLMs). Although existing methods attempt to mitigate this issue, they often lack a foundational theoretical explanation. We establish a gradient-based theoretical framework to explain catastrophic forgetting. We first prove that strongly negative gradient similarity is a fundamental cause of forgetting. We then use gradient similarity to identify two types of neurons: conflicting neurons that induce forgetting and account for 50%-75% of neurons, and collaborative neurons that mitigate forgetting and account for 25%-50%. Based on this analysis, we propose a knowledge injection method, Collaborative Neural Learning (CNL). By freezing conflicting neurons and updating only collaborative neurons, CNL theoretically eliminates catastrophic forgetting under an infinitesimal learning rate eta and an exactly known mastered set. Experiments on five LLMs, four datasets, and four optimizers show that CNL achieves zero forgetting in in-set settings and reduces forgetting by 59.1%-81.7% in out-of-set settings.
Authors: Morten Bl{\o}rstad, Herman Jangsett Mostein, Nello Blaser, Pekka Parviainen
Abstract: Deep learning models struggle with uncertainty estimation. Many approaches are either computationally infeasible or underestimate uncertainty. We investigate \textit{BatchEnsemble} as a general and scalable method for uncertainty estimation across both tabular and time series tasks. To extend BatchEnsemble to sequential modeling, we introduce GRUBE, a novel BatchEnsemble GRU cell. We compare the BatchEnsemble to Monte Carlo dropout and deep ensemble models. Our results show that BatchEnsemble matches the uncertainty estimation performance of deep ensembles, and clearly outperforms Monte Carlo dropout. GRUBE achieves similar or better performance in both prediction and uncertainty estimation. These findings show that BatchEnsemble and GRUBE achieve similar performance with fewer parameters and reduced training and inference time compared to traditional ensembles.
Authors: Tin Had\v{z}i Veljkovi\'c, Erik Bekkers, Michael Tiemann, Jan-Willem van de Meent
Abstract: Many learning problems require predicting sets of objects when the number of objects is not known beforehand. Examples include object detection, molecular modeling, and scientific inference tasks such as astrophysical source detection. Existing methods often rely on padded representations or must explicitly infer the set size, which often poses challenges. We present a novel strategy for addressing this challenge by casting prediction of variable-sized sets as a continuous inference problem. Our approach, CORDS (Continuous Representations of Discrete Structures), provides an invertible mapping that transforms a set of spatial objects into continuous fields: a density field that encodes object locations and count, and a feature field that carries their attributes over the same support. Because the mapping is invertible, models operate entirely in field space while remaining exactly decodable to discrete sets. We evaluate CORDS across molecular generation and regression, object detection, simulation-based inference, and a mathematical task involving recovery of local maxima, demonstrating robust handling of unknown set sizes with competitive accuracy.
Authors: Wentao Yu, Sheng Wan, Shuo Chen, Bo Han, Chen Gong
Abstract: Graph Federated Learning (GFL) enables distributed graph representation learning while protecting the privacy of graph data. However, GFL suffers from heterogeneity arising from diverse node features and structural topologies across multiple clients. To address both types of heterogeneity, we propose a novel graph Federated learning method via Semantic and Structural Alignment (FedSSA), which shares the knowledge of both node features and structural topologies. For node feature heterogeneity, we propose a novel variational model to infer class-wise node distributions, so that we can cluster clients based on inferred distributions and construct cluster-level representative distributions. We then minimize the divergence between local and cluster-level distributions to facilitate semantic knowledge sharing. For structural heterogeneity, we employ spectral Graph Neural Networks (GNNs) and propose a spectral energy measure to characterize structural information, so that we can cluster clients based on spectral energy and build cluster-level spectral GNNs. We then align the spectral characteristics of local spectral GNNs with those of cluster-level spectral GNNs to enable structural knowledge sharing. Experiments on six homophilic and five heterophilic graph datasets under both non-overlapping and overlapping partitioning settings demonstrate that FedSSA consistently outperforms eleven state-of-the-art methods.
Authors: Xiaotong Ji, Rasul Tutunov, Matthieu Zimmer, Haitham Bou Ammar
Abstract: Reinforcement learning (RL) post-training is a dominant approach for improving the reasoning performance of large language models (LLMs), yet growing evidence suggests that its gains arise primarily from distribution sharpening rather than the acquisition of new capabilities. Recent work has shown that sampling from the power distribution of LLMs using Markov chain Monte Carlo (MCMC) can recover performance comparable to RL post-training without relying on external rewards; however, the high computational cost of MCMC makes such approaches impractical for widespread adoption. In this work, we propose a theoretically grounded alternative that eliminates the need for iterative MCMC. We derive a novel formulation showing that the global power distribution can be approximated by a token-level scaled low-temperature one, where the scaling factor captures future trajectory quality. Leveraging this insight, we introduce a training-free and verifier-free algorithm that sharpens the base model's generative distribution autoregressively. Empirically, we evaluate our method on math, QA, and code tasks across four LLMs, and show that our method matches or surpasses one-shot GRPO without relying on any external rewards, while reducing inference latency by over 10x compared to MCMC-based sampling.
Authors: Hoyeon Chang, B\'alint Mucs\'anyi, Seong Joon Oh
Abstract: Neural networks adapt through first-order parameter updates, yet it remains unclear whether such updates preserve logical coherence. We investigate the geometric limits of the Linear Propagation Assumption (LPA), the premise that local updates coherently propagate to logical consequences. To formalize this, we adopt relation algebra and study three core operations on relations: negation flips truth values, converse swaps argument order, and composition chains relations. For negation and converse, we prove that guaranteeing direction-agnostic first-order propagation necessitates a tensor factorization separating entity-pair context from relation content. However, for composition, we identify a fundamental obstruction. We show that composition reduces to conjunction, and prove that any conjunction well-defined on linear features must be bilinear. Since bilinearity is incompatible with negation, this forces the feature map to collapse. These results suggest that failures in knowledge editing, the reversal curse, and multi-hop reasoning may stem from common structural limitations inherent to the LPA.
Authors: Jiaxin Zhang, Yiqi Wang, Siwei Wang, Xihong Yang, Yu Shi, Xinwang Liu, En Zhu
Abstract: Graph Neural Networks frequently exhibit significant performance degradation in the out-of-distribution test scenario. While test-time training (TTT) offers a promising solution, existing Parameter Finetuning (PaFT) paradigm suffer from catastrophic forgetting, hindering their real-world applicability. We propose TTReFT, a novel Test-Time Representation FineTuning framework that transitions the adaptation target from model parameters to latent representations. Specifically, TTReFT achieves this through three key innovations: (1) uncertainty-guided node selection for specific interventions, (2) low-rank representation interventions that preserve pre-trained knowledge, and (3) an intervention-aware masked autoencoder that dynamically adjust masking strategy to accommodate the node selection scheme. Theoretically, we establish guarantees for TTReFT in OOD settings. Empirically, extensive experiments across five benchmark datasets demonstrate that TTReFT achieves consistent and superior performance. Our work establishes representation finetuning as a new paradigm for graph TTT, offering both theoretical grounding and immediate practical utility for real-world deployment.
Authors: Yiming Wang, Zhuosheng Zhang, Rui Wang
Abstract: Parallel thinking enhances LLM reasoning by multi-path sampling and aggregation. In system-level evaluations, a global parallelism level N is allocated to all samples, typically set large to maximize overall dataset accuracy. However, due to sample heterogeneity, some samples can achieve comparable performance with a smaller N'< N, causing budget redundancy. This incompatibility between system-level efficacy and sample-level efficiency constitutes the overscaling curse. In this paper, we formalize and quantify the overscaling curse, showing its universality and severity in practice, and analyze its trigger mechanism. We then propose a lightweight method, T2, to break the overscaling curse, which utilizes latent representations to estimate the optimal parallelism level for each sample before decoding. Experiments show that T2 significantly reduces cost while maintaining comparable performance, enabling more efficient parallel thinking.
Authors: Stanislav Budzinskiy, Marian Gloser, Tolunay Yilmaz, Ying Hong Tham, Yuanyi Lin, Wenyi Fang, Fan Wu, Philipp Petersen
Abstract: Mixed-precision computations are a hallmark of the current stage of AI, driving the progress in large language models towards efficient, locally deployable solutions. This article addresses the floating-point computation of compositionally-rich functions, concentrating on transformer inference. Based on the rounding error analysis of a composition $f(g(\mathrm{x}))$, we provide an adaptive strategy that selects a small subset of components of $g(\mathrm{x})$ to be computed more accurately while all other computations can be carried out with lower accuracy. We then explain how this strategy can be applied to different compositions within a transformer and illustrate its overall effect on transformer inference. We study the effectiveness of this algorithm numerically on GPT-2 models and demonstrate that already very low recomputation rates allow for improvements of up to two orders of magnitude in accuracy.
Authors: Vasileios Sevetlidis, George Pavlidis
Abstract: Modern deep-learning training is not memoryless. Updates depend on optimizer moments and averaging, data-order policies (random reshuffling vs with-replacement, staged augmentations and replay), the nonconvex path, and auxiliary state (teacher EMA/SWA, contrastive queues, BatchNorm statistics). This survey organizes mechanisms by source, lifetime, and visibility. It introduces seed-paired, function-space causal estimands; portable perturbation primitives (carry/reset of momentum/Adam/EMA/BN, order-window swaps, queue/teacher tweaks); and a reporting checklist with audit artifacts (order hashes, buffer/BN checksums, RNG contracts). The conclusion is a protocol for portable, causal, uncertainty-aware measurement that attributes how much training history matters across models, data, and regimes.
Authors: Jinhao Zhang Yunquan Zhang, Zicheng yan, Boyang Zhang, Jun Sun, Daning Cheng
Abstract: Post Training Quantization (PTQ), a mainstream model compression technique, often leads to the paradoxical 'low error, high loss' phenomenon because it focuses solely on minimizing quantization error. The root cause lies in the Hessian matrix of the LLM loss landscape: a few high curvature directions are extremely sensitive to perturbations. To address this, we propose the Hessian Robust Quantization (HeRo Q) algorithm, which applies a lightweight, learnable rotation-compression matrix to the weight space prior to quantization. This joint framework reshapes the loss landscape by reducing the largest Hessian eigenvalue and reducing its max eigenvalue, thereby significantly enhancing robustness to quantization noise. HeRo-Q requires no architectural modifications, incurs negligible computational overhead, and integrates seamlessly into existing PTQ pipelines. Experiments on Llama and Qwen models show that HeRo Q consistently outperforms state of the art methods including GPTQ, AWQ, and SpinQuant not only achieving superior performance under standard W4A8 settings, but also excelling in the highly challenging W3A16 ultra low bit regime, where it boosts GSM8K accuracy on Llama3 8B to 70.15\% and effectively avoids the logical collapse commonly seen in aggressive quantization.
Authors: Jan Schuchardt, Nikita Kalinin
Abstract: We study privacy amplification for differentially private model training with matrix factorization under random allocation (also known as the balls-in-bins model). Recent work by Choquette-Choo et al. (2025) proposes a sampling-based Monte Carlo approach to compute amplification parameters in this setting. However, their guarantees either only hold with some high probability or require random abstention by the mechanism. Furthermore, the required number of samples for ensuring $(\epsilon,\delta)$-DP is inversely proportional to $\delta$. In contrast, we develop sampling-free bounds based on R\'enyi divergence and conditional composition. The former is facilitated by a dynamic programming formulation to efficiently compute the bounds. The latter complements it by offering stronger privacy guarantees for small $\epsilon$, where R\'enyi divergence bounds inherently lead to an over-approximation. Our framework applies to arbitrary banded and non-banded matrices. Through numerical comparisons, we demonstrate the efficacy of our approach across a broad range of matrix mechanisms used in research and practice.
Authors: Patrick Kruger, Rafael Diaz, Simon Hauschulz, Stefan Harries, Hanno Gottschalk
Abstract: In this paper, we explore the use of generative artificial intelligence (GenAI) for ship propeller design. While traditional forward machine learning models predict the performance of mechanical components based on given design parameters, GenAI models aim to generate designs that achieve specified performance targets. In particular, we employ conditional flow matching to establish a bidirectional mapping between design parameters and simulated noise that is conditioned on performance labels. This approach enables the generation of multiple valid designs corresponding to the same performance targets by sampling over the noise vector. To support model training, we generate data using a vortex lattice method for numerical simulation and analyze the trade-off between model accuracy and the amount of available data. We further propose data augmentation using pseudo-labels derived from less data-intensive forward surrogate models, which can often improve overall model performance. Finally, we present examples of distinct propeller geometries that exhibit nearly identical performance characteristics, illustrating the versatility and potential of GenAI in engineering design.
Authors: Evandro S. Ortigossa, Eran Segal
Abstract: Transformer-based models have recently made significant advances in accurate time-series forecasting, but even these architectures struggle to scale efficiently while capturing long-term temporal dynamics. Mixture-of-Experts (MoE) layers are a proven solution to scaling problems in natural language processing. However, existing MoE approaches for time-series forecasting rely on token-wise routing mechanisms, which may fail to exploit the natural locality and continuity of temporal data. In this work, we introduce Seg-MoE, a sparse MoE design that routes and processes contiguous time-step segments rather than making independent expert decisions. Token segments allow each expert to model intra-segment interactions directly, naturally aligning with inherent temporal patterns. We integrate Seg-MoE layers into a time-series Transformer and evaluate it on multiple multivariate long-term forecasting benchmarks. Seg-MoE consistently achieves state-of-the-art forecasting accuracy across almost all prediction horizons, outperforming both dense Transformers and prior token-wise MoE models. Comprehensive ablation studies confirm that segment-level routing is the key factor driving these gains. Our results show that aligning the MoE routing granularity with the inherent structure of time series provides a powerful, yet previously underexplored, inductive bias, opening new avenues for conditionally sparse architectures in sequential data modeling.
Authors: Vahid Shahverdi, Giovanni Luca Marchetti, Georg B\"okman, Kathl\'en Kohn
Abstract: We investigate the relation between end-to-end equivariance and layerwise equivariance in deep neural networks. We prove the following: For a network whose end-to-end function is equivariant with respect to group actions on the input and output spaces, there is a parameter choice yielding the same end-to-end function such that its layers are equivariant with respect to some group actions on the latent spaces. Our result assumes that the parameters of the model are identifiable in an appropriate sense. This identifiability property has been established in the literature for a large class of networks, to which our results apply immediately, while it is conjectural for others. The theory we develop is grounded in an abstract formalism, and is therefore architecture-agnostic. Overall, our results provide a mathematical explanation for the emergence of equivariant structures in the weights of neural networks during training -- a phenomenon that is consistently observed in practice.
Authors: Jinjun Peng, Magnus Saebo, Tianjun Zhong, Yi-Jie Cheng, Junfeng Yang, Baishakhi Ray, Simin Chen, Yangruibo Ding
Abstract: The deployment of coding agents in privacy-sensitive and resource-constrained environments drives the demand for capable open-weight Small Language Models (SLMs). However, they suffer from a fundamental capability gap: unlike frontier large models, they lack the inference-time strong generalization to work with complicated, unfamiliar codebases. We identify that the prevailing Task-Centric Learning (TCL) paradigm, which scales exposure across disparate repositories, fails to address this limitation. In response, we propose Repository-Centric Learning (RCL), a paradigm shift that prioritizes vertical repository depth over horizontal task breadth, suggesting SLMs must internalize the "physics" of a target software environment through parametric knowledge acquisition, rather than attempting to recover it via costly inference-time search. Following this new paradigm, we design a four-unit Repository-Centric Experience, transforming static codebases into interactive learning signals, to train SWE-Spot-4B, a family of highly compact models built as repo-specialized experts that breaks established scaling trends, outperforming open-weight models up to larger (e.g., CWM by Meta, Qwen3-Coder-30B) and surpassing/matching efficiency-focused commercial models (e.g., GPT-4.1-mini, GPT-5-nano) across multiple SWE tasks. Further analysis reveals that RCL yields higher training sample efficiency and lower inference costs, emphasizing that for building efficient intelligence, repository mastery is a distinct and necessary dimension that complements general coding capability.
Authors: Vasileios Sevetlidis, George Pavlidis
Abstract: Deep networks learn internal representations whose geometry--how features bend, rotate, and evolve--affects both generalization and robustness. Existing similarity measures such as CKA or SVCCA capture pointwise overlap between activation sets, but miss how representations change along input paths. Two models may appear nearly identical under these metrics yet respond very differently to perturbations or adversarial stress. We introduce representation holonomy, a gauge-invariant statistic that measures this path dependence. Conceptually, holonomy quantifies the "twist" accumulated when features are parallel-transported around a small loop in input space: flat representations yield zero holonomy, while nonzero values reveal hidden curvature. Our estimator fixes gauge through global whitening, aligns neighborhoods using shared subspaces and rotation-only Procrustes, and embeds the result back to the full feature space. We prove invariance to orthogonal (and affine, post-whitening) transformations, establish a linear null for affine layers, and show that holonomy vanishes at small radii. Empirically, holonomy increases with loop radius, separates models that appear similar under CKA, and correlates with adversarial and corruption robustness. It also tracks training dynamics as features form and stabilize. Together, these results position representation holonomy as a practical and scalable diagnostic for probing the geometric structure of learned representations beyond pointwise similarity.
Authors: Tianqi Zhao, Guanyang Wang, Yan Shuo Tan, Qiong Zhang
Abstract: Clustering tabular data is a fundamental yet challenging problem due to heterogeneous feature types, diverse data-generating mechanisms, and the absence of transferable inductive biases across datasets. Prior-fitted networks (PFNs) have recently demonstrated strong generalization in supervised tabular learning by amortizing Bayesian inference under a broad synthetic prior. Extending this paradigm to clustering is nontrivial: clustering is unsupervised, admits a combinatorial and permutation-invariant output space, and requires inferring the number of clusters. We introduce TabClustPFN, a prior-fitted network for tabular data clustering that performs amortized Bayesian inference over both cluster assignments and cluster cardinality. Pretrained on synthetic datasets drawn from a flexible clustering prior, TabClustPFN clusters unseen datasets in a single forward pass, without dataset-specific retraining or hyperparameter tuning. The model naturally handles heterogeneous numerical and categorical features and adapts to a wide range of clustering structures. Experiments on synthetic data and curated real-world tabular benchmarks show that TabClustPFN outperforms classical, deep, and amortized clustering baselines, while exhibiting strong robustness in out-of-the-box exploratory settings. Code is available at https://github.com/Tianqi-Zhao/TabClustPFN.
Authors: Li Ju, Mayank Nautiyal, Andreas Hellander, Ekta Vats, Prashant Singh
Abstract: Vision-Language Models (VLMs) are typically deterministic in nature and lack intrinsic mechanisms to quantify epistemic uncertainty, which reflects the model's lack of knowledge or ignorance of its own representations. We theoretically motivate negative log-density of an embedding as a proxy for the epistemic uncertainty, where low-density regions signify model ignorance. The proposed method REPVLM computes the probability density on the hyperspherical manifold of the VLM embeddings using Riemannian Flow Matching. We empirically demonstrate that REPVLM achieves near-perfect correlation between uncertainty and prediction error, significantly outperforming existing baselines. Beyond classification, we also demonstrate that the model also provides a scalable metric for out-of-distribution detection and automated data curation.
Authors: Xingyue Zhang, Yuxuan Bao, Mars Liyao Gao, J. Nathan Kutz
Abstract: Bridging the gap between data-rich training regimes and observation-sparse deployment conditions remains a central challenge in spatiotemporal field reconstruction, particularly when target domains exhibit distributional shifts, heterogeneous structure, and multi-scale dynamics absent from available training data. We present SENDAI, a hierarchical Sparse-measurement, EfficieNt Data AssImilation Framework that reconstructs full spatial states from hyper sparse sensor observations by combining simulation-derived priors with learned discrepancy corrections. We demonstrate the performance on satellite remote sensing, reconstructing MODIS (Moderate Resolution Imaging Spectroradiometer) derived vegetation index fields across six globally distributed sites. Using seasonal periods as a proxy for domain shift, the framework consistently outperforms established baselines that require substantially denser observations -- SENDAI achieves a maximum SSIM improvement of 185% over traditional baselines and a 36% improvement over recent high-frequency-based methods. These gains are particularly pronounced for landscapes with sharp boundaries and sub-seasonal dynamics; more importantly, the framework effectively preserves diagnostically relevant structures -- such as field topologies, land cover discontinuities, and spatial gradients. By yielding corrections that are more structurally and spectrally separable, the reconstructed fields are better suited for downstream inference of indirectly observed variables. The results therefore highlight a lightweight and operationally viable framework for sparse-measurement reconstruction that is applicable to physically grounded inference, resource-limited deployment, and real-time monitor and control.
Authors: Abhijeet Sinha, Sundari Elango, Dianbo Liu
Abstract: Many reinforcement learning (RL) problems admit multiple terminal solutions of comparable quality, where the goal is not to identify a single optimum but to represent a diverse set of high-quality outcomes. Nevertheless, policies trained by standard expected return maximization routinely collapse onto a small subset of outcomes, a phenomenon commonly attributed to insufficient exploration or weak regularization. We show that this explanation is incomplete: outcome level mode collapse is a structural consequence of the expected-return objective itself. Under idealized learning dynamics, the log-probability ratio between any two outcomes evolves linearly in their reward difference, implying exponential ratio divergence and inevitable collapse independent of the exploration strategy, entropy regularization, or optimization algorithm. We identify the source of this pathology as the probability multiplier inside the expectation and propose a minimal correction: inverse probability scaling, which removes outcome-frequency amplification from the learning signal, fundamentally changes the learning dynamics, and provably yields reward-proportional terminal distributions, preventing collapse in multimodal settings. We instantiate this principle in Group Relative Policy Optimization (GRPO) as a drop-in modification, IPS-GRPO, requiring no auxiliary models or architectural changes. Across different reasoning and molecular generation tasks, IPS-GRPO consistently reduces outcome-level mode collapse while matching or exceeding baseline performance, suggesting that correcting the objective rather than adding exploration heuristics is key to reliable multimodal policy optimization.
Authors: Qisong Xiao, Xinhai Chen, Qinglin Wang, Xiaowei Guo, Binglin Wang, Weifeng Chen, Zhichao Wang, Yunfei Liu, Rui Xia, Hang Zou, Gencheng Liu, Shuai Li, Jie Liu
Abstract: Deep learning has emerged as a promising paradigm for spatio-temporal modeling of fluid dynamics. However, existing approaches often suffer from limited generalization to unseen flow conditions and typically require retraining when applied to new scenarios. In this paper, we present LLM4Fluid, a spatio-temporal prediction framework that leverages Large Language Models (LLMs) as generalizable neural solvers for fluid dynamics. The framework first compresses high-dimensional flow fields into a compact latent space via reduced-order modeling enhanced with a physics-informed disentanglement mechanism, effectively mitigating spatial feature entanglement while preserving essential flow structures. A pretrained LLM then serves as a temporal processor, autoregressively predicting the dynamics of physical sequences with time series prompts. To bridge the modality gap between prompts and physical sequences, which can otherwise degrade prediction accuracy, we propose a dedicated modality alignment strategy that resolves representational mismatch and stabilizes long-term prediction. Extensive experiments across diverse flow scenarios demonstrate that LLM4Fluid functions as a robust and generalizable neural solver without retraining, achieving state-of-the-art accuracy while exhibiting powerful zero-shot and in-context learning capabilities. Code and datasets are publicly available at https://github.com/qisongxiao/LLM4Fluid.
Authors: Wu S. Zihan, Ariane Delrocq, Wulfram Gerstner, Guillaume Bellec
Abstract: While end-to-end self-supervised learning with backpropagation (global BP-SSL) has become central for training modern AI systems, theories of local self-supervised learning (local-SSL) have struggled to build functional representations in deep neural networks. To establish a link between global and local rules, we first develop a theory for deep linear networks: we identify conditions for local-SSL algorithms (like Forward-forward or CLAPP) to implement exactly the same weight update as a global BP-SSL. Starting from the theoretical insights, we then develop novel variants of local-SSL algorithms to approximate global BP-SSL in deep non-linear convolutional neural networks. Variants that improve the similarity between gradient updates of local-SSL with those of global BP-SSL also show better performance on image datasets (CIFAR-10, STL-10, and Tiny ImageNet). The best local-SSL rule with the CLAPP loss function matches the performance of a comparable global BP-SSL with InfoNCE or CPC-like loss functions, and improves upon state-of-the-art for local SSL on these benchmarks.
Authors: Luca Benfenati, Matteo Risso, Andrea Vannozzi, Ahmet Caner Y\"uz\"ug\"uler, Lukas Cavigelli, Enrico Macii, Daniele Jahier Pagliari, Alessio Burrello
Abstract: Key--value (KV) caching enables fast autoregressive decoding but at long contexts becomes a dominant bottleneck in High Bandwidth Memory (HBM) capacity and bandwidth. A common mitigation is to compress cached keys and values by projecting per-head matrixes to a lower rank, storing only the projections in the HBM. However, existing post-training approaches typically fit these projections using SVD-style proxy objectives, which may poorly reflect end-to-end reconstruction after softmax, value mixing, and subsequent decoder-layer transformations. For these reasons, we introduce StiefAttention, a post-training KV-cache compression method that learns \emph{orthonormal} projection bases by directly minimizing \emph{decoder-layer output reconstruction error}. StiefAttention additionally precomputes, for each layer, an error-rank profile over candidate ranks, enabling flexible layer-wise rank allocation under a user-specified error budget. Noteworthy, on Llama3-8B under the same conditions, StiefAttention outperforms EigenAttention by $11.9$ points on C4 perplexity and $5.4\%$ on 0-shot MMLU accuracy at iso-compression, yielding lower relative error and higher cosine similarity with respect to the original decoder-layer outputs.
Authors: Alexandre Myara, Nicolas Bourriez, Thomas Boyer, Thomas Lemercier, Ihab Bendidi, Auguste Genovesio
Abstract: Disentangled representation learning aims to map independent factors of variation to independent representation components. On one hand, purely unsupervised approaches have proven successful on fully disentangled synthetic data, but fail to recover semantic factors from real data without strong inductive biases. On the other hand, supervised approaches are unstable and hard to scale to large attribute sets because they rely on adversarial objectives or auxiliary classifiers. We introduce \textsc{XFactors}, a weakly-supervised VAE framework that disentangles and provides explicit control over a chosen set of factors. Building on the Disentangled Information Bottleneck perspective, we decompose the representation into a residual subspace $\mathcal{S}$ and factor-specific subspaces $\mathcal{T}_1,\ldots,\mathcal{T}_K$ and a residual subspace $\mathcal{S}$. Each target factor is encoded in its assigned $\mathcal{T}_i$ through contrastive supervision: an InfoNCE loss pulls together latents sharing the same factor value and pushes apart mismatched pairs. In parallel, KL regularization imposes a Gaussian structure on both $\mathcal{S}$ and the aggregated factor subspaces, organizing the geometry without additional supervision for non-targeted factors and avoiding adversarial training and classifiers. Across multiple datasets, with constant hyperparameters, \textsc{XFactors} achieves state-of-the-art disentanglement scores and yields consistent qualitative factor alignment in the corresponding subspaces, enabling controlled factor swapping via latent replacement. We further demonstrate that our method scales correctly with increasing latent capacity and evaluate it on the real-world dataset CelebA. Our code is available at \href{https://github.com/ICML26-anon/XFactors}{github.com/ICML26-anon/XFactors}.
Authors: Qinglun Li, Anke Tang, Miao Zhang, Mengzhu Wang, Quanjun Yin, Li Shen
Abstract: Model merging efficiently aggregates capabilities from multiple fine-tuned models into a single one, operating purely in parameter space without original data or expensive re-computation. Despite empirical successes, a unified theory for its effectiveness under heterogeneous finetuning hyperparameters (e.g., varying learning rates, batch sizes) remains missing. Moreover, the lack of hyperparameter transparency in open-source fine-tuned models makes it difficult to predict merged-model performance, leaving practitioners without guidance on how to fine-tune merge-friendly experts. To address those two challenges, we employ $L_2$-Stability theory under heterogeneous hyperparameter environments to analyze the generalization of the merged model $\boldsymbol{x}_{avg}$. This pioneering analysis yields two key contributions: (i) \textit{A unified theoretical framework} is provided to explain existing merging algorithms, revealing how they optimize specific terms in our bound, thus offering a strong theoretical foundation for empirical observations. (ii) \textit{Actionable recommendations} are proposed for practitioners to strategically fine-tune expert models, enabling the construction of merge-friendly models within the pretraining-to-finetuning pipeline. Extensive experiments on the ResNet/Vit family across 20/8 visual classification tasks, involving thousands of finetuning models, robustly confirm the impact of different hyperparameters on the generalization of $\boldsymbol{x}_{avg}$ predicted by our theoretical results.
Authors: Mohamed Elgaar, Hadi Amiri
Abstract: Curriculum learning changes the order of pre-training data, but it remains unclear whether it changes the learning trajectory or mainly reorders exposure over a fixed trajectory. We train Pythia models (14M-410M parameters) for 300B tokens under three linguistically motivated curricula-Age-of-Acquisition, word frequency, and Verb Variation (VV)-and compare each against Random ordering; at 1B parameters we compare Random and VV. Across orderings, training follows a shared sequence of latent phases, while curricula mainly change within-phase data exposure. In smaller models (up to 160M parameters), Random ordering exhibits higher gradient noise and stronger late-training output-head spectral saturation, alongside lower final accuracy; curricula reduce both effects at matched compute. At larger scales, saturation differences are smaller and curriculum gains shrink. We formalize the link between difficulty pacing and optimization stability in an idealized analysis based on gradient-variance control, and our results point to a practical takeaway: curricula help by stabilizing within-phase optimization rather than by creating new phases.
Authors: Tien Dang, The-Hai Nguyen, Dinh Mai Phuong, Nguyen Minh Phuong, Hoang Thanh-Tung, Le-Minh Nguyen, Naoya Inoue
Abstract: We consider representation misdirection (RM), a class of LLM unlearning methods that achieves forgetting by manipulating the forget-representations, that is, latent representations of forget samples. Despite being important, the roles of target vectors used in RM, however, remain underexplored. Here, we approach and revisit RM through the lens of the linear representation hypothesis. Specifically, if one can somehow identify a one-dimensional representation corresponding to a high-level concept, the linear representation hypothesis enables linear operations on this concept vector within the forget-representation space. Under this view, we hypothesize that, beyond forgetting, machine unlearning elicits controllable side behaviors and stronger side capabilities corresponding to the high-level concept. Our hypothesis is empirically validated across a wide range of tasks, including behavioral control (e.g., controlling unlearned models' truth, sentiment, and refusal) and capability enhancement (e.g., improving unlearned models' in-context learning capability). Our findings reveal that this fairly attractive phenomenon could be either a hidden risk if misused or a mechanism that can be harnessed for developing models that require stronger capabilities and controllable behaviors.
Authors: Nan Lin, Yanbo Wang, Jacco Heres, Peter Palensky, Pedro P. Vergara
Abstract: Smart meter data is the foundation for planning and operating the distribution network. Unfortunately, such data are not always available due to privacy regulations. Meanwhile, the collected data may be corrupted due to sensor or transmission failure, or it may not have sufficient resolution for downstream tasks. A wide range of generative tasks is formulated to address these issues, including synthetic data generation, missing data imputation, and super-resolution. Despite the success of machine learning models on these tasks, dedicated models need to be designed and trained for each task, leading to redundancy and inefficiency. In this paper, by recognizing the powerful modeling capability of flow matching models, we propose a new approach to unify diverse smart meter data generative tasks with a single model trained for conditional generation. The proposed flow matching models are trained to generate challenging, high-dimensional time series data, specifically monthly smart meter data at a 15 min resolution. By viewing different generative tasks as distinct forms of partial data observations and injecting them into the generation process, we unify tasks such as imputation and super-resolution with a single model, eliminating the need for re-training. The data generated by our model not only are consistent with the given observations but also remain realistic, showing better performance against interpolation and other machine learning based baselines dedicated to the tasks.
Authors: Lukas Sch\"afer, Pallavi Choudhury, Abdelhak Lemkhenter, Chris Lovett, Somjit Nath, Luis Fran\c{c}a, Matheus Ribeiro Furtado de Mendon\c{c}a, Alex Lamb, Riashat Islam, Siddhartha Sen, John Langford, Katja Hofmann, Sergio Valcarcel Macua
Abstract: Behavior cloning (BC) is a practical offline imitation learning method, but it often fails when expert demonstrations are limited. Recent works have introduced a class of architectures named predictive inverse dynamics models (PIDM) that combine a future state predictor with an inverse dynamics model (IDM). While PIDM often outperforms BC, the reasons behind its benefits remain unclear. In this paper, we provide a theoretical explanation: PIDM introduces a bias-variance tradeoff. While predicting the future state introduces bias, conditioning the IDM on the prediction can significantly reduce variance. We establish conditions on the state predictor bias for PIDM to achieve lower prediction error and higher sample efficiency than BC, with the gap widening when additional data sources are available. We validate the theoretical insights empirically in 2D navigation tasks, where BC requires up to five times (three times on average) more demonstrations than PIDM to reach comparable performance; and in a complex 3D environment in a modern video game with high-dimensional visual inputs and stochastic transitions, where BC requires over 66\% more samples than PIDM.
Authors: Yaxi Hu, Johanna D\"ungler, Bernhard Sch\"olkopf, Amartya Sanyal
Abstract: We introduce the (Wishart) projection mechanism, a randomized map of the form $S \mapsto M f(S)$ with $M \sim W_d(1/r I_d, r)$ and study its differential privacy properties. For vector-valued queries $f$, we prove non-asymptotic DP guarantees without any additive noise, showing that Wishart randomness alone can suffice. For matrix-valued queries, however, we establish a sharp negative result: in the noise-free setting, the mechanism is not DP, and we demonstrate its vulnerability by implementing a near perfect membership inference attack (AUC $> 0.99$). We then analyze a noisy variant and prove privacy amplification due to randomness and low rank projection, in both large- and small-rank regimes, yielding stronger privacy guarantees than additive noise alone. Finally, we show that LoRA-style updates are an instance of the matrix-valued mechanism, implying that LoRA is not inherently private despite its built-in randomness, but that low-rank fine-tuning can be more private than full fine-tuning at the same noise level. Preliminary experiments suggest that tighter accounting enables lower noise and improved accuracy in practice.
Authors: Kaustubh Sharma, Srijan Tiwari, Ojasva Nema, Parikshit Pareek
Abstract: Prior-Data Fitted Networks (PFNs) enable efficient amortized inference but lack transparent access to their learned priors and kernels. This opacity hinders their use in downstream tasks, such as surrogate-based optimization, that require explicit covariance models. We introduce an interpretability-driven framework for amortized spectral discovery from pre-trained PFNs with decoupled attention. We perform a mechanistic analysis on a trained PFN that identifies attention latent output as the key intermediary, linking observed function data to spectral structure. Building on this insight, we propose decoder architectures that map PFN latents to explicit spectral density estimates and corresponding stationary kernels via Bochner's theorem. We study this pipeline in both single-realization and multi-realization regimes, contextualizing theoretical limits on spectral identifiability and proving consistency when multiple function samples are available. Empirically, the proposed decoders recover complex multi-peak spectral mixtures and produce explicit kernels that support Gaussian process regression with accuracy comparable to PFNs and optimization-based baselines, while requiring only a single forward pass. This yields orders-of-magnitude reductions in inference time compared to optimization-based baselines.
Authors: Rebecca Pelke, Joel Klein, Jose Cubero-Cascante, Nils Bosbach, Jan Moritz Joseph, Rainer Leupers
Abstract: Computing-in-Memory (CIM) accelerators are a promising solution for accelerating Machine Learning (ML) workloads, as they perform Matrix-Vector Multiplications (MVMs) on crossbar arrays directly in memory. Although the bit widths of the crossbar inputs and cells are very limited, most CIM compilers do not support quantization below 8 bit. As a result, a single MVM requires many compute cycles, and weights cannot be efficiently stored in a single crossbar cell. To address this problem, we propose a mixed-precision training and compilation framework for CIM architectures. The biggest challenge is the massive search space, that makes it difficult to find good quantization parameters. This is why we introduce a reinforcement learning-based strategy to find suitable quantization configurations that balance latency and accuracy. In the best case, our approach achieves up to a 2.48x speedup over existing state-of-the-art solutions, with an accuracy loss of only 0.086 %.
Authors: Alberto Fern\'andez-Hern\'andez, Cristian P\'erez-Corral, Jose I. Mestre, Manuel F. Dolz, Enrique S. Quintana-Ort\'i
Abstract: Adam has been at the core of large-scale training for almost a decade, yet a simple empirical fact remains unaccounted for: both validation scores and the qualitative behaviour of the training runs improve when the momentum parameters satisfy $\beta_{1}=\beta_{2}$. Some recent studies have reported this pattern, but there is still no explanation for why this choice helps. We show that this choice is closely tied to a structural property that we refer to as \textit{gradient scale invariance}. We formalize this notion and prove that Adam becomes gradient scale invariant of first order if and only if $\beta_{1}=\beta_{2}$. This perspective places the balanced regime of Adam in direct alignment with the design principles underlying several recent optimizers that explicitly enforce scale-robust updates. The theory is supported by experiments across vision and language tasks, and across different architectural families, in which rescaling the gradient has a markedly smoother effect on the update when $\beta_{1}=\beta_{2}$. Overall, our results offer a coherent explanation for an open question in the behavior of Adam and provide a simple principle that helps guide the design of future optimizers.
Authors: Vincent Lemaire, N\'edra Meloulli, Pierre Jaquet
Abstract: Sepsis remains one of the most complex and heterogeneous syndromes in intensive care, characterized by diverse physiological trajectories and variable responses to treatment. While deep learning models perform well in the early prediction of sepsis, they often lack interpretability and ignore latent patient sub-phenotypes. In this work, we propose a machine learning framework by opening up a new avenue for addressing this issue: a relational approach. Temporal data from electronic medical records (EMRs) are viewed as multivariate patient logs and represented in a relational data schema. Then, a propositionalisation technique (based on classic aggregation/selection functions from the field of relational data) is applied to construct interpretable features to "flatten" the data. Finally, the flattened data is classified using a selective naive Bayesian classifier. Experimental validation demonstrates the relevance of the suggested approach as well as its extreme interpretability. The interpretation is fourfold: univariate, global, local, and counterfactual.
Authors: Chenrui Xu, Wenjing Yan, Ying-Jun Angela Zhang
Abstract: Training large-scale neural networks requires solving nonconvex optimization where the choice of optimizer fundamentally determines both convergence behavior and computational efficiency. While adaptive methods like Adam have long dominated practice, the recently proposed Muon optimizer achieves superior performance through orthogonalized momentum updates that enforce isotropic geometry with uniform singular values. However, this strict isotropy discards potentially valuable curvature information encoded in gradient spectra, motivating optimization methods that balance geometric structure with adaptivity. We introduce FISMO (Fisher-Structured Momentum-Orthogonalized) optimizer, which generalizes isotropic updates to incorporate anisotropic curvature information through Fisher information geometry. By reformulating the optimizer update as a trust-region problem constrained by a Kronecker-factored Fisher metric, FISMO achieves structured preconditioning that adapts to local loss landscape geometry while maintaining computational tractability. We establish convergence guarantees for FISMO in stochastic nonconvex settings, proving an $\mathcal{O}(1/\sqrt{T})$ rate for the expected squared gradient norm with explicit characterization of variance reduction through mini-batching. Empirical evaluation on image classification and language modeling benchmarks demonstrates that FISMO achieves superior training efficiency and final performance compared to established baselines.
Authors: Germain Vivier-Ardisson, Micha\"el E. Sander, Axel Parmentier, Mathieu Blondel
Abstract: Knapsack and Top-k operators are useful for selecting discrete subsets of variables. However, their integration into neural networks is challenging as they are piecewise constant, yielding gradients that are zero almost everywhere. In this paper, we propose a unified framework casting these operators as dynamic programs, and derive differentiable relaxations by smoothing the underlying recursions. On the algorithmic side, we develop efficient parallel algorithms supporting both deterministic and stochastic forward passes, and vector-Jacobian products for the backward pass. On the theoretical side, we prove that Shannon entropy is the unique regularization choice yielding permutation-equivariant operators, and characterize regularizers inducing sparse selections. Finally, on the experimental side, we demonstrate our framework on a decision-focused learning benchmark, a constrained dynamic assortment RL problem, and an extension of discrete VAEs.
Authors: Jun Qi, Chao-Han Huck Yang, Pin-Yu Chen, Min-Hsiu Hsieh, Hector Zenil, Jesper Tegner
Abstract: Hybrid quantum-classical learning models increasingly integrate neural networks with variational quantum circuits (VQCs) to exploit complementary inductive biases. However, many existing approaches rely on tightly coupled architectures or task-specific encoders, limiting conceptual clarity, generality, and transferability across learning settings. In this work, we introduce Quantum LEGO Learning, a modular and architecture-agnostic learning framework that treats classical and quantum components as reusable, composable learning blocks with well-defined roles. Within this framework, a pre-trained classical neural network serves as a frozen feature block, while a VQC acts as a trainable adaptive module that operates on structured representations rather than raw inputs. This separation enables efficient learning under constrained quantum resources and provides a principled abstraction for analyzing hybrid models. We develop a block-wise generalization theory that decomposes learning error into approximation and estimation components, explicitly characterizing how the complexity and training status of each block influence overall performance. Our analysis generalizes prior tensor-network-specific results and identifies conditions under which quantum modules provide representational advantages over comparably sized classical heads. Empirically, we validate the framework through systematic block-swap experiments across frozen feature extractors and both quantum and classical adaptive heads. Experiments on quantum dot classification demonstrate stable optimization, reduced sensitivity to qubit count, and robustness to realistic noise.
Authors: Adia Lumadjeng, Ilker Birbil, Erman Acar
Abstract: We introduce ECSEL, an explainable classification method that learns formal expressions in the form of signomial equations, motivated by the observation that many symbolic regression benchmarks admit compact signomial structure. ECSEL directly constructs a structural, closed-form expression that serves as both a classifier and an explanation. On standard symbolic regression benchmarks, our method recovers a larger fraction of target equations than competing state-of-the-art approaches while requiring substantially less computation. Leveraging this efficiency, ECSEL achieves classification accuracy competitive with established machine learning models without sacrificing interpretability. Further, we show that ECSEL satisfies some desirable properties regarding global feature behavior, decision-boundary analysis, and local feature attributions. Experiments on benchmark datasets and two real-world case studies i.e., e-commerce and fraud detection, demonstrate that the learned equations expose dataset biases, support counterfactual reasoning, and yield actionable insights.
Authors: Tongze Wang, Xiaohui Xie, Wenduo Wang, Chuyi Wang, Jinzhou Liu, Boyan Huang, Yannan Hu, Youjian Zhao, Yong Cui
Abstract: With the rapid growth of encrypted network traffic, effective traffic classification has become essential for network security and quality of service management. Current machine learning and deep learning approaches for traffic classification face three critical challenges: computational inefficiency of Transformer architectures, inadequate traffic representations with loss of crucial byte-level features while retaining detrimental biases, and poor handling of long-tail distributions in real-world data. We propose NetMamba+, a framework that addresses these challenges through three key innovations: (1) an efficient architecture considering Mamba and Flash Attention mechanisms, (2) a multimodal traffic representation scheme that preserves essential traffic information while eliminating biases, and (3) a label distribution-aware fine-tuning strategy. Evaluation experiments on massive datasets encompassing four main classification tasks showcase NetMamba+'s superior classification performance compared to state-of-the-art baselines, with improvements of up to 6.44\% in F1 score. Moreover, NetMamba+ demonstrates excellent efficiency, achieving 1.7x higher inference throughput than the best baseline while maintaining comparably low memory usage. Furthermore, NetMamba+ exhibits superior few-shot learning abilities, achieving better classification performance with fewer labeled data. Additionally, we implement an online traffic classification system that demonstrates robust real-world performance with a throughput of 261.87 Mb/s. As the first framework to adapt Mamba architecture for network traffic classification, NetMamba+ opens new possibilities for efficient and accurate traffic analysis in complex network environments.
Authors: Yejin Kim, Dongjun Hwang, Sungmin Cha, Junsuk Choe
Abstract: Large Vision-Language Models (LVLMs) are widely adopted for their strong multimodal capabilities, yet they raise serious concerns such as privacy leakage and harmful content generation. Machine unlearning has emerged as a promising solution for removing the influence of specific data from trained models. However, existing approaches largely rely on gradient-based optimization, incurring substantial computational costs for large-scale LVLMs. To address this limitation, we propose Knowledge Vector Weakening (KVW), a training-free unlearning method that directly intervenes in the full model without gradient computation. KVW identifies knowledge vectors that are activated during the model's output generation on the forget set and progressively weakens their contributions, thereby preventing the model from exploiting undesirable knowledge. Experiments on the MLLMU and CLEAR benchmarks demonstrate that KVW achieves a stable forget-retain trade-off while significantly improving computational efficiency over gradient-based and LoRA-based unlearning methods.
Authors: Akash Dhasade, Anne-Marie Kermarrec, Igor Pavlovic, Diana Petrescu, Rafael Pires, Mathis Randl, Martijn de Vos
Abstract: Low-rank adaptation (LoRA) enables parameter efficient specialization of large language models (LLMs) through modular adapters, resulting in rapidly growing public adapter pools spanning diverse tasks. Effectively using these adapters requires routing: selecting and composing the appropriate adapters for a query. We introduce LORAUTER, a novel routing framework that selects and composes LoRA adapters using task representations rather than adapter characteristics. Unlike existing approaches that map queries directly to adapters, LORAUTER routes queries via task embeddings derived from small validation sets and does not require adapter training data. By operating at the task level, LORAUTER achieves efficient routing that scales with the number of tasks rather than the number of adapters. Experiments across multiple tasks show that LORAUTER consistently outperforms baseline routing approaches, matching Oracle performance (101.2%) when task-aligned adapters exist and achieving state-of-the-art results on unseen tasks (+5.2 points). We further demonstrate the robustness of LORAUTER to very large, noisy adapter pools by scaling it to over 1500 adapters.
Authors: Dennis Frauen, Athiya Deviyani, Mihaela van der Schaar, Stefan Feuerriegel
Abstract: Evaluating the performance of large language models (LLMs) from human preference data is crucial for obtaining LLM leaderboards. However, many existing approaches either rely on restrictive parametric assumptions or lack valid uncertainty quantification when flexible machine learning methods are used. In this paper, we propose a nonparametric statistical framework, DMLEval, for comparing and ranking LLMs from preference data using debiased machine learning (DML). For this, we introduce generalized average ranking scores (GARS), which generalize commonly used ranking models, including the Bradley-Terry model or PageRank/ Rank centrality, with complex human responses such as ties. DMLEval comes with the following advantages: (i) It produces statistically efficient estimates of GARS ranking scores. (ii) It naturally allows the incorporation of black-box machine learning methods for estimation. (iii) It can be combined with pre-trained LLM evaluators (e.g., using LLM-as-a-judge). (iv) It suggests optimal policies for collecting preference data under budget constraints. We demonstrate these advantages both theoretically and empirically using both synthetic and real-world preference datasets. In summary, our framework provides practitioners with powerful, state-of-the-art methods for comparing or ranking LLMs.
Authors: Xinwei Qiang, Hongmin Chen, Shixuan Sun, Jingwen Leng, Xin Liu, Minyi Guo
Abstract: Determinism is indispensable for reproducibility in large language model (LLM) training, yet it often exacts a steep performance cost. In widely used attention implementations such as FlashAttention-3, the deterministic backward pass can incur up to a 37.9% throughput reduction relative to its non-deterministic counterpart, primarily because gradient accumulation operations must be serialized to guarantee numerical consistency. This performance loss stems from suboptimal scheduling of compute and gradient-reduction phases, leading to significant hardware underutilization. To address this challenge, we formulate the backward pass of deterministic attention as a scheduling problem on a Directed Acyclic Graph (DAG) and derive schedules that minimize the critical path length. Building on this formulation, we present DASH (Deterministic Attention Scheduling for High-Throughput), which encapsulates two complementary scheduling strategies: (i) Descending Q-Tile Iteration, a reversed query-block traversal that shrinks pipeline stalls in causal attention, and (ii) Shift Scheduling, a theoretically optimal schedule within our DAG model that reduces pipeline stalls for both full and causal masks. Our empirical evaluations on NVIDIA H800 GPUs demonstrate that DASH narrows the performance gap of deterministic attention. The proposed strategies improve the throughput of the attention backward pass by up to 1.28$\times$ compared to the baseline, significantly advancing the efficiency of reproducible LLM training. Our code is open-sourced at https://github.com/SJTU-Liquid/deterministic-FA3.
Authors: Jigar Parekh, Philipp Bekemeyer
Abstract: Machine learning models are widely regarded as a way forward to tackle multi-query challenges that arise once expensive black-box simulations such as computational fluid dynamics are investigated. However, ensuring the desired level of accuracy for a certain task at minimal computational cost, e.g. as few black-box samples as possible, remains a challenges. Active learning strategies are used for scalar quantities to overcome this challenges and different so-called infill criteria exists and are commonly employed in several scenarios. Even though needed in various field an extension of active learning strategies towards field predictions is still lacking or limited to very specific scenarios and/or model types. In this paper we propose an active learning strategy for machine learning models that are capable if predicting field which is agnostic to the model architecture itself. For doing so, we combine a well-established Gaussian process model for a scalar reference value and simultaneously aim at reducing the epistemic model error and the difference between scalar and field predictions. Different specific forms of the above-mentioned approach are introduced and compared to each other as well as only scalar-valued based infill. Results are presented for the NASA common research model for an uncertainty propagation task showcasing high level of accuracy at significantly smaller cost compared to an approach without active learning.
Authors: Luis A. Ortega, Sim\'on Rodr\'iguez-Santana, Daniel Hern\'andez-Lobato
Abstract: We introduce a scalable method to approximate the kernel of the Linearized Laplace Approximation (LLA). For this, we use a surrogate deep neural network (DNN) that learns a compact feature representation whose inner product replicates the Neural Tangent Kernel (NTK). This avoids the need to compute large Jacobians. Training relies solely on efficient Jacobian-vector products, allowing to compute predictive uncertainty on large-scale pre-trained DNNs. Experimental results show similar or improved uncertainty estimation and calibration compared to existing LLA approximations. Notwithstanding, biasing the learned kernel significantly enhances out-of-distribution detection. This remarks the benefits of the proposed method for finding better kernels than the NTK in the context of LLA to compute prediction uncertainty given a pre-trained DNN.
Authors: Tingting Ni, Maryam Kamgarpour
Abstract: Meta reinforcement learning (RL) allows agents to leverage experience across a distribution of tasks on which the agent can train at will, enabling faster learning of optimal policies on new test tasks. Despite its success in improving sample complexity on test tasks, many real-world applications, such as robotics and healthcare, impose safety constraints during testing. Constrained meta RL provides a promising framework for integrating safety into meta RL. An open question in constrained meta RL is how to ensure the safety of the policy on the real-world test task, while reducing the sample complexity and thus, enabling faster learning of optimal policies. To address this gap, we propose an algorithm that refines policies learned during training, with provable safety and sample complexity guarantees for learning a near optimal policy on the test tasks. We further derive a matching lower bound, showing that this sample complexity is tight.
Authors: Zechuan Huang, Zhiguang Cao, Hongshu Guo, Yue-Jiao Gong, Zeyuan Ma
Abstract: Meta-Black-Box Optimization (MetaBBO) is an emerging avenue within Optimization community, where algorithm design policy could be meta-learned by reinforcement learning to enhance optimization performance. So far, the reward functions in existing MetaBBO works are designed by human experts, introducing certain design bias and risks of reward hacking. In this paper, we use Large Language Model~(LLM) as an automated reward discovery tool for MetaBBO. Specifically, we consider both effectiveness and efficiency sides. On effectiveness side, we borrow the idea of evolution of heuristics, introducing tailored evolution paradigm in the iterative LLM-based program search process, which ensures continuous improvement. On efficiency side, we additionally introduce multi-task evolution architecture to support parallel reward discovery for diverse MetaBBO approaches. Such parallel process also benefits from knowledge sharing across tasks to accelerate convergence. Empirical results demonstrate that the reward functions discovered by our approach could be helpful for boosting existing MetaBBO works, underscoring the importance of reward design in MetaBBO. We provide READY's project at https://anonymous.4open.science/r/ICML_READY-747F.
Authors: Sidney Bender, Marco Morik
Abstract: Foundation models, despite their robust zero-shot capabilities, remain vulnerable to spurious correlations and 'Clever Hans' strategies. Existing mitigation methods often rely on unavailable group labels or computationally expensive gradient-based adversarial optimization. To address these limitations, we propose Visual Disentangled Diffusion Autoencoders (DiDAE), a novel framework integrating frozen foundation models with disentangled dictionary learning for efficient, gradient-free counterfactual generation directly for the foundation model. DiDAE first edits foundation model embeddings in interpretable disentangled directions of the disentangled dictionary and then decodes them via a diffusion autoencoder. This allows the generation of multiple diverse, disentangled counterfactuals for each factual, much faster than existing baselines, which generate single entangled counterfactuals. When paired with Counterfactual Knowledge Distillation, DiDAE-CFKD achieves state-of-the-art performance in mitigating shortcut learning, improving downstream performance on unbalanced datasets.
Authors: Evandro S. Ortigossa, Guy Lutsker, Eran Segal
Abstract: Real-world multivariate time series can exhibit intricate multi-scale structures, including global trends, local periodicities, and non-stationary regimes, which makes long-horizon forecasting challenging. Although sparse Mixture-of-Experts (MoE) approaches improve scalability and specialization, they typically rely on homogeneous MLP experts that poorly capture the diverse temporal dynamics of time series data. We address these limitations with MoHETS, an encoder-only Transformer that integrates sparse Mixture-of-Heterogeneous-Experts (MoHE) layers. MoHE routes temporal patches to a small subset of expert networks, combining a shared depthwise-convolution expert for sequence-level continuity with routed Fourier-based experts for patch-level periodic structures. MoHETS further improves robustness to non-stationary dynamics by incorporating exogenous information via cross-attention over covariate patch embeddings. Finally, we replace parameter-heavy linear projection heads with a lightweight convolutional patch decoder, improving parameter efficiency, reducing training instability, and allowing a single model to generalize across arbitrary forecast horizons. We validate across seven multivariate benchmarks and multiple horizons, with MoHETS consistently achieving state-of-the-art performance, reducing the average MSE by $12\%$ compared to strong recent baselines, demonstrating effective heterogeneous specialization for long-term forecasting.
Authors: Jinhang Chai, Xuyuan Liu, Elynn Chen, Yujun Yan
Abstract: Learning systems often expand their ambient features or latent representations over time, embedding earlier representations into larger spaces with limited new latent structure. We study transfer learning for structured matrix estimation under simultaneous growth of the ambient dimension and the intrinsic representation, where a well-estimated source task is embedded as a subspace of a higher-dimensional target task. We propose a general transfer framework in which the target parameter decomposes into an embedded source component, low-dimensional low-rank innovations, and sparse edits, and develop an anchored alternating projection estimator that preserves transferred subspaces while estimating only low-dimensional innovations and sparse modifications. We establish deterministic error bounds that separate target noise, representation growth, and source estimation error, yielding strictly improved rates when rank and sparsity increments are small. We demonstrate the generality of the framework by applying it to two canonical problems. For Markov transition matrix estimation from a single trajectory, we derive end-to-end theoretical guarantees under dependent noise. For structured covariance estimation under enlarged dimensions, we provide complementary theoretical analysis in the appendix and empirically validate consistent transfer gains.
Authors: Victor Spitzer, Francois Sanson
Abstract: Decision-focused learning integrates predictive modeling and combinatorial optimization by training models to directly improve decision quality rather than prediction accuracy alone. Differentiating through combinatorial optimization problems represents a central challenge, and recent approaches tackle this difficulty by introducing perturbation-based approximations. In this work, we focus on estimating the objective function coefficients of a combinatorial optimization problem. Our study demonstrates that fluctuations in perturbation intensity occurring during the learning phase can lead to ineffective training, by establishing a theoretical link to the notion of solution stability in combinatorial optimization. We propose addressing this issue by introducing a regularization of the estimated cost vectors which improves the robustness and reliability of the learning process, as demonstrated by extensive numerical experiments.
Authors: Lukas Twist, Shu Yang, Hanqi Yan, Jingzhi Gong, Di Wang, Helen Yannakoudakis, Jie M. Zhang
Abstract: Large Language Models (LLMs) increasingly exhibit strong reasoning abilities, often attributed to their capacity to generate chain-of-thought-style intermediate reasoning. Recent work suggests that exposure to code can further enhance these skills, but existing studies largely treat code as a generic training signal, leaving open the question of which properties of code actually contribute to improved reasoning. To address this gap, we study the structural complexity of code, which captures control flow and compositional structure that may shape how models internalise multi-step reasoning during fine-tuning. We examine two complementary settings: solution-driven complexity, where complexity varies across multiple solutions to the same problem, and problem-driven complexity, where complexity reflects variation in the underlying tasks. Using cyclomatic complexity and logical lines of code to construct controlled fine-tuning datasets, we evaluate a range of open-weight LLMs on diverse reasoning benchmarks. Our findings show that although code can improve reasoning, structural properties strongly determine its usefulness. In 83% of experiments, restricting fine-tuning data to a specific structural complexity range outperforms training on structurally diverse code, pointing to a data-centric path for improving reasoning beyond scaling.
Authors: Ali Hasanzadeh Karkan, Ahmed Ibrahim, Jean-Fran\c{c}ois Frigon, Fran\c{c}ois Leduc-Primeau
Abstract: Massive multiple-input multiple-output (mMIMO) downlink precoding offers high spectral efficiency but remains challenging to deploy in practice because near-optimal algorithms such as the weighted minimum mean squared error (WMMSE) are computationally expensive, and sensitive to SNR and channel-estimation quality, while existing deep learning (DL)-based solutions often lack robustness and require retraining for each deployment site. This paper proposes a plug-and-play precoder (PaPP), a DL framework with a backbone that can be trained for either fully digital (FDP) or hybrid beamforming (HBF) precoding and reused across sites, transmit-power levels, and with varying amounts of channel estimation error, avoiding the need to train a new model from scratch at each deployment. PaPP combines a high-capacity teacher and a compact student with a self-supervised loss that balances teacher imitation and normalized sum-rate, trained using meta-learning domain-generalization and transmit-power-aware input normalization. Numerical results on ray-tracing data from three unseen sites show that the PaPP FDP and HBF models both outperform conventional and deep learning baselines, after fine-tuning with a small set of local unlabeled samples. Across both architectures, PaPP achieves more than 21$\times$ reduction in modeled computation energy and maintains good performance under channel-estimation errors, making it a practical solution for energy-efficient mMIMO precoding.
Authors: Zhiqing Cui, Siru Zhong, Ming Jin, Shirui Pan, Qingsong Wen, Yuxuan Liang
Abstract: Global air quality forecasting grapples with extreme spatial heterogeneity and the poor generalization of existing transductive models to unseen regions. To tackle this, we propose OmniAir, a semantic topology learning framework tailored for global station-level prediction. By encoding invariant physical environmental attributes into generalizable station identities and dynamically constructing adaptive sparse topologies, our approach effectively captures long-range non-Euclidean correlations and physical diffusion patterns across unevenly distributed global networks. We further curate WorldAir, a massive dataset covering over 7,800 stations worldwide. Extensive experiments show that OmniAir achieves state-of-the-art performance against 18 baselines, maintaining high efficiency and scalability with speeds nearly 10 times faster than existing models, while effectively bridging the monitoring gap in data-sparse regions.
Authors: Jonas M\"oller, Erik Imgrund, Thorsten Eisenhofer, Konrad Rieck
Abstract: Machine learning models are routinely deployed on a wide range of computing hardware. Although such hardware is typically expected to produce identical results, differences in its design can lead to small numerical variations during inference. In this work, we show that these variations can be exploited to create backdoors in machine learning models. The core idea is to shape the model's decision function such that it yields different predictions for the same input when executed on different hardware. This effect is achieved by locally moving the decision boundary close to a target input and then refining numerical deviations to flip the prediction on selected hardware. We empirically demonstrate that these hardware-triggered backdoors can be created reliably across common GPU accelerators. Our findings reveal a novel attack vector affecting the use of third-party models, and we investigate different defenses to counter this threat.
Authors: Jinhang Chai, Enpei Zhang, Elynn Chen, Yujun Yan
Abstract: We study online transfer reinforcement learning (RL) in episodic Markov decision processes, where experience from related source tasks is available during learning on a target task. A fundamental difficulty is that task similarity is typically defined in terms of rewards or transitions, whereas online RL algorithms operate on Bellman regression targets. As a result, naively reusing source Bellman updates introduces systematic bias and invalidates regret guarantees. We identify one-step Bellman alignment as the correct abstraction for transfer in online RL and propose re-weighted targeting (RWT), an operator-level correction that retargets continuation values and compensates for transition mismatch via a change of measure. RWT reduces task mismatch to a fixed one-step correction and enables statistically sound reuse of source data. This alignment yields a two-stage RWT $Q$-learning framework that separates variance reduction from bias correction. Under RKHS function approximation, we establish regret bounds that scale with the complexity of the task shift rather than the target MDP. Empirical results in both tabular and neural network settings demonstrate consistent improvements over single-task learning and na\"{i}ve pooling, highlighting Bellman alignment as a model-agnostic transfer principle for online RL.
Authors: Shuangqi Li, Hieu Le, Jingyi Xu, Mathieu Salzmann
Abstract: Training data attribution (TDA) identifies which training examples most influenced a model's prediction. The best-performing TDA methods exploits gradients to define an influence function. To overcome the scalability challenge arising from gradient computation, the most popular strategy is random projection (e.g., TRAK, LoGRA). However, this still faces two bottlenecks when scaling to large training sets and high-quality attribution: \emph{(i)} storing and loading projected per-example gradients for all $N$ training examples, where query latency is dominated by I/O; and \emph{(ii)} forming the $D \times D$ inverse Hessian approximation, which costs $O(D^2)$ memory. Both bottlenecks scale with the projection dimension $D$, yet increasing $D$ is necessary for attribution quality -- creating a quality--scalability tradeoff. We introduce \textbf{LoRIF (Low-Rank Influence Functions)}, which exploits low-rank structures of gradient to address both bottlenecks. First, we store rank-$c$ factors of the projected per-example gradients rather than full matrices, reducing storage and query-time I/O from $O(D)$ to $O(c\sqrt{D})$ per layer per sample. Second, we use truncated SVD with the Woodbury identity to approximate the Hessian term in an $r$-dimensional subspace, reducing memory from $O(D^2)$ to $O(Dr)$. On models from 0.1B to 70B parameters trained on datasets with millions of examples, LoRIF achieves up to 20$\times$ storage reduction and query-time speedup compared to LoGRA, while matching or exceeding its attribution quality. LoRIF makes gradient-based TDA practical at frontier scale.
Authors: Xiaoguang Zhu, Linxiao Gong, Lianlong Sun, Yang Liu, Haoyu Wang, Jing Liu
Abstract: Medical multimodal representation learning aims to integrate heterogeneous data into unified patient representations to support clinical outcome prediction. However, real-world medical datasets commonly contain systematic biases from multiple sources, which poses significant challenges for medical multimodal representation learning. Existing approaches typically focus on effective multimodal fusion, neglecting inherent biased features that affect the generalization ability. To address these challenges, we propose a Dual-Stream Feature Decorrelation Framework that identifies and handles the biases through structural causal analysis introduced by latent confounders. Our method employs a causal-biased decorrelation framework with dual-stream neural networks to disentangle causal features from spurious correlations, utilizing generalized cross-entropy loss and mutual information minimization for effective decorrelation. The framework is model-agnostic and can be integrated into existing medical multimodal learning methods. Comprehensive experiments on MIMIC-IV, eICU, and ADNI datasets demonstrate consistent performance improvements.
Authors: Ahmad Aghapour, Erhan Bayraktar, Ziqing Zhang
Abstract: Diffusion generative models synthesize samples by discretizing reverse-time dynamics driven by a learned score (or denoiser). Existing convergence analyses of diffusion models typically scale at least linearly with the ambient dimension, and sharper rates often depend on intrinsic-dimension assumptions or other geometric restrictions on the target distribution. We develop an alternative, information-theoretic approach to dimension-free convergence that avoids any geometric assumptions. Under mild assumptions on the target distribution, we bound KL divergence between the target and generated distributions by $O(H^2/K)$ (up to endpoint factors), where $H$ is the Shannon entropy and $K$ is the number of sampling steps. Moreover, using a reformulation of the KL divergence, we propose a Loss-Adaptive Schedule (LAS) for efficient discretization of reverse SDE which is lightweight and relies only on the training loss, requiring no post-training heavy computation. Empirically, LAS improves sampling quality over common heuristic schedules.
Authors: Konstantinos P. Panousis, Diego Marcos
Abstract: The widespread adoption of Vision-Language Models (VLMs) across fields has amplified concerns about model interpretability. Distressingly, these models are often treated as black-boxes, with limited or non-existent investigation of their decision making process. Despite numerous post- and ante-hoc interepretability methods, systematic and objective evaluation of the learned representations remains limited, particularly for sparsity-aware methods that are increasingly considered to "induce interpretability". In this work, we focus on Concept Bottleneck Models and investigate how different modeling decisions affect the emerging representations. We introduce the notion of clarity, a measure, capturing the interplay between the downstream performance and the sparsity and precision of the concept representation, while proposing an interpretability assessment framework using datasets with ground truth concept annotations. We consider both VLM- and attribute predictor-based CBMs, and three different sparsity-inducing strategies: per example $\ell_1, \ell_0$ and Bernoulli-based formulations. Our experiments reveal a critical trade-off between flexibility and interpretability, under which a given method can exhibit markedly different behaviors even at comparable performance levels. The code will be made publicly available upon publication.
Authors: Qingshan Wang, Clara C. Wanjura, Florian Marquardt
Abstract: The rapid rise of artificial intelligence has led to an unsustainable growth in energy consumption. This has motivated progress in neuromorphic computing and physics-based training of learning machines as alternatives to digital neural networks. Many theoretical studies focus on simple architectures like all-to-all or densely connected layered networks. However, these may be challenging to realize experimentally, e.g. due to connectivity constraints. In this work, we investigate the performance of the widespread physics-based training method of equilibrium propagation for more realistic architectural choices, specifically, locally connected lattices. We train an XY model and explore the influence of architecture on various benchmark tasks, tracking the evolution of spatially distributed responses and couplings during training. Our results show that sparse networks with only local connections can achieve performance comparable to dense networks. Our findings provide guidelines for further scaling up architectures based on equilibrium propagation in realistic settings.
Authors: Linxiao Gong, Yang Liu, Lianlong Sun, Yulai Bi, Jing Liu, Xiaoguang Zhu
Abstract: Medical multimodal learning faces significant challenges with missing modalities prevalent in clinical practice. Existing approaches assume equal contribution of modality and random missing patterns, neglecting inherent uncertainty in medical data acquisition. In this regard, we propose the Aleatoric Uncertainty Modeling (AUM) that explicitly quantifies unimodal aleatoric uncertainty to address missing modalities. Specifically, AUM models each unimodal representation as a multivariate Gaussian distribution to capture aleatoric uncertainty and enable principled modality reliability quantification. To adaptively aggregate captured information, we develop a dynamic message-passing mechanism within a bipartite patient-modality graph using uncertainty-aware aggregation mechanism. Through this process, missing modalities are naturally accommodated, while more reliable information from available modalities is dynamically emphasized to guide representation generation. Our AUM framework achieves an improvement of 2.26% AUC-ROC on MIMIC-IV mortality prediction and 2.17% gain on eICU, outperforming existing state-of-the-art approaches.
Authors: Yunjia Yang, Runze Li, Yufei Zhang, Haixin Chen
Abstract: Data-based optimization (DBO) offers a promising approach for efficiently optimizing shape for better aerodynamic performance by leveraging a pretrained surrogate model for offline evaluations during iterations. However, DBO heavily relies on the quality of the training database. Samples outside the training distribution encountered during optimization can lead to significant prediction errors, potentially misleading the optimization process. Therefore, incorporating uncertainty quantification into optimization is critical for detecting outliers and enhancing robustness. This study proposes an uncertainty-aware data-based optimization (UA-DBO) framework to monitor and minimize surrogate model uncertainty during DBO. A probabilistic encoder-decoder surrogate model is developed to predict uncertainties associated with its outputs, and these uncertainties are integrated into a model-confidence-aware objective function to penalize samples with large prediction errors during data-based optimization process. The UA-DBO framework is evaluated on two multipoint optimization problems aimed at improving airfoil drag divergence and buffet performance. Results demonstrate that UA-DBO consistently reduces prediction errors in optimized samples and achieves superior performance gains compared to original DBO. Moreover, compared to multipoint optimization based on full computational simulations, UA-DBO offers comparable optimization effectiveness while significantly accelerating optimization speed.
Authors: Qianwei Yang, Dong Xu, Zhangfan Yang, Sisi Yuan, Zexuan Zhu, Jianqiang Li, Junkai Ji
Abstract: Drug discovery can be viewed as a combinatorial search over an immense chemical space, motivating the development of deep generative models for de novo molecular design. Among these, GPT-based molecular language models (MLM) have shown strong molecular design performance by learning chemical syntax and semantics from large-scale data. However, existing MLMs face two fundamental limitations: they inadequately capture the graph-structured nature of molecules when formulated as next-token prediction problems, and they typically lack explicit mechanisms for target-aware generation. Here, we propose SoftMol, a unified framework that co-designs molecular representation, model architecture, and search strategy for target-aware molecular generation. SoftMol introduces soft fragments, a rule-free block representation of SMILES that enables diffusion-native modeling, and develops SoftBD, the first block-diffusion molecular language model that combines local bidirectional diffusion with autoregressive generation under molecular structural constraints. To favor generated molecules with high drug-likeness and synthetic accessibility, SoftBD is trained on a carefully curated dataset named ZINC-Curated. SoftMol further integrates a gated Monte Carlo tree search to assemble fragments in a target-aware manner. Experimental results show that, compared with current state-of-the-art models, SoftMol achieves 100% chemical validity, improves binding affinity by 9.7%, yields a 2-3x increase in molecular diversity, and delivers a 6.6x speedup in inference efficiency. Code is available at https://github.com/szu-aicourse/softmol
Authors: Shiqi Fan, Quanming Yao, Hongyi Nie, Wentao Ma, Zhen Wang, Wen Hua
Abstract: Temporal knowledge graph reasoning (TKGR) aims to predict future events by inferring missing entities with dynamic knowledge structures. Existing LLM-based reasoning methods prioritize contextual over structural relations, struggling to extract relevant subgraphs from dynamic graphs. This limits structural information understanding, leading to unstructured, hallucination-prone inferences especially with temporal inconsistencies. To address this problem, we propose IGETR (Integration of Graph and Editing-enhanced Temporal Reasoning), a hybrid reasoning framework that combines the structured temporal modeling capabilities of Graph Neural Networks (GNNs) with the contextual understanding of LLMs. IGETR operates through a three-stage pipeline. The first stage aims to ground the reasoning process in the actual data by identifying structurally and temporally coherent candidate paths through a temporal GNN, ensuring that inference starts from reliable graph-based evidence. The second stage introduces LLM-guided path editing to address logical and semantic inconsistencies, leveraging external knowledge to refine and enhance the initial paths. The final stage focuses on integrating the refined reasoning paths to produce predictions that are both accurate and interpretable. Experiments on standard TKG benchmarks show that IGETR achieves state-of-the-art performance, outperforming strong baselines with relative improvements of up to 5.6% on Hits@1 and 8.1% on Hits@3 on the challenging ICEWS datasets. Additionally, we execute ablation studies and additional analyses confirm the effectiveness of each component.
Authors: Ciaran Bench, Vivek Desai, Carlijn Roozemond, Ruben van Engen, Spencer A. Thomas
Abstract: Feature embeddings acquired from pretrained models are widely used in medical applications of deep learning to assess the characteristics of datasets; e.g. to determine the quality of synthetic, generated medical images. The Fr\'{e}chet Inception Distance (FID) is one popular synthetic image quality metric that relies on the assumption that the characteristic features of the data can be detected and encoded by an InceptionV3 model pretrained on ImageNet1K (natural images). While it is widely known that this makes it less effective for applications involving medical images, the extent to which the metric fails to capture meaningful differences in image characteristics is not obviously known. Here, we use Monte Carlo dropout to compute the predictive variance in the FID as well as a supplemental estimate of the predictive variance in the feature embedding model's latent representations. We show that the magnitudes of the predictive variances considered exhibit varying degrees of correlation with the extent to which test inputs (ImageNet1K validation set augmented at various strengths, and other external datasets) are out-of-distribution relative to its training data, providing some insight into the effectiveness of their use as indicators of the trustworthiness of the FID.
Authors: Andrew Millard, Joshua Murphy, Peter Green, Simon Maskell
Abstract: Bayesian inference allows us to define a posterior distribution over the weights of a generic neural network (NN). Exact posteriors are usually intractable, in which case approximations can be employed. One such approximation - variational inference - is computationally efficient when using mini-batch stochastic gradient descent as subsets of the data are used for likelihood and gradient evaluations, though the approach relies on the selection of a variational distribution which sufficiently matches the form of the posterior. Particle-based methods such as Markov chain Monte Carlo and Sequential Monte Carlo (SMC) do not assume a parametric family for the posterior by typically require higher computational cost. These sampling methods typically use the full-batch of data for likelihood and gradient evaluations, which contributes to this computational expense. We explore several methods of gradually introducing more mini-batches of data (data annealing) into likelihood and gradient evaluations of an SMC sampler. We find that we can achieve up to $6\times$ faster training with minimal loss in accuracy on benchmark image classification problems using NNs.
Authors: Jian Gao, Yiwei Zou, Abhishek Pradhan, Wenhao Huang, Yumin Su, Kaiyuan Yang, Xuan Zhang
Abstract: Discovering superior circuit topologies requires navigating an exponentially large design space-a challenge traditionally reserved for human experts. Existing AI methods either select from predefined templates or generate novel topologies at a limited scale without rigorous verification, leaving large-scale performance-driven discovery underexplored. We present PowerGenie, a framework for automated discovery of higher-performance reconfigurable power converters at scale. PowerGenie introduces: (1) an automated analytical framework that determines converter functionality and theoretical performance limits without component sizing or SPICE simulation, and (2) an evolutionary finetuning method that co-evolves a generative model with its training distribution through fitness selection and uniqueness verification. Unlike existing methods that suffer from mode collapse and overfitting, our approach achieves higher syntax validity, function validity, novelty rate, and figure-of-merit (FoM). PowerGenie discovers a novel 8-mode reconfigurable converter with 23% higher FoM than the best training topology. SPICE simulations confirm average absolute efficiency gains of 10% across 8 modes and up to 17% at a single mode. Code is available at https://github.com/xz-group/PowerGenie.
Authors: Yunyang Li, Lin Huang, Luojia Xia, Wenhe Zhang, Mark Gerstein
Abstract: Generative models for 3D molecular conformations must respect Euclidean symmetries and concentrate probability mass on thermodynamically favorable, mechanically stable structures. However, E(3)-equivariant diffusion models often reproduce biases from semi-empirical training data rather than capturing the equilibrium distribution of a high-fidelity Hamiltonian. While physics-based guidance can correct this, it faces two computational bottlenecks: expensive quantum-chemical evaluations (e.g., DFT) and the need to repeat such queries at every sampling step. We present Elign, a post-training framework that amortizes both costs. First, we replace expensive DFT evaluations with a faster, pretrained foundational machine-learning force field (MLFF) to provide physical signals. Second, we eliminate repeated run-time queries by shifting physical steering to the training phase. To achieve the second amortization, we formulate reverse diffusion as a reinforcement learning problem and introduce Force--Energy Disentangled Group Relative Policy Optimization (FED-GRPO) to fine-tune the denoising policy. FED-GRPO includes a potential-based energy reward and a force-based stability reward, which are optimized and group-normalized independently. Experiments show that Elign generates conformations with lower gold-standard DFT energies and forces, while improving stability. Crucially, inference remains as fast as unguided sampling, since no energy evaluations are required during generation.
Authors: Fernando Palafox, Jingqi Li, Jesse Milzman, David Fridovich-Keil
Abstract: An agent operating in an unknown dynamical system must learn its dynamics from observations. Active information gathering accelerates this learning, but existing methods derive bespoke costs for specific modeling choices: dynamics models, belief update procedures, observation models, and planners. We present a unifying framework that decouples these choices from the information-gathering cost by explicitly exposing the causal dependencies between parameters, beliefs, and controls. Using this framework, we derive a general information-gathering cost based on Massey's directed information that assumes only Markov dynamics with additive noise and is otherwise agnostic to modeling choices. We prove that the mutual information cost used in existing literature is a special case of our cost. Then, we leverage our framework to establish an explicit connection between the mutual information cost and information gain in linearized Bayesian estimation, thereby providing theoretical justification for mutual information-based active learning approaches. Finally, we illustrate the practical utility of our framework through experiments spanning linear, nonlinear, and multi-agent systems.
Authors: Zuyuan Zhang, Mahdi Imani, Tian Lan
Abstract: Real-world reinforcement learning is often \emph{nonstationary}: rewards and dynamics drift, accelerate, oscillate, and trigger abrupt switches in the optimal action. Existing theory often represents nonstationarity with coarse-scale models that measure \emph{how much} the environment changes, not \emph{how} it changes locally -- even though acceleration and near-ties drive tracking error and policy chattering. We take a geometric view of nonstationary discounted Markov Decision Processes (MDPs) by modeling the environment as a differentiable homotopy path and tracking the induced motion of the optimal Bellman fixed point. This yields a length--curvature--kink signature of intrinsic complexity: cumulative drift, acceleration/oscillation, and action-gap-induced nonsmoothness. We prove a solver-agnostic path-integral stability bound and derive gap-safe feasible regions that certify local stability away from switch regimes. Building on these results, we introduce \textit{Homotopy-Tracking RL (HT-RL)} and \textit{HT-MCTS}, lightweight wrappers that estimate replay-based proxies of length, curvature, and near-tie proximity online and adapt learning or planning intensity accordingly. Experiments show improved tracking and dynamic regret over matched static baselines, with the largest gains in oscillatory and switch-prone regimes.
Authors: Meng Cao, Jiexi Liu, Songcan Chen
Abstract: Imbalanced Domain Generalization (IDG) focuses on mitigating both domain and label shifts, both of which fundamentally shape the model's decision boundaries, particularly under heterogeneous long-tailed distributions across domains. Despite its practical significance, it remains underexplored, primarily due to the technical complexity of handling their entanglement and the paucity of theoretical foundations. In this paper, we begin by theoretically establishing the generalization bound for IDG, highlighting the role of posterior discrepancy and decision margin. This bound motivates us to focus on directly steering decision boundaries, marking a clear departure from existing methods. Subsequently, we technically propose a novel Negative-Dominant Contrastive Learning (NDCL) for IDG to enhance discriminability while enforce posterior consistency across domains. Specifically, inter-class decision-boundary separation is enhanced by placing greater emphasis on negatives as the primary signal in our contrastive learning, naturally amplifying gradient signals for minority classes to avoid the decision boundary being biased toward majority classes. Meanwhile, intra-class compactness is encouraged through a re-weighted cross-entropy strategy, and posterior consistency across domains is enforced through a prediction-central alignment strategy. Finally, rigorous yet challenging experiments on benchmarks validate the effectiveness of our NDCL. The code is available at https://github.com/Alrash/NDCL.
Authors: Anderson de Andrade, Alon Harell, Ivan V. Baji\'c
Abstract: Transformers achieve superior performance on many tasks, but impose heavy compute and memory requirements during inference. This inference can be made more efficient by partitioning the process across multiple devices, which, in turn, requires compressing its intermediate representations. In this work, we introduce a principled rate-distortion-based framework for lossy compression that learns compact encodings that explicitly trade off bitrate against accuracy. Experiments on language benchmarks show that the proposed codec achieves substantial savings with improved accuracy in some cases, outperforming more complex baseline methods. We characterize and analyze the rate-distortion performance of transformers, offering a unified lens for understanding performance in representation coding. This formulation extends information-theoretic concepts to define the gap between rate and entropy, and derive some of its bounds. We further develop probably approximately correct (PAC)-style bounds for estimating this gap. For different architectures and tasks, we empirically demonstrate that their rates are driven by these bounds, adding to the explainability of the formulation.
Authors: Dongxuan Zhu, Ly Tran Ho Khanh, Andy Yat-Ming Cheung, Man-Chung Yue, Viet Anh Nguyen
Abstract: Language models often default to a narrow set of high-probability outputs, leaving their generation paths homogeneous and prone to mode collapse. Sampling-based strategies inject randomness but still struggle to guarantee diversity across multiple concurrent generation runs. We address this limitation by introducing STARS ($\textbf{St}$iefel-based $\textbf{A}$ctivation Steering for Diverse $\textbf{R}$ea$\textbf{S}$oning), a training-free, inference-time intervention method that transforms activation steering into an exploration engine. At each token, STARS collects the hidden activations of concurrent generation runs and optimizes multiple additive steering directions jointly on the Stiefel manifold. STARS maximizes the geometric volume of the steered activations, while the Stiefel manifold induces orthogonality of the steering interventions. This formulation explicitly promotes divergent activation vectors of concurrent generation runs, and implicitly promotes divergent generation trajectories. This manifold optimization formulation can be solved using a Riemannian gradient descent algorithm with convergence guarantees, but this algorithm is too time-consuming for real-time inference. To guarantee low latency, we further design a lightweight one-step update with an aggressive, closed-form stepsize. For test case generation and scientific discovery benchmarks, STARS consistently outperforms standard sampling methods, achieving greater diversity without sacrificing qualitative performance.
Authors: Sergi Masip, Gido M. van de Ven, Javier Ferrando, Tinne Tuytelaars
Abstract: Catastrophic forgetting in continual learning is often measured at the performance or last-layer representation level, overlooking the underlying mechanisms. We introduce a mechanistic framework that offers a geometric interpretation of catastrophic forgetting as the result of transformations to the encoding of individual features. These transformations can lead to forgetting by reducing the allocated capacity of features (worse representation) and disrupting their readout by downstream computations. Analysis of a tractable model formalizes this view, allowing us to identify best- and worst-case scenarios. Through experiments on this model, we empirically test our formal analysis and highlight the detrimental effect of depth. Finally, we demonstrate how our framework can be used in the analysis of practical models through the use of Crosscoders. We present a case study of a Vision Transformer trained on sequential CIFAR-10. Our work provides a new, feature-centric vocabulary for continual learning.
Authors: Robert Istvan Busa-Fekete, Julian Zimmert, Anne Xiangyi Zheng, Claudio Gentile, Andras Gyorgy
Abstract: The quality of machine learning models depends heavily on their training data. Selecting high-quality, diverse training sets for large language models (LLMs) is a difficult task, due to the lack of cheap and reliable quality metrics. While querying existing LLMs for document quality is common, this is not scalable to the large number (billions) of documents used in training. Instead, practitioners often use classifiers trained on sparse quality signals. In this paper, we propose a text-embedding-based hierarchical clustering approach that adaptively selects the documents to be evaluated by the LLM to estimate cluster quality. We prove that our method is query efficient: under the assumption that the hierarchical clustering contains a subtree such that each leaf cluster in the tree is pure enough (i.e., it mostly contains either only good or only bad documents), with high probability, the method can correctly predict the quality of each document after querying a small number of documents. The number of such documents is proportional to the size of the smallest subtree with (almost) pure leaves, without the algorithm knowing this subtree in advance. Furthermore, in a comprehensive experimental study, we demonstrate the benefits of our algorithm compared to other classifier-based filtering methods.
Authors: Chengyi Cai, Zesheng Ye, Peike Li, Bo Han, Jianzhong Qi, Feng Liu
Abstract: Unlearning in Multimodal Large Language Models (MLLMs) prevents the model from revealing private information when queried about target images. Existing MLLM unlearning methods largely adopt approaches developed for LLMs. They treat all answer tokens uniformly, disregarding their varying importance in the unlearning process. Moreover, these methods focus exclusively on the language modality, disregarding visual cues that indicate key tokens in answers. In this paper, after formulating the problem of unlearning in multimodal question answering for MLLMs, we propose Visual-Guided Key-Token Regularization (ViKeR). We leverage irrelevant visual inputs to predict ideal post-unlearning token-level distributions and use these distributions to regularize the unlearning process, thereby prioritizing key tokens. Further, we define key tokens in unlearning via information entropy and discuss ViKeR's effectiveness through token-level gradient reweighting, which amplifies updates on key tokens. Experiments on MLLMU and CLEAR benchmarks demonstrate that our method effectively performs unlearning while mitigating forgetting and maintaining response coherence.
Authors: Haoran Tang, Rajiv Khanna
Abstract: Most LLM unlearning methods aim to approximate retrain-from-scratch behaviors with minimal distribution shift, often via alignment-style objectives defined in the prediction space. While effective at reducing forgotten content generation, such approaches may act as suppression: forgotten concepts can persist in representations and remain entangled with retained knowledge. We introduce CLReg, a contrastive representation regularizer that identifies forget features while pushing them away from retain features, explicitly reducing forget-retain interference with minimal shifts on retain features. We provide first theoretical insights that relate representation shaping to entanglement reduction. Across unlearning benchmarks and LLMs of different sizes, CLReg decreases forget-retain representation entanglement that facilitates mainstream unlearning methods without positing extra privacy risks, inspiring future work that reshapes the representation space to remove forget concepts.
Authors: Zhengyan Huan, Camila Pazos, Martin Klassen, Vincent Croft, Pierre-Hugues Beauchemin, Shuchin Aeron
Abstract: We introduce a new multivariate statistical problem that we refer to as the Ensemble Inverse Problem (EIP). The aim of EIP is to invert for an ensemble that is distributed according to the pushforward of a prior under a forward process. In high energy physics (HEP), this is related to a widely known problem called unfolding, which aims to reconstruct the true physics distribution of quantities, such as momentum and angle, from measurements that are distorted by detector effects. In recent applications, the EIP also arises in full waveform inversion (FWI) and inverse imaging with unknown priors. We propose non-iterative inference-time methods that construct posterior samplers based on a new class of conditional generative models, which we call ensemble inverse generative models. For the posterior modeling, these models additionally use the ensemble information contained in the observation set on top of single measurements. Unlike existing methods, our proposed methods avoid explicit and iterative use of the forward model at inference time via training across several sets of truth-observation pairs that are consistent with the same forward model, but originate from a wide range of priors. We demonstrate that this training procedure implicitly encodes the likelihood model. The use of ensemble information helps posterior inference and enables generalization to unseen priors. We benchmark the proposed method on several synthetic and real datasets in inverse imaging, HEP, and FWI. The codes are available at https://github.com/ZhengyanHuan/The-Ensemble-Inverse-Problem--Applications-and-Methods.
URLs: https://github.com/ZhengyanHuan/The-Ensemble-Inverse-Problem--Applications-and-Methods.
Authors: Chengyi Cai, Zesheng Ye, Jiangchao Yao, Jianzhong Qi, Bo Han, Xiaolu Zhang, Feng Liu, Jun Zhou
Abstract: In large language model (LLM) unlearning, private information is required to be removed. Task arithmetic unlearns by subtracting a specific task vector (TV)--defined as the parameter difference between a privacy-information-tuned model and the original model. While efficient, it can cause over-forgetting by disrupting parameters essential for retaining other information. Motivated by the observation that each parameter exhibits different importance for forgetting versus retention, we propose a per-parameter task arithmetic (PerTA) mechanism to rescale the TV, allowing per-parameter adjustment. These weights quantify the relative importance of each parameter for forgetting versus retention, estimated via gradients (i.e., PerTA-grad) or the diagonal Fisher information approximation (i.e., PerTA-fisher). Moreover, we discuss the effectiveness of PerTA, extend it to a more general form, and provide further analysis. Extensive experiments demonstrate that PerTA consistently improves upon standard TV, and in many cases surpasses widely used training-based unlearning methods in both forgetting effectiveness and overall model utility. By retaining the efficiency of task arithmetic while mitigating over-forgetting, PerTA offers a principled and practical framework for LLM unlearning.
Authors: Ehsan Mirafzali, Sanjit Shashi, Sanya Murdeshwar, Edgar Shaghoulian, Daniele Venturi, Razvan Marinescu
Abstract: We present a framework for generative machine learning that leverages the holographic principle of quantum gravity, or to be more precise its manifestation as the anti-de Sitter/conformal field theory (AdS/CFT) correspondence, with techniques for deep learning and transport theory. Our proposal is to represent the flow of data from a base distribution to some learned distribution using the bulk-to-boundary mapping of scalar fields in AdS. In the language of machine learning, we are representing and augmenting the flow-matching algorithm with AdS physics. Using a checkerboard toy dataset and MNIST, we find that our model achieves faster and higher quality convergence than comparable physics-free flow-matching models. Our method provides a physically interpretable version of flow matching. More broadly, it establishes the utility of AdS physics and geometry in the development of novel paradigms in generative modeling.
Authors: Xiaolong Zhang, Jianwei Zhang, Xubo Song
Abstract: Quantifying degrees of fusion and separability between data groups in representation space is a fundamental problem in representation learning, particularly under domain shift. A meaningful metric should capture fusion-altering factors like geometric displacement between representation groups, whose variations change the extent of fusion, while remaining invariant to fusion-preserving factors such as global scaling and sampling-induced layout changes, whose variations do not. Existing distributional distance metrics conflate these factors, leading to measures that are not informative of the true extent of fusion between data groups. We introduce Cross-Fusion Distance (CFD), a principled measure that isolates fusion-altering geometry while remaining robust to fusion-preserving variations, with linear computational complexity. We characterize the invariance and sensitivity properties of CFD theoretically and validate them in controlled synthetic experiments. For practical utility on real-world datasets with domain shift, CFD aligns more closely with downstream generalization degradation than commonly used alternatives. Overall, CFD provides a theoretically grounded and interpretable distance measure for representation learning.
Authors: Mehmet Ozgur Turkoglu, Dominik J. M\"uhlematter, Alexander Becker, Konrad Schindler, Helge Aasen
Abstract: Foundation models have become a dominant paradigm in machine learning, achieving remarkable performance across diverse tasks through large-scale pretraining. However, these models often yield overconfident, uncalibrated predictions. The standard approach to quantifying epistemic uncertainty, training an ensemble of independent models, incurs prohibitive computational costs that scale linearly with ensemble size, making it impractical for large foundation models. We propose Singular Value Ensemble (SVE), a parameter-efficient implicit ensemble method that builds on a simple, but powerful core assumption: namely, that the singular vectors of the weight matrices constitute meaningful subspaces of the model's knowledge. Pretrained foundation models encode rich, transferable information in their weight matrices. If the singular vectors are indeed meaningful (orthogonal) "knowledge directions". To obtain a model ensemble, we modulate only how strongly each direction contributes to the output. Rather than learning entirely new parameters, we freeze the singular vectors and only train per-member singular values that rescale the contribution of each direction in that shared knowledge basis. Ensemble diversity emerges naturally as stochastic initialization and random sampling of mini-batches during joint training cause different members to converge to different combinations of the same underlying knowledge. SVE achieves uncertainty quantification comparable to explicit deep ensembles while increasing the parameter count of the base model by less than 1%, making principled uncertainty estimation accessible in resource-constrained settings. We validate SVE on NLP and vision tasks with various different backbones and show that it improves calibration while maintaining predictive accuracy.
Authors: Jae-Won Chung, Ruofan Wu, Jeff J. Ma, Mosharaf Chowdhury
Abstract: Energy is now a critical ML computing resource. While measuring energy consumption and observing trends is a valuable first step, accurately understanding and diagnosing why those differences occur is crucial for optimization. To that end, we begin by presenting a large-scale measurement study of inference time and energy across the generative AI landscape with 46 models, 7 tasks, and 1,858 different configurations on NVIDIA H100 and B200 GPUs. Our empirical findings span order-of-magnitude variations: LLM task type can lead to 25$\times$ energy differences, video generation sometimes consumes more than 100$\times$ the energy of images, and GPU utilization differences can result in 3--5$\times$ energy differences. Based on our observations, we present a framework for reasoning about the underlying mechanisms that govern time and energy consumption. The essence is that time and energy are determined by latent metrics like memory and utilization, which are in turn affected by various factors across the algorithm, software, and hardware layers. Our framework also extends directly to throughput per watt, a critical metric for power-constrained datacenters.
Authors: Enyi Jiang, Yibo Jacky Zhang, Yinglun Xu, Andreas Haupt, Nancy Amato, Sanmi Koyejo
Abstract: Learning from human feedback typically relies on preference optimization that constrains policy updates through token-level regularization. However, preference optimization for language models is particularly challenging because token-space similarity does not imply semantic or behavioral similarity. To address this challenge, we leverage latent-space regularization for language model preference optimization. We introduce GANPO, which achieves latent-space regularization by penalizing divergence between the internal representations of a policy model and a reference model. Given that latent representations are not associated with explicit probability densities, we adopt an adversarial approach inspired by GANs to minimize latent-space divergence. We integrate GANPO as a regularizer into existing offline preference optimization objectives. Experiments across multiple model architectures and tasks show consistent improvements from latent-space regularization. Further, by comparing GANPO-induced inferential biases with those from token-level regularization, we find that GANPO provides more robust structural feedback under distributional shift and noise while maintaining comparable downstream performance with minor computational overhead.
Authors: Chuanyang Zheng, Jiankai Sun, Yihang Gao, Chi Wang, Yuehao Wang, Jing Xiong, Liliang Ren, Bo Peng, Qingmei Wang, Xiaoran Shang, Mac Schwager, Anderson Schneider, Yuriy Nevmyvaka, Xiaodong Liu
Abstract: The placement of normalization layers, specifically Pre-Norm and Post-Norm, remains an open question in Transformer architecture design. In this work, we rethink these approaches through the lens of manifold optimization, interpreting the outputs of the Feed-Forward Network (FFN) and attention layers as update directions in optimization. Building on this perspective, we introduce GeoNorm, a novel method that replaces standard normalization with geodesic updates on the manifold. Furthermore, analogous to learning rate schedules, we propose a layer-wise update decay for the FFN and attention components. Comprehensive experiments demonstrate that GeoNorm consistently outperforms existing normalization methods in Transformer models. Crucially, GeoNorm can be seamlessly integrated into standard Transformer architectures, achieving performance improvements with negligible additional computational cost.
Authors: Yudong Luo, Erick Delage
Abstract: Optimizing Conditional Value-at-risk (CVaR) using policy gradient (a.k.a CVaR-PG) faces significant challenges of sample inefficiency. This inefficiency stems from the fact that it focuses on tail-end performance and overlooks many sampled trajectories. We address this problem by augmenting CVaR with an expected quantile term. Quantile optimization admits a dynamic programming formulation that leverages all sampled data, thus improves sample efficiency. This does not alter the CVaR objective since CVaR corresponds to the expectation of quantile over the tail. Empirical results in domains with verifiable risk-averse behavior show that our algorithm within the Markovian policy class substantially improves upon CVaR-PG and consistently outperforms other existing methods.
Authors: Harvey Chen, Nicolas Zilberstein, Santiago Segarra
Abstract: We introduce Prior-Informed Flow Matching (PIFM), a conditional flow model for graph reconstruction. Reconstructing graphs from partial observations remains a key challenge; classical embedding methods often lack global consistency, while modern generative models struggle to incorporate structural priors. PIFM bridges this gap by integrating embedding-based priors with continuous-time flow matching. Grounded in a permutation equivariant version of the distortion-perception theory, our method first uses a prior, such as graphons or GraphSAGE/node2vec, to form an informed initial estimate of the adjacency matrix based on local information. It then applies rectified flow matching to refine this estimate, transporting it toward the true distribution of clean graphs and learning a global coupling. Experiments on different datasets demonstrate that PIFM consistently enhances classical embeddings, outperforming them and state-of-the-art generative baselines in reconstruction accuracy.
Authors: Shuqi Ke, Giulia Fanti
Abstract: Can a small amount of verified goal information steer the expensive self-supervised pretraining of foundation models? Standard pretraining optimizes a fixed proxy objective (e.g., next-token prediction), which can misallocate compute away from downstream capabilities of interest. We introduce V-Pretraining: a value-based, modality-agnostic method for controlled continued pretraining in which a lightweight task designer reshapes the pretraining task to maximize the value of each gradient step. For example, consider self-supervised learning (SSL) with sample augmentation. The V-Pretraining task designer selects pretraining tasks (e.g., augmentations) for which the pretraining loss gradient is aligned with a gradient computed over a downstream task (e.g., image segmentation). This helps steer pretraining towards relevant downstream capabilities. Notably, the pretrained model is never updated on downstream task labels; they are used only to shape the pretraining task. Under matched learner update budgets, V-Pretraining of 0.5B--7B language models improves reasoning (GSM8K test Pass@1) by up to 18% relative over standard next-token prediction using only 12% of GSM8K training examples as feedback. In vision SSL, we improve the state-of-the-art results on ADE20K by up to 1.07 mIoU and reduce NYUv2 RMSE while improving ImageNet linear accuracy, and we provide pilot evidence of improved token efficiency in continued pretraining.
Authors: Abdullah Tasim, Wei Sun
Abstract: Accurate reconstruction of atmospheric wind fields is essential for applications such as weather forecasting, hazard prediction, and wind energy assessment, yet conventional instruments leave spatio-temporal gaps within the lower atmospheric boundary layer. Unmanned aircraft systems (UAS) provide flexible in situ measurements, but individual platforms sample wind only along their flight trajectories, limiting full wind-field recovery. This study presents a framework for reconstructing four-dimensional atmospheric wind fields using measurements obtained from a coordinated UAS swarm. A synthetic turbulence environment and high-fidelity multirotor simulation are used to generate training and evaluation data. Local wind components are estimated from UAS dynamics using a bidirectional long short-term memory network (Bi-LSTM) and assimilated into a physics-informed neural network (PINN) to reconstruct a continuous wind field in space and time. For local wind estimation, the bidirectional LSTM achieves root-mean-square errors (RMSE) of 0.064 and 0.062 m/s for the north and east components in low-wind conditions, increasing to 0.122 to 0.129 m/s under moderate winds and 0.271 to 0.273 m/s in high-wind conditions, while the vertical component exhibits higher error, with RMSE values of 0.029 to 0.091 m/s. The physics-informed reconstruction recovers the dominant spatial and temporal structure of the wind field up to 1000 m altitude while preserving mean flow direction and vertical shear. Under moderate wind conditions, the reconstructed mean wind field achieves an overall RMSE between 0.118 and 0.154 m/s across evaluated UAS configurations, with the lowest error obtained using a five-UAS swarm. These results demonstrate that coordinated UAS measurements enable accurate and scalable four-dimensional wind-field reconstruction without dedicated wind sensors or fixed infrastructure.
Authors: Winfried Ripken, Michael Plainer, Gregor Lied, Thorben Frank, Oliver T. Unke, Stefan Chmiela, Frank No\'e, Klaus Robert M\"uller
Abstract: Simulating the long-time evolution of Hamiltonian systems is limited by the small timesteps required for stable numerical integration. To overcome this constraint, we introduce a framework to learn Hamiltonian Flow Maps by predicting the mean phase-space evolution over a chosen time span $\Delta t$, enabling stable large-timestep updates far beyond the stability limits of classical integrators. To this end, we impose a Mean Flow consistency condition for time-averaged Hamiltonian dynamics. Unlike prior approaches, this allows training on independent phase-space samples without access to future states, avoiding expensive trajectory generation. Validated across diverse Hamiltonian systems, our method in particular improves upon molecular dynamics simulations using machine-learned force fields (MLFF). Our models maintain comparable training and inference cost, but support significantly larger integration timesteps while trained directly on widely-available trajectory-free MLFF datasets.
Authors: Leonard Papenmeier, Petru Tighineanu
Abstract: Multi-objective optimization aims to solve problems with competing objectives, often with only black-box access to a problem and a limited budget of measurements. In many applications, historical data from related optimization tasks is available, creating an opportunity for meta-learning to accelerate the optimization. Bayesian optimization, as a promising technique for black-box optimization, has been extended to meta-learning and multi-objective optimization independently, but methods that simultaneously address both settings - meta-learned priors for multi-objective Bayesian optimization - remain largely unexplored. We propose SMOG, a scalable and modular meta-learning model based on a multi-output Gaussian process that explicitly learns correlations between objectives. SMOG builds a structured joint Gaussian process prior across meta- and target tasks and, after conditioning on metadata, yields a closed-form target-task prior augmented by a flexible residual multi-output kernel. This construction propagates metadata uncertainty into the target surrogate in a principled way. SMOG supports hierarchical, parallel training: meta-task Gaussian processes are fit once and then cached, achieving linear scaling with the number of meta-tasks. The resulting surrogate integrates seamlessly with standard multi-objective Bayesian optimization acquisition functions.
Authors: Ziming Dong, Hardik Sharma, Evan O'Toole, Jaya Prakash Champati, Kui Wu
Abstract: Large Language Models (LLMs) deliver state-of-the-art performance on complex reasoning tasks, but their inference costs limit deployment at scale. Small Language Models (SLMs) offer dramatic cost savings yet lag substantially in accuracy. Existing approaches - routing and cascading - treat the LLM as an all-or-nothing resource: either the query bypasses the LLM entirely, or the LLM generates a complete response at full cost. We introduce LLM Shepherding, a framework that requests only a short prefix (a hint) from the LLM and provides it to SLM. This simple mechanism is surprisingly effective for math and coding tasks: even hints comprising 10-30% of the full LLM response improve SLM accuracy significantly. Shepherding generalizes both routing and cascading, and it achieves lower cost under oracle decision-making. We develop a two-stage predictor that jointly determines whether a hint is needed and how many tokens to request. On the widely-used mathematical reasoning (GSM8K, CNK12) and code generation (HumanEval, MBPP) benchmarks, Shepherding reduces costs by 42-94% relative to LLM-only inference. Compared to state-of-the-art routing and cascading baselines, shepherding delivers up to 2.8x cost reduction while matching accuracy. To our knowledge, this is the first work to exploit token-level budget control for SLM-LLM collaboration.
Authors: Gloria Felicia (University of Virginia), Michael Eniolade (University of the Cumberlands), Jinfeng He (Cornell University), Zitha Sasindran (Indian Institute of Science Bangalore), Hemant Kumar (University of Arizona), Milan Hussain Angati (California State University Northridge), Sandeep Bandarupalli (University of Cincinnati)
Abstract: Existing agent safety benchmarks report binary accuracy, conflating early intervention with post-mortem analysis. A detector that flags a violation at step 8 enables intervention; one that reports it at step 48 provides only forensic value. This distinction is critical, yet current benchmarks cannot measure it. We introduce StepShield, the first benchmark to evaluate when violations are detected, not just whether. StepShield contains 9,213 code agent trajectories, including 1,278 meticulously annotated training pairs and a 7,935-trajectory test set with a realistic 8.1% rogue rate. Rogue behaviors are grounded in real-world security incidents across six categories. We propose three novel temporal metrics: Early Intervention Rate (EIR), Intervention Gap, and Tokens Saved. Surprisingly, our evaluation reveals that an LLM-based judge achieves 59% EIR while a static analyzer achieves only 26%, a 2.3x performance gap that is entirely invisible to standard accuracy metrics. We further show that early detection has direct economic benefits: our cascaded HybridGuard detector reduces monitoring costs by 75% and projects to $108M in cumulative savings over five years at enterprise scale. By shifting the focus of evaluation from whether to when, StepShield provides a new foundation for building safer and more economically viable AI agents. The code and data are released under an Apache 2.0 license.
Authors: Shenghao Yang, Zhichao Wang, Oleg Balabanov, N. Benjamin Erichson, Michael W. Mahoney
Abstract: Matrix functions such as square root, inverse roots, and orthogonalization play a central role in preconditioned gradient methods for neural network training. This has motivated the development of iterative algorithms that avoid explicit eigendecompositions and rely primarily on matrix multiplications, making them well suited for modern GPU accelerators. We present PRISM (Polynomial-fitting and Randomized Iterative Sketching for Matrix functions computation), a general framework for accelerating iterative algorithms for computing matrix functions. PRISM combines adaptive polynomial approximation with randomized sketching: at each iteration, it fits a polynomial surrogate to the current spectrum via a sketched least-squares problem, adapting to the instance at hand with minimal overhead. We apply PRISM to accelerate Newton-Schulz-like iterations for matrix square roots and orthogonalization, which are core primitives in machine learning. Unlike prior methods, PRISM requires no explicit spectral bounds or singular value estimates; and it adapts automatically to the evolving spectrum. Empirically, PRISM accelerates training when integrated into Shampoo and Muon optimizers.
Authors: Daniel Stein, Shaoyi Huang, Rolf Drechsler, Bing Li, Grace Li Zhang
Abstract: Neural networks have been successfully applied in various resource-constrained edge devices, where usually central processing units (CPUs) instead of graphics processing units exist due to limited power availability. State-of-the-art research still focuses on efficiently executing enormous numbers of multiply-accumulate (MAC) operations. However, CPUs themselves are not good at executing such mathematical operations on a large scale, since they are more suited to execute control flow logic, i.e., computer algorithms. To enhance the computation efficiency of neural networks on CPUs, in this paper, we propose to convert them into logic flows for execution. Specifically, neural networks are first converted into equivalent decision trees, from which decision paths with constant leaves are then selected and compressed into logic flows. Such logic flows consist of if and else structures and a reduced number of MAC operations. Experimental results demonstrate that the latency can be reduced by up to 14.9 % on a simulated RISC-V CPU without any accuracy degradation. The code is open source at https://github.com/TUDa-HWAI/NN2Logic
Authors: Jonathan Kahana, Eliahu Horwitz, Yedid Hoshen
Abstract: Public repositories host millions of fine-tuned models, yet community usage remains disproportionately concentrated on a small number of foundation checkpoints. We investigate whether this concentration reflects efficient market selection or if superior models are systematically overlooked. Through an extensive evaluation of over 2,000 models, we show the prevalence of "hidden gems", unpopular fine-tunes that significantly outperform their popular counterparts. Notably, within the Llama-3.1-8B family, we find rarely downloaded checkpoints that improve math performance from 83.2% to 96.0% without increasing inference costs. However, discovering these models through exhaustive evaluation of every uploaded model is computationally infeasible. We therefore formulate model discovery as a Multi-Armed Bandit problem and accelerate the Sequential Halving search algorithm by using shared query sets and aggressive elimination schedules. Our method retrieves top models with as few as 50 queries per candidate, accelerating discovery by over 50x.
Authors: Matteo Gianferrari, Omayma Moussadek, Riccardo Salami, Cosimo Fiorini, Lorenzo Tartarini, Daniela Gandolfi, Simone Calderara
Abstract: Spiking Neural Networks (SNNs) are inherently suited for continuous learning due to their event-driven temporal dynamics; however, their application to Class-Incremental Learning (CIL) has been hindered by catastrophic forgetting and the temporal misalignment of spike patterns. In this work, we introduce Spiking Temporal Alignment with Experience Replay (STAER), a novel framework that explicitly preserves temporal structure to bridge the performance gap between SNNs and ANNs. Our approach integrates a differentiable Soft-DTW alignment loss to maintain spike timing fidelity and employs a temporal expansion and contraction mechanism on output logits to enforce robust representation learning. Implemented on a deep ResNet19 spiking backbone, STAER achieves state-of-the-art performance on Sequential-MNIST and Sequential-CIFAR10. Empirical results demonstrate that our method matches or outperforms strong ANN baselines (ER, DER++) while preserving biologically plausible dynamics. Ablation studies further confirm that explicit temporal alignment is critical for representational stability, positioning STAER as a scalable solution for spike-native lifelong learning. Code is available at https://github.com/matteogianferrari/staer.
Authors: Mustafa Cavus, Przemys{\l}aw Biecek, Julian Tejada, Fernando Marmolejo-Ramos, Andre Faro
Abstract: This paper introduces a new modeling perspective in the public mental health domain to provide a robust interpretation of the relations between anxiety and depression, and the demographic and temporal factors. This perspective particularly leverages the Rashomon Effect, where multiple models exhibit similar predictive performance but rely on diverse internal structures. Instead of considering these multiple models, choosing a single best model risks masking alternative narratives embedded in the data. To address this, we employed this perspective in the interpretation of a large-scale psychological dataset, specifically focusing on the Patient Health Questionnaire-4. We use a random forest model combined with partial dependence profiles to rigorously assess the robustness and stability of predictive relationships across the resulting Rashomon set, which consists of multiple models that exhibit similar predictive performance. Our findings confirm that demographic variables \texttt{age}, \texttt{sex}, and \texttt{education} lead to consistent structural shifts in anxiety and depression risk. Crucially, we identify significant temporal effects: risk probability demonstrates clear diurnal and circaseptan fluctuations, peaking during early morning hours. This work demonstrates the necessity of moving beyond the best model to analyze the entire Rashomon set. Our results highlight that the observed variability, particularly due to circadian and circaseptan rhythms, must be meticulously considered for robust interpretation in psychological screening. We advocate for a multiplicity-aware approach to enhance the stability and generalizability of ML-based conclusions in mental health research.
Authors: Md Muhtasim Munif Fahim, Md Jahid Hasan Imran, Luknath Debnath, Tonmoy Shill, Md. Naim Molla, Ehsanul Bashar Pranto, Md Shafin Sanyan Saad, Md Rezaul Karim
Abstract: The achievement of the 2030 Sustainable Development Goals (SDGs) is dependent upon strategic resource distribution. We propose a causal discovery framework using Panel Vector Autoregression, along with both country-specific fixed effects and PCMCI+ conditional independence testing on 168 countries (2000-2025) to develop the first complete causal architecture of SDG dependencies. Utilizing 8 strategically chosen SDGs, we identify a distributed causal network (i.e., no single 'hub' SDG), with 10 statistically significant Granger-causal relationships identified as 11 unique direct effects. Education to Inequality is identified as the most statistically significant direct relationship (r = -0.599; p < 0.05), while effect magnitude significantly varies depending on income levels (e.g., high-income: r = -0.65; lower-middle-income: r = -0.06; non-significant). We also reject the idea that there exists a single 'keystone' SDG. Additionally, we offer a proposed tiered priority framework for the SDGs namely, identifying upstream drivers (Education, Growth), enabling goals (Institutions, Energy), and downstream outcomes (Poverty, Health). Therefore, we conclude that effective SDG acceleration can be accomplished through coordinated multi-dimensional intervention(s), and that single-goal sequential strategies are insufficient.
Authors: Stefano Maria Iacus, Haodong Qi, Devika Jain
Abstract: Recent advances in Generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), enable scalable extraction of spatial information from unstructured text and offer new methodological opportunities for studying climate geography. This study develops a spatial framework to examine how cumulative climate risk relates to multidimensional human flourishing across U.S. counties. High-resolution climate hazard indicators are integrated with a Human Flourishing Geographic Index (HFGI), an index derived from classification of 2.6 billion geotagged tweets using fine-tuned open-source Large Language Models (LLMs). These indicators are aggregated to the US county-level and mapped to a structural equation model to infer overall climate risk and human flourishing dimensions, including expressed well-being, meaning and purpose, social connectedness, psychological distress, physical condition, economic stability, religiosity, character and virtue, and institutional trust. The results reveal spatially heterogeneous associations between greater cumulative climate risk and lower levels of expressed human flourishing, with coherent spatial patterns corresponding to recurrent exposure to heat, flooding, wind, drought, and wildfire hazards. The study demonstrates how Generative AI can be combined with latent construct modeling for geographical analysis and for spatial knowledge extraction.
Authors: Bharath Krishnamurthy, Ajita Rattani
Abstract: Morphing techniques generate artificial biometric samples that combine features from multiple individuals, allowing each contributor to be verified against a single enrolled template. While extensively studied in face recognition, this vulnerability remains largely unexplored in voice biometrics. Prior work on voice morphing is computationally expensive, non-scalable, and limited to acoustically similar identity pairs, constraining practical deployment. Moreover, existing sound-morphing methods target audio textures, music, or environmental sounds and are not transferable to voice identity manipulation. We propose VoxMorph, a zero-shot framework that produces high-fidelity voice morphs from as little as five seconds of audio per subject without model retraining. Our method disentangles vocal traits into prosody and timbre embeddings, enabling fine-grained interpolation of speaking style and identity. These embeddings are fused via Spherical Linear Interpolation (Slerp) and synthesized using an autoregressive language model coupled with a Conditional Flow Matching network. VoxMorph achieves state-of-the-art performance, delivering a 2.6x gain in audio quality, a 73% reduction in intelligibility errors, and a 67.8% morphing attack success rate on automated speaker verification systems under strict security thresholds. This work establishes a practical and scalable paradigm for voice morphing with significant implications for biometric security. The code and dataset are available on our project page: https://vcbsl.github.io/VoxMorph/
Authors: Spencer Mateega, Jeff Yang, Tiana Costello, Shaurya Jadhav, Nicole Tian, Agustin Garcinu\~no
Abstract: IDE-Bench is a comprehensive framework for evaluating AI IDE agents on real-world software engineering tasks through an IDE-native tool interface. We present a Dockerized test harness that goes beyond raw terminal execution, granting models a structured tool ecosystem that represents AI-native IDEs like Cursor and Windsurf. By providing high-level abstractions for codebase search, structured file editing, and tools for testing full-stack applications, IDE-Bench evaluates an agent's ability to act as a true engineering collaborator. For evaluation and to prevent training data contamination, we created 80 tasks across eight never-published repositories spanning C/C++, Java, and MERN stacks, representing modern tech stack production scenarios, including feature implementation, bug fixing, refactoring, and performance optimization that mirror daily developer workflows in private codebases. Our benchmark is the first to systematically correlate agent-reported intent with successful project-level modifications in a multi-language, full-stack environment on completely uncontaminated code.
Authors: Youguang Chen, George Biros
Abstract: We study sampling from posterior distributions in Bayesian linear inverse problems where $A$, the parameters to observables operator, is computationally expensive. In many applications, $A$ can be factored in a manner that facilitates the construction of a cost-effective approximation $\tilde{A}$. In this framework, we introduce Latent-IMH, a sampling method based on the Metropolis-Hastings independence (IMH) sampler. Latent-IMH first generates intermediate latent variables using the approximate $\tilde{A}$, and then refines them using the exact $A$. Its primary benefit is that it shifts the computational cost to an offline phase. We theoretically analyze the performance of Latent-IMH using KL divergence and mixing time bounds. Using numerical experiments on several model problems, we show that, under reasonable assumptions, it outperforms state-of-the-art methods such as the No-U-Turn sampler (NUTS) in computational efficiency. In some cases, Latent-IMH can be orders of magnitude faster than existing schemes.
Authors: Hajung Kim, Eunha Lee, Sohyun Chung, Jueon Park, Seungheun Baek, Jaewoo Kang
Abstract: Identifying metabolic sites where cytochrome P450 enzymes metabolize small-molecule drugs is essential for drug discovery. Although existing computational approaches have been proposed for site-of-metabolism prediction, they typically ignore cytochrome P450 isoform identity or model isoforms independently, thereby failing to fully capture inherent cross-isoform metabolic patterns. In addition, prior evaluations often rely on top-k metrics, where false positive atoms may be included among the top predictions, underscoring the need for complementary metrics that more directly assess binary atom-level discrimination under severe class imbalance. We propose ATTNSOM, an atom-level site-of-metabolism prediction framework that integrates intrinsic molecular reactivity with cross-isoform relationships. The model combines a shared graph encoder, molecule-conditioned atom representations, and a cross-attention mechanism to capture correlated metabolic patterns across cytochrome P450 isoforms. The model is evaluated on two benchmark datasets annotated with site-of-metabolism labels at atom resolution. Across these benchmarks, the model achieves consistently strong top-k performance across multiple cytochrome P450 isoforms. Relative to ablated variants, the model yields higher Matthews correlation coefficient, indicating improved discrimination of true metabolic sites. These results support the importance of explicitly modeling cross-isoform relationships for site-of-metabolism prediction. The code and datasets are available at https://github.com/dmis-lab/ATTNSOM.
Authors: Sergio Burdisso, Esa\'u Villatoro-Tello, Shashi Kumar, Srikanth Madikeri, Andr\'es Carofilis, Pradeep Rangappa, Manjunath K E, Kadri Hacioglu, Petr Motlicek, Andreas Stolcke
Abstract: LLM-based automatic speech recognition (ASR), a well-established approach, connects speech foundation models to large language models (LLMs) through a speech-to-LLM projector, yielding promising results. A common design choice in these architectures is the use of a fixed, manually defined prompt during both training and inference. This setup not only enables applicability across a range of practical scenarios, but also helps maximize model performance. However, the impact of prompt design remains underexplored. This paper presents a comprehensive analysis of commonly used prompts across diverse datasets, showing that prompt choice significantly affects ASR performance and introduces instability, with no single prompt performing best across all cases. Inspired by the speech-to-LLM projector, we propose a prompt projector module, a simple, model-agnostic extension that learns to project prompt embeddings to more effective regions of the LLM input space, without modifying the underlying LLM-based ASR model. Experiments on four datasets show that the addition of a prompt projector consistently improves performance, reduces variability, and outperforms the best manually selected prompts.
Authors: Sergio Burdisso, Esa\'u Villatoro-Tello, Andr\'es Carofilis, Shashi Kumar, Kadri Hacioglu, Srikanth Madikeri, Pradeep Rangappa, Manjunath K E, Petr Motlicek, Shankar Venkatesan, Andreas Stolcke
Abstract: Adapting automatic speech recognition (ASR) systems based on large language models (LLMs) to new domains using text-only data is a significant yet underexplored challenge. Standard fine-tuning of the LLM on target-domain text often disrupts the critical alignment between speech and text modalities learned by the projector, degrading performance. We introduce a novel text-only adaptation method that emulates the audio projection task by treating it as a text denoising task. Our approach thus trains the LLM to recover clean transcripts from noisy inputs. This process effectively adapts the model to a target domain while preserving cross-modal alignment. Our solution is lightweight, requiring no architectural changes or additional parameters. Extensive evaluation on two datasets demonstrates up to 22.1% relative improvement, outperforming recent state-of-the-art text-only adaptation methods.
Authors: Xiaocheng Fang, Zhengyao Ding, Jieyi Cai, Yujie Xiao, Bo Liu, Jiarui Jin, Haoyu Wang, Guangkun Nie, Shun Huang, Ting Chen, Hongyan Li, Shenda Hong
Abstract: Cardiac Magnetic Resonance (CMR) imaging provides a comprehensive assessment of cardiac structure and function but remains constrained by high acquisition costs and reliance on expert annotations, limiting the availability of large-scale labeled datasets. In contrast, electrocardiograms (ECGs) are inexpensive, widely accessible, and offer a promising modality for conditioning the generative synthesis of cine CMR. To this end, we propose ECGFlowCMR, a novel ECG-to-CMR generative framework that integrates a Phase-Aware Masked Autoencoder (PA-MAE) and an Anatomy-Motion Disentangled Flow (AMDF) to address two fundamental challenges: (1) the cross-modal temporal mismatch between multi-beat ECG recordings and single-cycle CMR sequences, and (2) the anatomical observability gap due to the limited structural information inherent in ECGs. Extensive experiments on the UK Biobank and a proprietary clinical dataset demonstrate that ECGFlowCMR can generate realistic cine CMR sequences from ECG inputs, enabling scalable pretraining and improving performance on downstream cardiac disease classification and phenotype prediction tasks.
Authors: Azadeh Mokari, Shravan Raghunathan, Artem Shydliukh, Oleg Ryabchykov, Christoph Krafft, Thomas Bocklitz
Abstract: High-quality Fourier Transform Infrared (FTIR) imaging usually needs extensive signal averaging to reduce noise and drift which severely limits clinical speed. Deep learning can accelerate imaging by reconstructing spectra from rapid, single-scan inputs. However, separating noise and baseline drift simultaneously without ground truth is an ill-posed inverse problem. Standard black-box architectures often rely on statistical approximations that introduce spectral hallucinations or fail to generalize to unstable atmospheric conditions. To solve these issues we propose a physics-informed cascade Unet that separates denoising and baseline correction tasks using a new, deterministic Physics Bridge. This architecture forces the network to separate random noise from chemical signals using an embedded SNIP layer to enforce spectroscopic constraints instead of learning statistical approximations. We benchmarked this approach against a standard single Unet and a traditional Savitzky-Golay/SNIP workflow. We used a dataset of human hypopharyngeal carcinoma cells (FaDu). The cascade model outperformed all other methods, achieving a 51.3% reduction in RMSE compared to raw single-scan inputs, surpassing both the single Unet (40.2%) and the traditional workflow (33.7%). Peak-aware metrics show that the cascade architecture eliminates spectral hallucinations found in standard deep learning. It also preserves peak intensity with much higher fidelity than traditional smoothing. These results show that the cascade Unet is a robust solution for diagnostic-grade FTIR imaging. It enables imaging speeds 32 times faster than current methods.
Authors: G\"otz-Henrik Wiegand, Filip Stepniak, Patrick Baier
Abstract: Domain-Driven Design (DDD) is a key framework for developing customer-oriented software, focusing on the precise modeling of an application's domain. Traditionally, metamodels that describe these domains are created manually by system designers, forming the basis for iterative software development. This paper explores the partial automation of metamodel generation using generative AI, particularly for producing domain-specific JSON objects. By training a model on real-world DDD project data, we demonstrate that generative AI can produce syntactically correct JSON objects based on simple prompts, offering significant potential for streamlining the design process. To address resource constraints, the AI model was fine-tuned on a consumer-grade GPU using a 4-bit quantized version of Code Llama and Low-Rank Adaptation (LoRA). Despite limited hardware, the model achieved high performance, generating accurate JSON objects with minimal post-processing. This research illustrates the viability of incorporating generative AI into the DDD process, improving efficiency and reducing resource requirements, while also laying the groundwork for further advancements in AI-driven software development.
Authors: Simon Tonner, Viet T. Tran, Richard Kueng
Abstract: Quantum state tomography (QST) is essential for validating quantum devices but suffers from exponential scaling in system size. Neural-network quantum states, such as Restricted Boltzmann Machines (RBMs), can efficiently parameterize individual many-body quantum states and have been successfully used for QST. However, existing approaches are point-wise and require retraining at every parameter value in a phase diagram. We introduce a parametric QST framework based on a hypernetwork that conditions an RBM on Hamiltonian control parameters, enabling a single model to represent an entire family of quantum ground states. Applied to the transverse-field Ising model, our HyperRBM achieves high-fidelity reconstructions from local Pauli measurements on 1D and 2D lattices across both phases and through the critical region. Crucially, the model accurately reproduces the fidelity susceptibility and identifies the quantum phase transition without prior knowledge of the critical point. These results demonstrate that hypernetwork-modulated neural quantum states provide an efficient and scalable route to tomographic reconstruction across full phase diagrams.
Authors: Gabriel Ponte, Marcia Fampa, Jon Lee
Abstract: The maximum-entropy remote sampling problem (MERSP) is to select a subset of s random variables from a set of n random variables, so as to maximize the information concerning a set of target random variables that are not directly observable. We assume throughout that the set of all of these random variables follows a joint Gaussian distribution, and that we have the covariance matrix available. Finally, we measure information using Shannon's differential entropy. The main approach for exact solution of moderate-sized instances of MERSP has been branch-and-bound, and so previous work concentrated on upper bounds. Prior to our work, there were two upper-bounding methods for MERSP: the so-called NLP bound and the spectral bound, both introduced 25 years ago. We are able now to establish domination results between these two upper bounds. We propose an ``augmented NLP bound'' based on a subtle convex relaxation. We provide theoretical guarantees, giving sufficient conditions under which the augmented NLP bound strictly dominates the ordinary NLP bound. In addition, the augmented NLP formulation allows us to derive upper bounds for rank-deficient covariance matrices when they satisfy a technical condition. This is in contrast to the earlier work on the ordinary NLP bound that worked with only positive definite covariance matrices. Finally, we introduce a novel and very effective diagonal-scaling technique for MERSP, employing a positive vector of parameters. Numerical experiments on benchmark instances demonstrate the effectiveness of our approaches in advancing the state of the art for calculating upper bounds on MERSP.
Authors: Haixiang Sun, Pengchao Tian, Zihan Zhou, Jielei Zhang, Peiyi Li, Andrew L. Liu
Abstract: Learning causal structures from observational data remains a fundamental yet computationally intensive task, particularly in high-dimensional settings where existing methods face challenges such as the super-exponential growth of the search space and increasing computational demands. To address this, we introduce VISTA (Voting-based Integration of Subgraph Topologies for Acyclicity), a modular framework that decomposes the global causal structure learning problem into local subgraphs based on Markov Blankets. The global integration is achieved through a weighted voting mechanism that penalizes low-support edges via exponential decay, filters unreliable ones with an adaptive threshold, and ensures acyclicity using a Feedback Arc Set (FAS) algorithm. The framework is model-agnostic, imposing no assumptions on the inductive biases of base learners, is compatible with arbitrary data settings without requiring specific structural forms, and fully supports parallelization. We also theoretically establish finite-sample error bounds for VISTA, and prove its asymptotic consistency under mild conditions. Extensive experiments on both synthetic and real datasets consistently demonstrate the effectiveness of VISTA, yielding notable improvements in both accuracy and efficiency over a wide range of base learners.
Authors: Markus Holzleitner, Sergiy Pereverzyev Jr., Sergei V. Pereverzyev, Vaibhav Silmana, S. Sivananthan
Abstract: This paper investigates a general regularization framework for unsupervised domain adaptation in vector-valued regression under the covariate shift assumption, utilizing vector-valued reproducing kernel Hilbert spaces (vRKHS). Covariate shift occurs when the input distributions of the training and test data differ, introducing significant challenges for reliable learning. By restricting the hypothesis space, we develop a practical operator learning algorithm capable of handling functional outputs. We establish optimal convergence rates for the proposed framework under a general source condition, providing a theoretical foundation for regularized learning in this setting. We also propose an aggregation-based approach that forms a linear combination of estimators corresponding to different regularization parameters and different kernels. The proposed approach addresses the challenge of selecting appropriate tuning parameters, which is crucial for constructing a good estimator, and we provide a theoretical justification for its effectiveness. Furthermore, we illustrate the proposed method on a real-world face image dataset, demonstrating robustness and effectiveness in mitigating distributional discrepancies under covariate shift.
Authors: Louis Grenioux, RuiKang OuYang, Jos\'e Miguel Hern\'andez-Lobato
Abstract: Score-based generative models have recently achieved remarkable success. While they are usually parameterized by the score, an alternative way is to use a series of time-dependent energy-based models (EBMs), where the score is obtained from the negative input-gradient of the energy. Crucially, EBMs can be leveraged not only for generation, but also for tasks such as compositional sampling or building Boltzmann Generators via Monte Carlo methods. However, training EBMs remains challenging. Direct maximum likelihood is computationally prohibitive due to the need for nested sampling, while score matching, though efficient, suffers from mode blindness. To address these issues, we introduce the Diffusive Classification (DiffCLF) objective, a simple method that avoids blindness while remaining computationally efficient. DiffCLF reframes EBM learning as a supervised classification problem across noise levels, and can be seamlessly combined with standard score-based objectives. We validate the effectiveness of DiffCLF by comparing the estimated energies against ground truth in analytical Gaussian mixture cases, and by applying the trained models to tasks such as model composition and Boltzmann Generator sampling. Our results show that DiffCLF enables EBMs with higher fidelity and broader applicability than existing approaches.
Authors: Louis Grenioux, Maxence Noble
Abstract: Sampling configurations at thermodynamic equilibrium is a central challenge in statistical physics. Boltzmann Generators (BGs) tackle it by combining a generative model with a Monte Carlo (MC) correction step to obtain asymptotically unbiased samples from an unnormalized target. Most current BGs use classic MC mechanisms such as importance sampling, which both require tractable likelihoods from the backbone model and scale poorly in high-dimensional, multi-modal targets. We study BGs built on annealed Monte Carlo (aMC), which is designed to overcome these limitations by bridging a simple reference to the target through a sequence of intermediate densities. Diffusion models (DMs) are powerful generative models and have already been incorporated into aMC-based recalibration schemes via the diffusion-induced density path, making them appealing backbones for aMC-BGs. We provide an empirical meta-analysis of DM-based aMC-BGs on controlled multi-modal Gaussian mixtures (varying mode separation, number of modes, and dimension), explicitly disentangling inference effects from learning effects by comparing (i) a perfectly learned DM and (ii) a DM trained from data. Even with a perfect DM, standard integrations using only first-order stochastic denoising kernels fail systematically, whereas second-order denoising kernels can substantially improve performance when covariance information is available. We further propose a deterministic aMC integration based on first-order transport maps derived from DMs, which outperforms the stochastic first-order variant at higher computational cost. Finally, in the learned-DM setting, all DM-aMC variants struggle to produce accurate BGs; we trace the main bottleneck to inaccurate DM log-density estimation.
Authors: Zhuoran Yang, Ed Li, Jianliang He, Aman Priyanshu, Baturay Saglam, Paul Kassianik, Sajana Weerawardhena, Anu Vellore, Blaine Nelson, Neusha Javidnia, Arthur Goldblatt, Fraser Burch, Avi Zohary, Assaf Eisenman, Mahdi Sabbaghi, Supriti Vijay, Rahim Dharssi, Dhruv Kedia, Kojin Oshiba, Yaron Singer, Amin Karbasi
Abstract: We present Foundation-Sec-8B-Reasoning, the first open-source native reasoning model for cybersecurity. Built upon our previously released Foundation-Sec-8B base model (derived from Llama-3.1-8B-Base), the model is trained through a two-stage process combining supervised fine-tuning (SFT) and reinforcement learning from verifiable rewards (RLVR). Our training leverages proprietary reasoning data spanning cybersecurity analysis, instruction-following, and mathematical reasoning. Evaluation across 10 cybersecurity benchmarks and 10 general-purpose benchmarks demonstrates performance competitive with significantly larger models on cybersecurity tasks while maintaining strong general capabilities. The model shows effective generalization on multi-hop reasoning tasks and strong safety performance when deployed with appropriate system prompts and guardrails. This work demonstrates that domain-specialized reasoning models can achieve strong performance on specialized tasks while maintaining broad general capabilities. We release the model publicly at https://huggingface.co/fdtn-ai/Foundation-Sec-8B-Reasoning.
URLs: https://huggingface.co/fdtn-ai/Foundation-Sec-8B-Reasoning.
Authors: Thomas Warford, Fabian L. Thiemann, G\'abor Cs\'anyi
Abstract: The training of foundational machine learning interatomic potentials (fMLIPs) relies on diverse databases with energies and forces calculated using ab initio methods. We show that fMLIPs trained on large datasets such as MPtrj, Alexandria, and OMat24 encode inconsistencies from the Materials Project's selective use of the Hubbard U correction, which is applied to certain transition metals only if O or F atoms are present in the simulation cell. This inconsistent use of +U creates two incompatible potential-energy surfaces (PES): a lower-energy GGA surface and a higher-energy GGA+U one. When trained on both, MLIPs interpolate between them, leading to systematic underbinding, or even spurious repulsion, between U-corrected metals and oxygen- or fluorine-containing species. Models such as MACE-OMAT and -MPA exhibit repulsion between U-corrected metals and their oxides, limiting their value for studying catalysis and oxidation. We link the severity of this pathology to the oxygen number density in U-corrected training configurations. This explains why OMAT-trained models are most affected and suggests the issue might worsen as expanding future datasets increasingly include configurations with low oxygen content, such as those generated through combinatorial exploration of multi-element or defect-containing systems. Our simple per-U-corrected-atom shift aligns PBE+U and PBE energies for identical structures, yielding a smoother PES compared to existing correction schemes, which target phase diagram accuracy. As a result, models trained on datasets with our shift applied exhibit smaller mean absolute errors for the adsorption energies of oxygen on U-corrected elemental slabs. Since datasets omitting +U entirely (e.g. MatPES, MP-ALOE) avoid these pathologies, we recommend excluding +U in future fMLIP datasets. For existing datasets, our post-hoc correction provides a low-cost improvement.
Authors: Lizuo Liu, Lu Zhang, Anne Gelb
Abstract: We propose a parametric hyperbolic conservation law (SymCLaw) for learning hyperbolic systems directly from data while ensuring conservation, entropy stability, and hyperbolicity by design. Unlike existing approaches that typically enforce only conservation or rely on prior knowledge of the governing equations, our method parameterizes the flux functions in a form that guarantees real eigenvalues and complete eigenvectors of the flux Jacobian, thereby preserving hyperbolicity. At the same time, we embed entropy-stable design principles by jointly learning a convex entropy function and its associated flux potential, ensuring entropy dissipation and the selection of physically admissible weak solutions. A corresponding entropy-stable numerical flux scheme provides compatibility with standard discretizations, allowing seamless integration into classical solvers. Numerical experiments on benchmark problems, including Burgers, shallow water, Euler, and KPP equations, demonstrate that SymCLaw generalizes to unseen initial conditions, maintains stability under noisy training data, and achieves accurate long-time predictions, highlighting its potential as a principled foundation for data-driven modeling of hyperbolic conservation laws.
Authors: Jacob Zhu, Donald Estep
Abstract: We introduce a novel machine learning method called the Penalized Profile Support Vector Machine based on the Gabriel edited set for the computation of the probability of failure for a complex system as determined by a threshold condition on a computer model of system behavior. The method is designed to minimize the number of evaluations of the computer model while preserving the geometry of the decision boundary that determines the probability. It employs an adaptive sampling strategy designed to strategically allocate points near the boundary determining failure and builds a locally linear surrogate boundary that remains consistent with its geometry by strategic clustering of training points. We prove two convergence results and we compare the performance of the method against a number of state of the art classification methods on four test problems. We also apply the method to determine the probability of survival using the Lotka--Volterra model for competing species.
Authors: Zhou Fan, Leda Wang
Abstract: We study the learning dynamics of a multi-pass, mini-batch Stochastic Gradient Descent (SGD) procedure for empirical risk minimization in high-dimensional multi-index models with isotropic random data. In an asymptotic regime where the sample size $n$ and data dimension $d$ increase proportionally, for any sub-linear batch size $\kappa \asymp n^\alpha$ where $\alpha \in [0,1)$, and for a commensurate ``critical'' scaling of the learning rate, we provide an asymptotically exact characterization of the coordinate-wise dynamics of SGD. This characterization takes the form of a system of dynamical mean-field equations, driven by a scalar Poisson jump process that represents the asymptotic limit of SGD sampling noise. We develop an analogous characterization of the Stochastic Modified Equation (SME) which provides a Gaussian diffusion approximation to SGD. Our analyses imply that the limiting dynamics for SGD are the same for any batch size scaling $\alpha \in [0,1)$, and that under a commensurate scaling of the learning rate, dynamics of SGD, SME, and gradient flow are mutually distinct, with those of SGD and SME coinciding in the special case of a linear model. We recover a known dynamical mean-field characterization of gradient flow in a limit of small learning rate, and of one-pass/online SGD in a limit of increasing sample size $n/d \to \infty$.
Authors: Hongzheng Chen, Alexander Novikov, Ng\^an V\~u, Hanna Alam, Zhiru Zhang, Aiden Grossman, Mircea Trofin, Amir Yazdanbakhsh
Abstract: Modern compilers rely on hand-crafted heuristics to guide optimization passes. These human-designed rules often struggle to adapt to the complexity of modern software and hardware and lead to high maintenance burden. To address this challenge, we present Magellan, an agentic framework that evolves the compiler pass itself by synthesizing executable C++ decision logic. Magellan couples an LLM coding agent with evolutionary search and autotuning in a closed loop of generation, evaluation on user-provided macro-benchmarks, and refinement, producing compact heuristics that integrate directly into existing compilers. Across several production optimization tasks, Magellan discovers policies that match or surpass expert baselines. In LLVM function inlining, Magellan synthesizes new heuristics that outperform decades of manual engineering for both binary-size reduction and end-to-end performance. In register allocation, it learns a concise priority rule for live-range processing that matches intricate human-designed policies on a large-scale workload. We also report preliminary results on XLA problems, demonstrating portability beyond LLVM with reduced engineering effort.
Authors: Aidan Gleich, Scott C. Schmidler
Abstract: We address the problem of accurate, training-free guidance for conditional generation in trained diffusion models. Existing methods typically rely on point-estimates to approximate the posterior score, often resulting in biased approximations that fail to capture multimodality inherent to the reverse process of diffusion models. We propose a sequential Monte Carlo (SMC) framework that constructs an unbiased estimator of $p_\theta(y|x_t)$ by integrating over the full denoising distribution via Monte Carlo approximation. To ensure computational tractability, we incorporate variance-reduction schemes based on Multi-Level Monte Carlo (MLMC). Our approach achieves new state-of-the-art results for training-free guidance on CIFAR-10 class-conditional generation, achieving $95.6\%$ accuracy with $3\times$ lower cost-per-success than baselines. On ImageNet, our algorithm achieves $1.5\times$ cost-per-success advantage over existing methods.
Authors: Qiyang Han
Abstract: Adaptive sampling schemes are well known to create complex dependence that may invalidate conventional inference methods. A recent line of work shows that this need not be the case for UCB-type algorithms in multi-armed bandits. A central emerging theme is a `stability' property with asymptotically deterministic arm-pull counts in these algorithms, making inference as easy as in the i.i.d. setting. In this paper, we study the precise arm-pull dynamics in another canonical class of Thompson-sampling type algorithms. We show that the phenomenology is qualitatively different: the arm-pull count is asymptotically deterministic if and only if the arm is suboptimal or is the unique optimal arm; otherwise it converges in distribution to the unique invariant law of an SDE. This dichotomy uncovers a unifying principle behind many existing (in)stability results: an arm is stable if and only if its interaction with statistical noise is asymptotically negligible. As an application, we show that normalized arm means obey the same dichotomy, with Gaussian limits for stable arms and a semi-universal, non-Gaussian limit for unstable arms. This not only enables the construction of confidence intervals for the unknown mean rewards despite non-normality, but also reveals the potential of developing tractable inference procedures beyond the stable regime. The proofs rely on two new approaches. For suboptimal arms, we develop an `inverse process' approach that characterizes the inverse of the arm-pull count process via a Stieltjes integral. For optimal arms, we adopt a reparametrization of the arm-pull and noise processes that reduces the singularity in the natural SDE to proving the uniqueness of the invariant law of another SDE. We prove the latter by a set of analytic tools, including the parabolic H\"ormander condition and the Stroock-Varadhan support theorem.
Authors: Liang Hong
Abstract: The extant insurance literature demonstrates a paucity of finite-sample valid prediction intervals of future insurance claims in the regression setting. To address this challenge, this article proposes a new strategy that converts a predictive method in the unsupervised iid (independent identically distributed) setting to a predictive method in the regression setting. In particular, it enables an actuary to obtain infinitely many finite-sample valid prediction intervals in the regression setting.
Authors: Miles Wang, Robi Lin, Kat Hu, Joy Jiao, Neil Chowdhury, Ethan Chang, Tejal Patwardhan
Abstract: We introduce FrontierScience, a benchmark evaluating expert-level scientific reasoning in frontier language models. Recent model progress has nearly saturated existing science benchmarks, which often rely on multiple-choice knowledge questions or already published information. FrontierScience addresses this gap through two complementary tracks: (1) Olympiad, consisting of international olympiad problems at the level of IPhO, IChO, and IBO, and (2) Research, consisting of PhD-level, open-ended problems representative of sub-tasks in scientific research. FrontierScience contains several hundred questions (including 160 in the open-sourced gold set) covering subfields across physics, chemistry, and biology, from quantum electrodynamics to synthetic organic chemistry. All Olympiad problems are originally produced by international Olympiad medalists and national team coaches to ensure standards of difficulty, originality, and factuality. All Research problems are research sub-tasks written and verified by PhD scientists (doctoral candidates, postdoctoral researchers, or professors). For Research, we introduce a granular rubric-based evaluation framework to assess model capabilities throughout the process of solving a research task, rather than judging only a standalone final answer.
Authors: Jacek Duszenko
Abstract: Reasoning models frequently agree with incorrect user suggestions -- a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. To localize and quantify this behavior, we introduce \emph{sycophantic anchors} -- sentences that causally lock models into user agreement. Analyzing over 10,000 counterfactual rollouts on a distilled reasoning model, we show that anchors can be reliably detected and quantified mid-inference. Linear probes distinguish sycophantic anchors with 84.6\% balanced accuracy, while activation-based regressors predict the magnitude of the commitment ($R^2 = 0.74$). We further observe asymmetry where sycophantic anchors are significantly more distinguishable than correct reasoning anchors, and find that sycophancy builds gradually during reasoning, revealing a potential window for intervention. These results offer sentence-level mechanisms for localizing model misalignment mid-inference.
Authors: Chenyu Huang, Peng Ye, Xudong Tan, Jinhan Mu, Shenghe Zheng, Li Shen, Tao Chen
Abstract: Efficiently enhancing the reasoning capabilities of Vision-Language Models (VLMs) by merging them with Large Reasoning Models (LRMs) has emerged as a promising direction. However, existing methods typically operate at a coarse-grained layer level, which often leads to a trade-off between injecting reasoning capabilities and preserving visual capabilities. To address this limitation, we propose {FRISM} (Fine-grained Reasoning Injection via Subspace-level model Merging), a fine-grained reasoning injection framework based on subspace-level model merging. Observing that reasoning capabilities are encoded in distinct subspaces, FRISM decomposes LRM task vectors via Singular Value Decomposition (SVD) and adaptively tunes the scaling coefficients of each subspace through learning to realize fine-grained reasoning injection. Furthermore, we introduce a label-free self-distillation learning strategy with a dual-objective optimization using common vision-language perception datasets. Extensive experiments demonstrate that FRISM effectively improves reasoning capabilities without compromising the model's original visual capabilities by consistently achieving state-of-the-art performance across diverse visual reasoning benchmarks.
Authors: Yuchen Yang, Yaru Zhao, Pu Yang, Shaowei Wang, Zhi-Hua Zhou
Abstract: While Mixture-of-Experts (MoE) architectures substantially bolster the expressive power of large-language models, their prohibitive memory footprint severely impedes the practical deployment on resource-constrained edge devices, especially when model behavior must be preserved without relying on lossy quantization. In this paper, we present ZipMoE, an efficient and semantically lossless on-device MoE serving system. ZipMoE exploits the synergy between the hardware properties of edge devices and the statistical redundancy inherent to MoE parameters via a caching-scheduling co-design with provable performance guarantee. Fundamentally, our design shifts the paradigm of on-device MoE inference from an I/O-bound bottleneck to a compute-centric workflow that enables efficient parallelization. We implement a prototype of ZipMoE and conduct extensive experiments on representative edge computing platforms using popular open-source MoE models and real-world workloads. Our evaluation reveals that ZipMoE achieves up to $72.77\%$ inference latency reduction and up to $6.76\times$ higher throughput than the state-of-the-art systems.
Authors: Sharan Sahu, Arisina Banerjee, Yuchen Wu
Abstract: Classifier-guided diffusion models generate conditional samples by augmenting the reverse-time score with the gradient of a learned classifier, yet it remains unclear whether standard classifier training procedures yield effective diffusion guidance. We address this gap by showing that, under mild smoothness assumptions on the classifiers, controlling the cross-entropy error at each diffusion step also controls the error of the resulting guidance vectors: classifiers achieving conditional KL divergence $\varepsilon^2$ from the ground-truth conditional label probabilities induce guidance vectors with mean squared error $\widetilde{O}(d \varepsilon )$. Our result yields an upper bound on the sampling error under classifier guidance and bears resemblance to a reverse log-Sobolev-type inequality. Moreover, we show that the classifier smoothness assumption is essential, by constructing simple counterexamples demonstrating that, without it, control of the guidance vector can fail for almost all distributions. To our knowledge, our work establishes the first quantitative link between classifier training and guidance alignment, yielding both a theoretical foundation for classifier guidance and principled guidelines for classifier selection.
Authors: Hong Liu, Jiaqi Zhang, Chao Wang, Xing Hu, Linkun Lyu, Jiaqi Sun, Xurui Yang, Bo Wang, Fengcun Li, Yulei Qian, Lingtong Si, Yerui Sun, Rumei Li, Peng Pei, Yuchen Xie, Xunliang Cai
Abstract: While Mixture-of-Experts (MoE) architectures have become the standard for sparsity scaling in large language models, they increasingly face diminishing returns and system-level bottlenecks. In this work, we explore embedding scaling as a potent, orthogonal dimension for scaling sparsity. Through a comprehensive analysis and experiments, we identify specific regimes where embedding scaling achieves a superior Pareto frontier compared to expert scaling. We systematically characterize the critical architectural factors governing this efficacy -- ranging from parameter budgeting to the interplay with model width and depth. Moreover, by integrating tailored system optimizations and speculative decoding, we effectively convert this sparsity into tangible inference speedups. Guided by these insights, we introduce LongCat-Flash-Lite, a 68.5B parameter model with ~3B activated trained from scratch. Despite allocating over 30B parameters to embeddings, LongCat-Flash-Lite not only surpasses parameter-equivalent MoE baselines but also exhibits exceptional competitiveness against existing models of comparable scale, particularly in agentic and coding domains.
Authors: Zhaoyi Li, Jiatong Li, Gangwei Jiang, Linqi Song, Defu Lian, Ying Wei
Abstract: Chain-of-thought (CoT) reasoning has become the standard paradigm for enabling Large Language Models (LLMs) to solve complex problems. However, recent studies reveal a sharp performance drop in reasoning hop generalization scenarios, where the required number of reasoning steps exceeds training distributions while the underlying algorithm remains unchanged. The internal mechanisms driving this failure remain poorly understood. In this work, we conduct a systematic study on tasks from multiple domains, and find that errors concentrate at token positions of a few critical error types, rather than being uniformly distributed. Closer inspection reveals that these token-level erroneous predictions stem from internal competition mechanisms: certain attention heads, termed erroneous processing heads (ep heads), tip the balance by amplifying incorrect reasoning trajectories while suppressing correct ones. Notably, removing individual ep heads during inference can often restore the correct predictions. Motivated by these insights, we propose test-time correction of reasoning, a lightweight intervention method that dynamically identifies and deactivates ep heads in the reasoning process. Extensive experiments across different tasks and LLMs show that it consistently improves reasoning hop generalization, highlighting both its effectiveness and potential.
Authors: Dongyue Xie, Wanrong Zhu, Matthew Stephens
Abstract: We introduce a flexible empirical Bayes approach for fitting Bayesian generalized linear models. Specifically, we adopt a novel mean-field variational inference (VI) method and the prior is estimated within the VI algorithm, making the method tuning-free. Unlike traditional VI methods that optimize the posterior density function, our approach directly optimizes the posterior mean and prior parameters. This formulation reduces the number of parameters to optimize and enables the use of scalable algorithms such as L-BFGS and stochastic gradient descent. Furthermore, our method automatically determines the optimal posterior based on the prior and likelihood, distinguishing it from existing VI methods that often assume a Gaussian variational. Our approach represents a unified framework applicable to a wide range of exponential family distributions, removing the need to develop unique VI methods for each combination of likelihood and prior distributions. We apply the framework to solve sparse logistic regression and demonstrate the superior predictive performance of our method in extensive numerical studies, by comparing it to prevalent sparse logistic regression approaches.
Authors: Alex Ross, Tianqi Zhao, Sanjay Reddy
Abstract: We develop a machine learning model based on a structured variational autoencoder (VAE) framework to reconstruct and generate neutron star (NS) equations of state (EOS). The VAE consists of an encoder network that maps high-dimensional EOS data into a lower-dimensional latent space and a decoder network that reconstructs the full EOS from the latent representation. The latent space includes supervised NS observables derived from the training EOS data, as well as latent random variables corresponding to additional unspecified EOS features learned automatically. Sampling the latent space enables the generation of new, causal, and stable EOS models that satisfy astronomical constraints on the supervised NS observables, while allowing Bayesian inference of the EOS incorporating additional multimessenger data, including gravitational waves from LIGO/Virgo and mass and radius measurements of pulsars. Based on a VAE trained on a Skyrme EOS dataset, we find that a latent space with two supervised NS observables, the maximum mass $(M_{\max})$ and the canonical radius $(R_{1.4})$, together with one latent random variable controlling the EOS near the crust--core transition, can already reconstruct Skyrme EOSs with high fidelity, achieving mean absolute percentage errors of approximately $(0.15\%)$ for $(M_{\max})$ and $(R_{1.4})$ derived from the decoder-reconstructed EOS.
Authors: Aaron Li, Ian Zhang
Abstract: Kleinberg and Mullainathan recently proposed a formal framework for studying the phenomenon of language generation, called language generation in the limit. In this model, an adversary gives an enumeration of example strings from an unknown target language, and the algorithm is tasked with correctly generating unseen strings from the target language within finite time. Refined notions of non-uniform and uniform generation were later introduced by Li, Raman, and Tewari (2025), and a noisy model was introduced by Raman and Raman (2025), which allows the adversary to insert extraneous strings. A natural question in the noisy model is to quantify the effect of noise, by studying the impact of each additional extraneous string. We show two complementary results in this setting. We first show that for both uniform and non-uniform generation, a single noisy string strictly reduces the set of collections that can be generated, thus answering an open question in Raman and Raman (2025). Then, we show for both uniform and non-uniform generation that generation with a single noisy string is equivalent to generation with any finite amount of noise, sharply contrasting with the strict hierarchy for noisy generation in the limit shown by Bai, Panigrahi, and Zhang (2026). Finally, we leverage our previous results to provide the first known characterization for non-uniform noise-dependent generatability.
Authors: Amir Ali Farzin, Yuen-Man Pun, Philipp Braun, Tyler Summers, Iman Shames
Abstract: We consider max-min and min-max problems with objective functions that are possibly non-smooth, submodular with respect to the minimiser and concave with respect to the maximiser. We investigate the performance of a zeroth-order method applied to this problem. The method is based on the subgradient of the Lov\'asz extension of the objective function with respect to the minimiser and based on Gaussian smoothing to estimate the smoothed function gradient with respect to the maximiser. In expectation sense, we prove the convergence of the algorithm to an $\epsilon$-saddle point in the offline case. Moreover, we show that, in the expectation sense, in the online setting, the algorithm achieves $O(\sqrt{N\bar{P}_N})$ online duality gap, where $N$ is the number of iterations and $\bar{P}_N$ is the path length of the sequence of optimal decisions. The complexity analysis and hyperparameter selection are presented for all the cases. The theoretical results are illustrated via numerical examples.
Authors: Enzo Nicol\'as Spotorno, Ant\^onio Augusto Medeiros Fr\"ohlich
Abstract: The machine learning community has achieved remarkable success with universal foundation models for time-series and physical dynamics, largely overcoming earlier approximation barriers in smooth or slowly varying regimes through scale and specialized architectures. However, deploying these monolithic models in safety-critical Cyber-Physical Systems (CPS), governed by non-stationary lifecycle dynamics and strict reliability requirements, reveals persistent challenges. Recent evidence shows that fine-tuning time-series foundation models induces catastrophic forgetting, degrading performance on prior regimes. Standard models continue to exhibit residual spectral bias, smoothing high-frequency discontinuities characteristic of incipient faults, while their opacity hinders formal verification and traceability demanded by safety standards (e.g., ISO 26262, IEC 61508). This position paper argues that the plasticity-stability paradox cannot be fully resolved by global parameter updates (whether via offline fine-tuning or online adaptation). Instead, we advocate a Modular Sovereignty paradigm: a library of compact, frozen regime-specific specialists combined via uncertainty-aware blending, which we term "HYDRA" (Hierarchical uncertaintY-aware Dynamics for Rapidly-Adapting systems). This paradigm ensures regime-conditional validity, rigorous disentanglement of aleatoric and epistemic uncertainties, and modular auditability, offering a certifiable path for robust state integrity across the CPS lifecycle.
Authors: Esteban Rodr\'iguez-Betancourt, Edgar Casasola-Murillo
Abstract: A recurring challenge in self-supervised learning is preventing representation collapse. Existing solutions typically rely on global regularization, such as maximizing distances, decorrelating dimensions or enforcing certain distributions. We instead reinterpret representation learning as a discrete packing problem, where preserving information simplifies to maintaining injectivity. We operationalize this in Hypersolid, a method using short-range hard-ball repulsion to prevent local collisions. This constraint results in a high-separation geometric regime that preserves augmentation diversity, excelling on fine-grained and low-resolution classification tasks.
Authors: Seonghyeon Go, Yumin Kim
Abstract: Recently, the problem of music plagiarism has emerged as an even more pressing social issue. As music information retrieval research advances, there is a growing effort to address issues related to music plagiarism. However, many studies, including our previous work, have conducted research without clearly defining what the music plagiarism detection task actually involves. This lack of a clear definition has slowed research progress and made it hard to apply results to real-world scenarios. To fix this situation, we defined how Music Plagiarism Detection is different from other MIR tasks and explained what problems need to be solved. We introduce the Similar Music Pair dataset to support this newly defined task. In addition, we propose a method based on segment transcription as one way to solve the task. Our demo and dataset are available at https://github.com/Mippia/ICASSP2026-MPD.
Authors: Mohit Talreja, Joshua Diao, Jim Thannikary James, Radu Casapu, Tejas Santanam, Ethan Mendes, Alan Ritter, Wei Xu, James Hays
Abstract: Vision Language Models (VLMs) are good at recognizing the global location of a photograph -- their geolocation prediction accuracy rivals the best human experts. But many VLMs are startlingly bad at explaining which image evidence led to their prediction, even when their location prediction is correct. The reasoning chains produced by VLMs frequently hallucinate scene attributes to support their location prediction (e.g. phantom writing, imagined infrastructure, misidentified flora). In this paper, we introduce the first benchmark for geolocation reasoning chains. We focus on the global location prediction task in the popular GeoGuessr game which draws from Google Street View spanning more than 100 countries. We collaborate with expert GeoGuessr players, including the reigning world champion, to produce 800 ground truth reasoning chains for 500 query scenes. These expert reasoning chains address hundreds of different discriminative visual attributes such as license plate shape, architecture, and soil properties to name just a few. We evaluate LLM-as-a-judge and VLM-as-a-judge strategies for scoring VLM-generated reasoning chains against our expert reasoning chains and find that Qwen 3 LLM-as-a-judge correlates best with human scoring. Our benchmark reveals that while large, closed-source VLMs such as Gemini and GPT 5 rival human experts at prediction locations, they still lag behind human experts when it comes to producing auditable reasoning chains. Open weights VLMs such as Llama and Qwen catastrophically fail on our benchmark -- they perform only slightly better than a baseline in which an LLM hallucinates a reasoning chain with oracle knowledge of the photo location but no visual information at all. We believe the gap between human experts and VLMs on this task points to VLM limitations at extracting fine-grained visual attributes from high resolution images.
Authors: Keke Tang, Ziyong Du, Xiaofei Wang, Weilong Peng, Peican Zhu, Zhihong Tian
Abstract: Deep neural networks (DNNs) often produce overconfident predictions on out-of-distribution (OOD) inputs, undermining their reliability in open-world environments. Singularities in semi-discrete optimal transport (OT) mark regions of semantic ambiguity, where classifiers are particularly prone to unwarranted high-confidence predictions. Motivated by this observation, we propose a principled framework to mitigate OOD overconfidence by leveraging the geometry of OT-induced singular boundaries. Specifically, we formulate an OT problem between a continuous base distribution and the latent embeddings of training data, and identify the resulting singular boundaries. By sampling near these boundaries, we construct a class of OOD inputs, termed optimal transport-induced OOD samples (OTIS), which are geometrically grounded and inherently semantically ambiguous. During training, a confidence suppression loss is applied to OTIS to guide the model toward more calibrated predictions in structurally uncertain regions. Extensive experiments show that our method significantly alleviates OOD overconfidence and outperforms state-of-the-art methods.
Authors: Mengqi Chen, Thomas B. Berrett, Theodoros Damoulas, Michele Caprio
Abstract: Distributionally robust optimisation (DRO) minimises the worst-case expected loss over an ambiguity set that can capture distributional shifts in out-of-sample environments. While Huber (linear-vacuous) contamination is a classical minimal-assumption model for an $\varepsilon$-fraction of arbitrary perturbations, including it in an ambiguity set can make the worst-case risk infinite and the DRO objective vacuous unless one imposes strong boundedness or support assumptions. We address these challenges by introducing bulk-calibrated credal ambiguity sets: we learn a high-mass bulk set from data while considering contamination inside the bulk and bounding the remaining tail contribution separately. This leads to a closed-form, finite $\mathrm{mean}+\sup$ robust objective and tractable linear or second-order cone programs for common losses and bulk geometries. Through this framework, we highlight and exploit the equivalence between the imprecise probability (IP) notion of upper expectation and the worst-case risk, demonstrating how IP credal sets translate into DRO objectives with interpretable tolerance levels. Experiments on heavy-tailed inventory control, geographically shifted house-price regression, and demographically shifted text classification show competitive robustness-accuracy trade-offs and efficient optimisation times, using Bayesian, frequentist, or empirical reference distributions.
Authors: Ellen Xiaoqing Tan, Shehzaad Dhuliawala, Jing Xu, Ping Yu, Sainbayar Sukhbaatar, Jason Weston, Olga Golovneva
Abstract: Ensuring safety, factuality and overall quality in the generations of large language models is a critical challenge, especially as these models are increasingly deployed in real-world applications. The prevailing approach to addressing these issues involves collecting expensive, carefully curated datasets and applying multiple stages of fine-tuning and alignment. However, even this complex pipeline cannot guarantee the correction of patterns learned during pretraining. Therefore, addressing these issues during pretraining is crucial, as it shapes a model's core behaviors and prevents unsafe or hallucinated outputs from becoming deeply embedded. To tackle this issue, we introduce a new pretraining method that streams documents and uses reinforcement learning (RL) to improve the next K generated tokens at each step. A strong, post-trained model judges candidate generations -- including model rollouts, the original suffix, and a rewritten suffix -- for quality, safety, and factuality. Early in training, the process relies on the original and rewritten suffixes; as the model improves, RL rewards high-quality rollouts. This approach builds higher quality, safer, and more factual models from the ground up. In experiments, our method gives 36.2% and 18.5% relative improvements over standard pretraining in terms of factuality and safety, and up to 86.3% win rate improvements in overall generation quality.
Authors: Devanshu Sahoo, Manish Prasad, Vasudev Majhi, Arjun Neekhra, Yash Sinha, Murari Mandal, Vinay Chamola, Dhruv Kumar
Abstract: The rapid integration of Large Language Models (LLMs) into educational assessment rests on the unverified assumption that instruction following capability translates directly to objective adjudication. We demonstrate that this assumption is fundamentally flawed. Instead of evaluating code quality, models frequently decouple from the submission's logic to satisfy hidden directives, a systemic vulnerability we term the Compliance Paradox, where models fine-tuned for extreme helpfulness are vulnerable to adversarial manipulation. To expose this, we introduce the Semantic-Preserving Adversarial Code Injection (SPACI) Framework and the Abstract Syntax Tree-Aware Semantic Injection Protocol (AST-ASIP). These methods exploit the Syntax-Semantics Gap by embedding adversarial directives into syntactically inert regions (trivia nodes) of the Abstract Syntax Tree. Through a large-scale evaluation of 9 SOTA models across 25,000 submissions in Python, C, C++, and Java, we reveal catastrophic failure rates (>95%) in high-capacity open-weights models like DeepSeek-V3, which systematically prioritize hidden formatting constraints over code correctness. We quantify this failure using our novel tripartite framework measuring Decoupling Probability, Score Divergence, and Pedagogical Severity to demonstrate the widespread "False Certification" of functionally broken code. Our findings suggest that current alignment paradigms create a "Trojan" vulnerability in automated grading, necessitating a shift from standard RLHF toward domain-specific Adjudicative Robustness, where models are conditioned to prioritize evidence over instruction compliance. We release our complete dataset and injection framework to facilitate further research on the topic.
Authors: Wenjia Hua, Kejie Zhao, Luziwei Leng, Ran Cheng, Yuxin Ma, Qinghai Guo
Abstract: Backpropagation algorithm has driven the remarkable success of deep neural networks, but its lack of biological plausibility and high computational costs have motivated the ongoing search for alternative training methods. Hebbian learning has attracted considerable interest as a biologically plausible alternative to backpropagation. Nevertheless, its exclusive reliance on local information, without consideration of global task objectives, fundamentally limits its scalability. Inspired by the biological synergy between neuromodulators and local plasticity, we introduce a novel model-agnostic Global-guided Hebbian Learning (GHL) framework, which seamlessly integrates local and global information to scale up across diverse networks and tasks. In specific, the local component employs Oja's rule with competitive learning to ensure stable and effective local updates. Meanwhile, the global component introduces a sign-based signal that guides the direction of local Hebbian plasticity updates. Extensive experiments demonstrate that our method consistently outperforms existing Hebbian approaches. Notably, on large-scale network and complex datasets like ImageNet, our framework achieves the competitive results and significantly narrows the gap with standard backpropagation.
Authors: Zihan Su, Hongyang Wei, Kangrui Cen, Yong Wang, Guanhua Chen, Chun Yuan, Xiangxiang Chu
Abstract: Unified Multimodal Models (UMMs) integrate both visual understanding and generation within a single framework. Their ultimate aspiration is to create a cycle where understanding and generation mutually reinforce each other. While recent post-training methods have successfully leveraged understanding to enhance generation, the reverse direction of utilizing generation to improve understanding remains largely unexplored. In this work, we propose UniMRG (Unified Multi-Representation Generation), a simple yet effective architecture-agnostic post-training method. UniMRG enhances the understanding capabilities of UMMs by incorporating auxiliary generation tasks. Specifically, we train UMMs to generate multiple intrinsic representations of input images, namely pixel (reconstruction), depth (geometry), and segmentation (structure), alongside standard visual understanding objectives. By synthesizing these diverse representations, UMMs capture complementary information regarding appearance, spatial relations, and structural layout. Consequently, UMMs develop a deeper and more comprehensive understanding of visual inputs. Extensive experiments across diverse UMM architectures demonstrate that our method notably enhances fine-grained perception, reduces hallucinations, and improves spatial understanding, while simultaneously boosting generation capabilities.
Authors: Erica Zhang, Naomi Sagan, Danny Tse, Fangzhao Zhang, Mert Pilanci, Jose Blanchet
Abstract: We introduce Statsformer, a principled framework for integrating large language model (LLM)-derived knowledge into supervised statistical learning. Existing approaches are limited in adaptability and scope: they either inject LLM guidance as an unvalidated heuristic, which is sensitive to LLM hallucination, or embed semantic information within a single fixed learner. Statsformer overcomes both limitations through a guardrailed ensemble architecture. We embed LLM-derived feature priors within an ensemble of linear and nonlinear learners, adaptively calibrating their influence via cross-validation. This design yields a flexible system with an oracle-style guarantee that it performs no worse than any convex combination of its in-library base learners, up to statistical error. Empirically, informative priors yield consistent performance improvements, while uninformative or misspecified LLM guidance is automatically downweighted, mitigating the impact of hallucinations across a diverse range of prediction tasks.
Authors: Yizhou Min, Yizhou Lu, Lanqi Li, Zhen Zhang, Jiaye Teng
Abstract: Conformal prediction (CP) has become a cornerstone of distribution-free uncertainty quantification, conventionally evaluated by its coverage and interval length. This work critically examines the sufficiency of these standard metrics. We demonstrate that the interval length might be deceptively improved through a counter-intuitive approach termed Prejudicial Trick (PT), while the coverage remains valid. Specifically, for any given test sample, PT probabilistically returns an interval, which is either null or constructed using an adjusted confidence level, thereby preserving marginal coverage. While PT potentially yields a deceptively lower interval length, it introduces practical vulnerabilities: the same input can yield completely different prediction intervals across repeated runs of the algorithm. We formally derive the conditions under which PT achieves these misleading improvements and provides extensive empirical evidence across various regression and classification tasks. Furthermore, we introduce a new metric interval stability which helps detect whether a new CP method implicitly improves the length based on such PT-like techniques.
Authors: Chao Wang, Licheng Jiao, Lingling Li, Jiaxuan Zhao, Guanchun Wang, Fang Liu, Shuyuan Yang
Abstract: Handcrafted optimizers become prohibitively inefficient for complex black-box optimization (BBO) tasks. MetaBBO addresses this challenge by meta-learning to automatically configure optimizers for low-level BBO tasks, thereby eliminating heuristic dependencies. However, existing methods typically require extensive handcrafted training tasks to learn meta-strategies that generalize to target tasks, which poses a critical limitation for realistic applications with unknown task distributions. To overcome the issue, we propose the Adaptive meta Black-box Optimization Model (ABOM), which performs online parameter adaptation using solely optimization data from the target task, obviating the need for predefined task distributions. Unlike conventional metaBBO frameworks that decouple meta-training and optimization phases, ABOM introduces a closed-loop adaptive parameter learning mechanism, where parameterized evolutionary operators continuously self-update by leveraging generated populations during optimization. This paradigm shift enables zero-shot optimization: ABOM achieves competitive performance on synthetic BBO benchmarks and realistic unmanned aerial vehicle path planning problems without any handcrafted training tasks. Visualization studies reveal that parameterized evolutionary operators exhibit statistically significant search patterns, including natural selection and genetic recombination.
Authors: Tobias Schmidt, Kai Cui
Abstract: Mean-field control (MFC) offers a scalable solution to the curse of dimensionality in multi-agent systems but traditionally hinges on the restrictive assumption of exchangeability via dense, all-to-all interactions. In this work, we bridge the gap to real-world network structures by proposing a rigorous framework for MFC on large sparse graphs. We redefine the system state as a probability measure over decorated rooted neighborhoods, effectively capturing local heterogeneity. Our central contribution is a theoretical foundation for scalable reinforcement learning in this setting. We prove horizon-dependent locality: for finite-horizon problems, an agent's optimal policy at time t depends strictly on its (T-t)-hop neighborhood. This result renders the infinite-dimensional control problem tractable and underpins a novel Dynamic Programming Principle (DPP) on the lifted space of neighborhood distributions. Furthermore, we formally and experimentally justify the use of Graph Neural Networks (GNNs) for actor-critic algorithms in this context. Our framework naturally recovers classical MFC as a degenerate case while enabling efficient, theoretically grounded control on complex sparse topologies.
Authors: Kaiwei Yang, Lexiao Lai
Abstract: Norm-constrained linear minimization oracle (LMO)-based optimizers such as spectral gradient descent and Muon are attractive in large-scale learning, but extending them to manifold-constrained problems is nontrivial and often leads to nested-loop schemes that solve tangent-space subproblems iteratively. We propose \emph{Manifold Constrained Steepest Descent} (MCSD), a single-loop framework for optimization over manifolds that selects a norm-induced steepest-descent direction via an LMO applied to the Riemannian gradient, and then returns to the manifold via projection. Under standard smoothness assumptions, we establish convergence guarantees for MCSD and a stochastic momentum variant. We further introduce \emph{SPEL}, the spectral-norm specialization of MCSD on the Stiefel manifold, which admits scalable implementations via fast matrix sign computations. Experiments on PCA, orthogonality-constrained CNNs, and manifold-constrained LLM adapter tuning demonstrate improved stability and competitive performance relative to standard Riemannian baselines and existing manifold-aware LMO methods.
Authors: Junhong Cai, Guiqin Wang, Kejie Zhao, Jianxiong Tang, Xiang Wang, Luziwei Leng, Ran Cheng, Yuxin Ma, Qinghai Guo
Abstract: Large Language Models (LLMs) excel across diverse domains but suffer from high energy costs due to quadratic attention and dense Feed-Forward Network (FFN) operations. To address these issues, we propose Module-aware Architecture Refinement (MAR), a two-stage framework that integrates State Space Models (SSMs) for linear-time sequence modeling and applies activation sparsification to reduce FFN costs. In addition, to mitigate low information density and temporal mismatch in integrating Spiking Neural Networks (SNNs) with SSMs, we design the Adaptive Ternary Multi-step Neuron (ATMN) and the Spike-aware Bidirectional Distillation Strategy (SBDS). Extensive experiments demonstrate that MAR effectively restores the performance of its dense counterpart under constrained resources while substantially reducing inference energy consumption. Furthermore, it outperforms efficient models of comparable or even larger scale, underscoring its potential for building efficient and practical LLMs.
Authors: Wuyang Zhou, Yuxuan Gu, Giorgos Iacovides, Danilo Mandic
Abstract: The success of Hyper-Connections (HC) in neural networks (NN) has also highlighted issues related to its training instability and restricted scalability. The Manifold-Constrained Hyper-Connections (mHC) mitigate these challenges by projecting the residual connection space onto a Birkhoff polytope, however, it faces two issues: 1) its iterative Sinkhorn-Knopp (SK) algorithm does not always yield exact doubly stochastic residual matrices; 2) mHC incurs a prohibitive $\mathcal{O}(n^3C)$ parameter complexity with $n$ as the width of the residual stream and $C$ as the feature dimension. The recently proposed mHC-lite reparametrizes the residual matrix via the Birkhoff-von-Neumann theorem to guarantee double stochasticity, but also faces a factorial explosion in its parameter complexity, $\mathcal{O} \left( nC \cdot n! \right)$. To address both challenges, we propose \textbf{KromHC}, which uses the \underline{Kro}necker products of smaller doubly stochastic matrices to parametrize the residual matrix in \underline{mHC}. By enforcing manifold constraints across the factor residual matrices along each mode of the tensorized residual stream, KromHC guarantees exact double stochasticity of the residual matrices while reducing parameter complexity to $\mathcal{O}(n^2C)$. Comprehensive experiments demonstrate that KromHC matches or even outperforms state-of-the-art (SOTA) mHC variants, while requiring significantly fewer trainable parameters. The code is available at \texttt{https://github.com/wz1119/KromHC}.
Authors: Jonas Knupp, Jan Hendrik Metzen, Jeremias Bohn, Georg Groh, Kristian Kersting
Abstract: Depth-recurrence facilitates latent reasoning by sharing parameters across depths. However, prior work lacks combined FLOP-, parameter-, and memory-matched baselines, underutilizes depth-recurrence due to partially fixed layer stacks, and ignores the bottleneck of constant hidden-sizes that restricts many-step latent reasoning. To address this, we introduce a modular framework of depth-recurrent attention mixtures (Dreamer), combining sequence attention, depth attention, and sparse expert attention. It alleviates the hidden-size bottleneck through attention along depth, decouples scaling dimensions, and allows depth-recurrent models to scale efficiently and effectively. Across language reasoning benchmarks, our models require 2 to 8x fewer training tokens for the same accuracy as FLOP-, parameter-, and memory-matched SOTA, and outperform ca. 2x larger SOTA models with the same training tokens. We further present insights into knowledge usage across depths, e.g., showing 2 to 11x larger expert selection diversity than SOTA MoEs.
Authors: Shashiwadana Nirmania, Garima Sharma, Hourieh Khalajzadeh, Mojtaba Shahin
Abstract: In recent years, mobile applications have become indispensable tools for managing various aspects of life. From enhancing productivity to providing personalized entertainment, mobile apps have revolutionized people's daily routines. Despite this rapid growth and popularity, gaps remain in how these apps address the needs of users from different age groups. Users of varying ages face distinct challenges when interacting with mobile apps, from younger users dealing with inappropriate content to older users having difficulty with usability due to age-related vision and cognition impairments. Although there have been initiatives to create age-inclusive apps, a limited understanding of user perspectives on age-related issues may hinder developers from recognizing specific challenges and implementing effective solutions. In this study, we explore age discussions in app reviews to gain insights into how mobile apps should cater to users across different age groups.We manually curated a dataset of 4,163 app reviews from the Google Play Store and identified 1,429 age-related reviews and 2,734 non-age-related reviews. We employed eight machine learning, deep learning, and large language models to automatically detect age discussions, with RoBERTa performing the best, achieving a precision of 92.46%. Additionally, a qualitative analysis of the 1,429 age-related reviews uncovers six dominant themes reflecting user concerns.
Authors: Puwei Lian, Yujun Cai, Songze Li, Bingkun Bao
Abstract: Diffusion models have achieved remarkable progress in image generation, but their increasing deployment raises serious concerns about privacy. In particular, fine-tuned models are highly vulnerable, as they are often fine-tuned on small and private datasets. Membership inference attacks (MIAs) are used to assess privacy risks by determining whether a specific sample was part of a model's training data. Existing MIAs against diffusion models either assume obtaining the intermediate results or require auxiliary datasets for training the shadow model. In this work, we utilized a critical yet overlooked vulnerability: the widely used noise schedules fail to fully eliminate semantic information in the images, resulting in residual semantic signals even at the maximum noise step. We empirically demonstrate that the fine-tuned diffusion model captures hidden correlations between the residual semantics in initial noise and the original images. Building on this insight, we propose a simple yet effective membership inference attack, which injects semantic information into the initial noise and infers membership by analyzing the model's generation result. Extensive experiments demonstrate that the semantic initial noise can strongly reveal membership information, highlighting the vulnerability of diffusion models to MIAs.
Authors: Daniel Mills, Ifan Williams, Jacob Swain, Gabriel Matos, Enrico Rinaldi, Alexander Koziell-Pipe
Abstract: Many quantum software development kits provide a suite of circuit optimisation passes. These passes have been highly optimised and tested in isolation. However, the order in which they are applied is left to the user, or else defined in general-purpose default pass sequences. While general-purpose sequences miss opportunities for optimisation which are particular to individual circuits, designing pass sequences bespoke to particular circuits requires exceptional knowledge about quantum circuit design and optimisation. Here we propose and demonstrate training a reinforcement learning agent to compose optimisation-pass sequences. In particular the agent's action space consists of passes for two-qubit gate count reduction used in default PyTKET pass sequences. For the circuits in our diverse test set, the (mean, median) fraction of two-qubit gates removed by the agent is $(57.7\%, \ 56.7 \%)$, compared to $(41.8 \%, \ 50.0 \%)$ for the next best default pass sequence.
Authors: Eden Avrahami, Eliya Nachmani
Abstract: Discrete Diffusion Language Models (DLMs) offer a promising non-autoregressive alternative for text generation, yet effective mechanisms for inference-time control remain relatively underexplored. Existing approaches include sampling-level guidance procedures or trajectory optimization mechanisms. In this work, we introduce Iterative Latent Representation Refinement (ILRR), a learning-free framework for steering DLMs using a single reference sequence. ILRR guides generation by dynamically aligning the internal activations of the generated sequence with those of a given reference throughout the denoising process. This approach captures and transfers high-level semantic properties, with a tunable steering scale enabling flexible control over attributes such as sentiment. We further introduce Spatially Modulated Steering, an extension that enables steering long texts using shorter references by regulating guidance intensity across the sequence. Empirically, we demonstrate that ILRR achieves effective attribute steering on LLaDA and MDLM architectures with a minor computational overhead, requiring only one additional parallel forward pass per denoising step. Under the same compute budget, ILRR improves attribute accuracy over comparable baselines by 10$\%$ to 60$\%$ points, while maintaining high generation quality.
Authors: Zixuan Xia, Hao Wang, Pengcheng Weng, Yanyu Qian, Yangxin Xu, William Dan, Fei Wang
Abstract: Multimodal learning aims to integrate complementary information from heterogeneous modalities, yet strong optimization alone does not guaranty well-structured representations. Even under carefully balanced training schemes, multimodal models often exhibit geometric pathologies, including intra-modal representation collapse and sample-level cross-modal inconsistency, which degrade both unimodal robustness and multimodal fusion. We identify representation geometry as a missing control axis in multimodal learning and propose \regName, a lightweight geometry-aware regularization framework. \regName enforces two complementary constraints on intermediate embeddings: an intra-modal dispersive regularization that promotes representation diversity, and an inter-modal anchoring regularization that bounds sample-level cross-modal drift without rigid alignment. The proposed regularizer is plug-and-play, requires no architectural modifications, and is compatible with various training paradigms. Extensive experiments across multiple multimodal benchmarks demonstrate consistent improvements in both multimodal and unimodal performance, showing that explicitly regulating representation geometry effectively mitigates modality trade-offs.
Authors: Xiaoyu Xu, Minxin Du, Kun Fang, Zi Liang, Yaxin Xiao, Zhicong Huang, Cheng Hong, Qingqing Ye, Haibo Hu
Abstract: Large language models (LLMs) demonstrate impressive capabilities across diverse tasks but raise concerns about privacy, copyright, and harmful materials. Existing LLM unlearning methods rarely consider the continual and high-volume nature of real-world deletion requests, which can cause utility degradation and catastrophic forgetting as requests accumulate. To address this challenge, we introduce \fit, a framework for continual unlearning that handles large numbers of deletion requests while maintaining robustness against both catastrophic forgetting and post-unlearning recovery. \fit mitigates degradation through rigorous data \underline{F}iltering, \underline{I}mportance-aware updates, and \underline{T}argeted layer attribution, enabling stable performance across long sequences of unlearning operations and achieving a favorable balance between forgetting effectiveness and utility retention. To support realistic evaluation, we present \textbf{PCH}, a benchmark covering \textbf{P}ersonal information, \textbf{C}opyright, and \textbf{H}armful content in sequential deletion scenarios, along with two symmetric metrics, Forget Degree (F.D.) and Retain Utility (R.U.), which jointly assess forgetting quality and utility preservation. Extensive experiments on four open-source LLMs with hundreds of deletion requests show that \fit achieves the strongest trade-off between F.D. and R.U., surpasses existing methods on MMLU, CommonsenseQA, and GSM8K, and remains resistant against both relearning and quantization recovery attacks.
Authors: Xinglin Wang, Jiayi Shi, Shaoxiong Feng, Peiwen Yuan, Yiwei Li, Yueqi Zhang, Chuyi Tan, Ji Zhang, Boyuan Pan, Yao Hu, Kan Li
Abstract: Test-Time Scaling enhances the reasoning capabilities of Large Language Models by allocating additional inference compute to broaden the exploration of the solution space. However, existing search strategies typically treat rollouts as disposable samples, where valuable intermediate insights are effectively discarded after each trial. This systemic memorylessness leads to massive computational redundancy, as models repeatedly re-derive discovered conclusions and revisit known dead ends across extensive attempts. To bridge this gap, we propose \textbf{Recycling Search Experience (RSE)}, a self-guided, training-free strategy that turns test-time search from a series of isolated trials into a cumulative process. By actively distilling raw trajectories into a shared experience bank, RSE enables positive recycling of intermediate conclusions to shortcut redundant derivations and negative recycling of failure patterns to prune encountered dead ends. Theoretically, we provide an analysis that formalizes the efficiency gains of RSE, validating its advantage over independent sampling in solving complex reasoning tasks. Empirically, extensive experiments on HMMT24, HMMT25, IMO-Bench, and HLE show that RSE consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art scaling efficiency.
Authors: Liangze Jiang, Zachary Shinnick, Anton van den Hengel, Hemanth Saratchandran, Damien Teney
Abstract: Pretraining directly on web-scale corpora is the de facto paradigm for building language models. We study an alternative setting where the model is initially exposed to abstract structured data, as a means to ease the subsequent acquisition of rich semantic knowledge, much like humans learn simple logic and mathematics before higher reasoning. We specifically focus on procedural data, generated by formal languages and other simple algorithms, as such abstract data. We first diagnose the algorithmic skills that different forms of procedural data can improve, often significantly. For example, on context recall (Needle-in-a-haystack), the accuracy jumps from 10 to 98% when pretraining on Dyck sequences (balanced brackets). Second, we study how these gains are reflected in pretraining larger models (up to 1.3B). We find that front-loading as little as 0.1% procedural data significantly outperforms standard pretraining on natural language, code, and informal mathematics (C4, CodeParrot, and DeepMind-Math datasets). Notably, this procedural pretraining enables the models to reach the same loss value with only 55, 67, 86% of the original data. Third, we explore the mechanisms behind and find that procedural pretraining instils non-trivial structure in both attention and MLP layers. The former is particularly important for structured domains (e.g. code), and the latter for language. Finally, we lay a path for combining multiple forms of procedural data. Our results show that procedural pretraining is a simple, lightweight means to improving performance and accelerating language model pretraining, ultimately suggesting the promise of disentangling knowledge acquisition from reasoning in LLMs.
Authors: Zijie Xu, Zihan Huang, Yiting Dong, Kang Chen, Wenxuan Liu, Zhaofei Yu
Abstract: Spiking Neural Networks (SNNs) can achieve competitive performance by converting already existing well-trained Artificial Neural Networks (ANNs), avoiding further costly training. This property is particularly attractive in Reinforcement Learning (RL), where training through environment interaction is expensive and potentially unsafe. However, existing conversion methods perform poorly in continuous control, where suitable baselines are largely absent. We identify error amplification as the key cause: small action approximation errors become temporally correlated across decision steps, inducing cumulative state distribution shift and severe performance degradation. To address this issue, we propose Cross-Step Residual Potential Initialization (CRPI), a lightweight training-free mechanism that carries over residual membrane potentials across decision steps to suppress temporally correlated errors. Experiments on continuous control benchmarks with both vector and visual observations demonstrate that CRPI can be integrated into existing conversion pipelines and substantially recovers lost performance. Our results highlight continuous control as a critical and challenging benchmark for ANN-to-SNN conversion, where small errors can be strongly amplified and impact performance.
Authors: Francisco Caldas, Sahil Kumar, Cl\'audia Soares
Abstract: We introduce a model-agnostic forward diffusion process for time-series forecasting that decomposes signals into spectral components, preserving structured temporal patterns such as seasonality more effectively than standard diffusion. Unlike prior work that modifies the network architecture or diffuses directly in the frequency domain, our proposed method alters only the diffusion process itself, making it compatible with existing diffusion backbones (e.g., DiffWave, TimeGrad, CSDI). By staging noise injection according to component energy, it maintains high signal-to-noise ratios for dominant frequencies throughout the diffusion trajectory, thereby improving the recoverability of long-term patterns. This strategy enables the model to maintain the signal structure for a longer period in the forward process, leading to improved forecast quality. Across standard forecasting benchmarks, we show that applying spectral decomposition strategies, such as the Fourier or Wavelet transform, consistently improves upon diffusion models using the baseline forward process, with negligible computational overhead. The code for this paper is available at https://anonymous.4open.science/r/D-FDP-4A29.
Authors: Mingyuan Xu, Xinzi Tan, Jiawei Wu, Doudou Zhou
Abstract: Evaluating large language models (LLMs) on open-ended tasks without ground-truth labels is increasingly done via the LLM-as-a-judge paradigm. A critical but under-modeled issue is that judge LLMs differ substantially in reliability; treating all judges equally can yield biased leaderboards and misleading uncertainty estimates. More data can make evaluation more confidently wrong under misspecified aggregation. We propose a judge-aware ranking framework that extends the Bradley-Terry-Luce model by introducing judge-specific discrimination parameters, jointly estimating latent model quality and judge reliability from pairwise comparisons without reference labels. We establish identifiability up to natural normalizations and prove consistency and asymptotic normality of the maximum likelihood estimator, enabling confidence intervals for score differences and rank comparisons. Across multiple public benchmarks and a newly collected dataset, our method improves agreement with human preferences, achieves higher data efficiency than unweighted baselines, and produces calibrated uncertainty quantification for LLM rankings.
Authors: Daniel Gonzalez-Alvarado, Jonas Cassel, Stefania Petra, Christoph Schn\"orr
Abstract: We introduce the use of latent subspaces in the exponential parameter space of product manifolds of categorial distributions, as a tool for learning generative models of discrete data. The low-dimensional latent space encodes statistical dependencies and removes redundant degrees of freedom among the categorial variables. We equip the parameter domain with a Riemannian geometry such that the spaces and distances are related by isometries which enables consistent flow matching. In particular, geodesics become straight lines which makes model training by flow matching effective. Empirical results demonstrate that reduced latent dimensions suffice to represent data for generative modeling.
Authors: Ander Artola Velasco, Dimitrios Rontogiannis, Stratis Tsirtsis, Manuel Gomez-Rodriguez
Abstract: Test-time compute has emerged as a promising strategy to enhance the reasoning abilities of large language models (LLMs). However, this strategy has in turn increased how much users pay cloud-based providers offering LLM-as-a-service, since providers charge users for the amount of test-time compute they use to generate an output. In our work, we show that the market of LLM-as-a-service is socially inefficient: providers have a financial incentive to increase the amount of test-time compute, even if this increase contributes little to the quality of the outputs. To address this inefficiency, we introduce a reverse second-price auction mechanism where providers bid their offered price and (expected) quality for the opportunity to serve a user, and users pay proportionally to the marginal value generated by the winning provider relative to the second-highest bidder. To illustrate and complement our theoretical results, we conduct experiments with multiple instruct models from the $\texttt{Llama}$ and $\texttt{Qwen}$ families, as well as reasoning models distilled from $\texttt{DeepSeek-R1}$, on math and science benchmark datasets.
Authors: Elias J\"a\"asaari, Ville Hyv\"onen, Teemu Roos
Abstract: Multi-vector representations generated by late interaction models, such as ColBERT, enable superior retrieval quality compared to single-vector representations in information retrieval applications. In multi-vector retrieval systems, both queries and documents are encoded using one embedding for each token, and similarity between queries and documents is measured by the MaxSim similarity measure. However, the improved recall of multi-vector retrieval comes at the expense of significantly increased latency. This necessitates designing efficient approximate nearest neighbor search (ANNS) algorithms for multi-vector search. In this work, we introduce LEMUR, a simple-yet-efficient framework for multi-vector similarity search. LEMUR consists of two consecutive problem reductions: We first formulate multi-vector similarity search as a supervised learning problem that can be solved using a one-hidden-layer neural network. Second, we reduce inference under this model to single-vector similarity search in its latent space, which enables the use of existing single-vector ANNS methods for speeding up retrieval. In addition to performance evaluation on ColBERTv2 embeddings, we evaluate LEMUR on embeddings generated by modern multi-vector text models and multi-vector visual document retrieval models. LEMUR is an order of magnitude faster than earlier multi-vector similarity search methods.
Authors: Stanislas Strasman, Gabriel Cardoso, Sylvain Le Corff, Vincent Lemaire, Antonio Ocello
Abstract: Understanding the stability and long-time behavior of generative models is a fundamental problem in modern machine learning. This paper provides quantitative bounds on the sampling error of score-based generative models by leveraging stability and forgetting properties of the Markov chain associated with the reverse-time dynamics. Under weak assumptions, we provide the two structural properties to ensure the propagation of initialization and discretization errors of the backward process: a Lyapunov drift condition and a Doeblin-type minorization condition. A practical consequence is quantitative stability of the sampling procedure, as the reverse diffusion dynamics induces a contraction mechanism along the sampling trajectory. Our results clarify the role of stochastic dynamics in score-based models and provide a principled framework for analyzing propagation of errors in such approaches.
Authors: Arie Soeteman, Michael Benedikt, Martin Grohe, Balder ten Cate
Abstract: Graph neural networks (GNNs) are a widely used class of machine learning models for graph-structured data, based on local aggregation over neighbors. GNNs have close connections to logic. In particular, their expressive power is linked to that of modal logics and bounded-variable logics with counting. In many practical scenarios, graphs processed by GNNs have node features that act as unique identifiers. In this work, we study how such identifiers affect the expressive power of GNNs. We initiate a study of the key-invariant expressive power of GNNs, inspired by the notion of order-invariant definability in finite model theory: which node queries that depend only on the underlying graph structure can GNNs express on graphs with unique node identifiers? We provide answers for various classes of GNNs with local max- or sum-aggregation.
Authors: Gustav Nor\'en, Anubhab Ghosh, Fredrik Cumlin, Saikat Chatterjee
Abstract: We design a variational state estimation (VSE) method that provides a closed-form Gaussian posterior of an underlying complex dynamical process from (noisy) nonlinear measurements. The complex process is model-free. That is, we do not have a suitable physics-based model characterizing the temporal evolution of the process state. The closed-form Gaussian posterior is provided by a recurrent neural network (RNN). The use of RNN is computationally simple in the inference phase. For learning the RNN, an additional RNN is used in the learning phase. Both RNNs help each other learn better based on variational inference principles. The VSE is demonstrated for a tracking application - state estimation of a stochastic Lorenz system (a benchmark process) using a 2-D camera measurement model. The VSE is shown to be competitive against a particle filter that knows the Lorenz system model and a recently proposed data-driven state estimation method that does not know the Lorenz system model.
Authors: Amir Ali Ahmadi, Georgina Hall
Abstract: We show that computing even very coarse approximations of critical points is intractable for simple classes of nonconvex functions. More concretely, we prove that if there exists a polynomial-time algorithm that takes as input a polynomial in $n$ variables of constant degree (as low as three) and outputs a point whose gradient has Euclidean norm at most $2^n$ whenever the polynomial has a critical point, then P=NP. The algorithm is permitted to return an arbitrary point when no critical point exists. We also prove hardness results for approximate computation of critical points under additional structural assumptions, including settings in which existence and uniqueness of a critical point are guaranteed, the function is lower bounded, and approximation is measured in terms of distance to a critical point. Overall, our results stand in contrast to the commonly-held belief that, in nonconvex optimization, approximate computation of critical points is a tractable task.
Authors: Lev Fedorov, Micha\"el E. Sander, Romuald Elie, Pierre Marion, Mathieu Lauri\`ere
Abstract: Transformers have revolutionized deep learning across various domains but understanding the precise token dynamics remains a theoretical challenge. Existing theories of deep Transformers with layer normalization typically predict that tokens cluster to a single point; however, these results rely on deterministic weight assumptions, which fail to capture the standard initialization scheme in Transformers. In this work, we show that accounting for the intrinsic stochasticity of random initialization alters this picture. More precisely, we analyze deep Transformers where noise arises from the random initialization of value matrices. Under diffusion scaling and token-wise RMS normalization, we prove that, as the number of Transformer layers goes to infinity, the discrete token dynamics converge to an interacting-particle system on the sphere where tokens are driven by a \emph{common} matrix-valued Brownian noise. In this limit, we show that initialization noise prevents the collapse to a single cluster predicted by deterministic models. For two tokens, we prove a phase transition governed by the interaction strength and the token dimension: unlike deterministic attention flows, antipodal configurations become attracting with positive probability. Numerical experiments confirm the predicted transition, reveal that antipodal formations persist for more than two tokens, and demonstrate that suppressing the intrinsic noise degrades accuracy.
Authors: James Matthew Young, Paula Cordero-Encinar, Sebastian Reich, Andrew Duncan, O. Deniz Akyildiz
Abstract: We develop a diffusion-based sampler for target distributions known up to a normalising constant. To this end, we rely on the well-known diffusion path that smoothly interpolates between a (simple) base distribution and the target distribution, widely used in diffusion models. Our approach is based on a practical implementation of diffusion-annealed Langevin Monte Carlo, which approximates the diffusion path with convergence guarantees. We tackle the score estimation problem by developing an efficient sequential Monte Carlo sampler that evolves auxiliary variables from conditional distributions along the path, which provides principled score estimates for time-varying distributions. We further develop novel control variate schedules that minimise the variance of these score estimates. Finally, we provide theoretical guarantees and empirically demonstrate the effectiveness of our method on several synthetic and real-world datasets.
Authors: Yu-Wei Chen, Raghu Pasupathy, Jordan Awan
Abstract: We develop a near-optimal testing procedure under the framework of Gaussian differential privacy for simple as well as one- and two-sided tests under monotone likelihood ratio conditions. Our mechanism is based on a private mean estimator with data-driven clamping bounds, whose population risk matches the private minimax rate up to logarithmic factors. Using this estimator, we construct private test statistics that achieve the same asymptotic relative efficiency as the non-private, most powerful tests while maintaining conservative type I error control. In addition to our theoretical results, our numerical experiments show that our private tests outperform competing DP methods and offer comparable power to the non-private most powerful tests, even at moderately small sample sizes and privacy loss budgets.
Authors: Lorenzo Mazza, Ariel Rodriguez, Rayan Younis, Martin Lelis, Ortrun Hellig, Chenpan Li, Sebastian Bodenstedt, Martin Wagner, Stefanie Speidel
Abstract: Imitation learning has achieved remarkable success in robotic manipulation, yet its application to surgical robotics remains challenging due to data scarcity, constrained workspaces, and the need for an exceptional level of safety and predictability. We present a supervised Mixture-of-Experts (MoE) architecture designed for phase-structured surgical manipulation tasks, which can be added on top of any autonomous policy. Unlike prior surgical robot learning approaches that rely on multi-camera setups or thousands of demonstrations, we show that a lightweight action decoder policy like Action Chunking Transformer (ACT) can learn complex, long-horizon manipulation from less than 150 demonstrations using solely stereo endoscopic images, when equipped with our architecture. We evaluate our approach on the collaborative surgical task of bowel grasping and retraction, where a robot assistant interprets visual cues from a human surgeon, executes targeted grasping on deformable tissue, and performs sustained retraction. We benchmark our method against state-of-the-art Vision-Language-Action (VLA) models and the standard ACT baseline. Our results show that generalist VLAs fail to acquire the task entirely, even under standard in-distribution conditions. Furthermore, while standard ACT achieves moderate success in-distribution, adopting a supervised MoE architecture significantly boosts its performance, yielding higher success rates in-distribution and demonstrating superior robustness in out-of-distribution scenarios, including novel grasp locations, reduced illumination, and partial occlusions. Notably, it generalizes to unseen testing viewpoints and also transfers zero-shot to ex vivo porcine tissue without additional training, offering a promising pathway toward in vivo deployment. To support this, we present qualitative preliminary results of policy roll-outs during in vivo porcine surgery.
Authors: Nicolas Blin, Stefano Gualandi, Christopher Maes, Andrea Lodi, Bartolomeo Stellato
Abstract: We present a batched first-order method for solving multiple linear programs in parallel on GPUs. Our approach extends the primal-dual hybrid gradient algorithm to efficiently solve batches of related linear programming problems that arise in mixed-integer programming techniques such as strong branching and bound tightening. By leveraging matrix-matrix operations instead of repeated matrix-vector operations, we obtain significant computational advantages on GPU architectures. We demonstrate the effectiveness of our approach on various case studies and identify the problem sizes where first-order methods outperform traditional simplex-based solvers depending on the computational environment one can use. This is a significant step for the design and development of integer programming algorithms tightly exploiting GPU capabilities where we argue that some specific operations should be allocated to GPUs and performed in full instead of using light-weight heuristic approaches on CPUs.
Authors: Jianhui Chen, Yuzhang Luo, Liangming Pan
Abstract: While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.
Authors: James Cuin, Davide Carbone, Yanbo Tang, O. Deniz Akyildiz
Abstract: The problem of optimising functions with intractable gradients frequently arise in machine learning and statistics, ranging from maximum marginal likelihood estimation procedures to fine-tuning of generative models. Stochastic approximation methods for this class of problems typically require inner sampling loops to obtain (biased) stochastic gradient estimates, which rapidly becomes computationally expensive. In this work, we develop sequential Monte Carlo (SMC) samplers for optimisation of functions with intractable gradients. Our approach replaces expensive inner sampling methods with efficient SMC approximations, which can result in significant computational gains. We establish convergence results for the basic recursions defined by our methodology which SMC samplers approximate. We demonstrate the effectiveness of our approach on the reward-tuning of energy-based models within various settings.
Authors: Anand Babu, Rog\'erio Almeida Gouv\^ea, Pierre Vandergheynst, Gian-Marco Rignanese
Abstract: In this work, we present Multimodal Equivariant Inverse Design Network (MEIDNet), a framework that jointly learns structural information and materials properties through contrastive learning, while encoding structures via an equivariant graph neural network (EGNN). By combining generative inverse design with multimodal learning, our approach accelerates the exploration of chemical-structural space and facilitates the discovery of materials that satisfy predefined property targets. MEIDNet exhibits strong latent-space alignment with cosine similarity 0.96 by fusion of three modalities through cross-modal learning. Through implementation of curriculum learning strategies, MEIDNet achieves ~60 times higher learning efficiency than conventional training techniques. The potential of our multimodal approach is demonstrated by generating low-bandgap perovskite structures at a stable, unique, and novel (SUN) rate of 13.6 %, which are further validated by ab initio methods. Our inverse design framework demonstrates both scalability and adaptability, paving the way for the universal learning of chemical space across diverse modalities.
Authors: Sami Abuzakuk, Anne-Marie Kermarrec, Rishi Sharma, Rasmus Moorits Veski, Martijn de Vos
Abstract: Agentic AI enables LLM to dynamically reason, plan, and interact with tools to solve complex tasks. However, agentic workflows often require many iterative reasoning steps and tool invocations, leading to significant operational expense, end-to-end latency and failures due to hallucinations. This work introduces Agent Workflow Optimization (AWO), a framework that identifies and optimizes redundant tool execution patterns to improve the efficiency and robustness of agentic workflows. AWO analyzes existing workflow traces to discover recurring sequences of tool calls and transforms them into meta-tools, which are deterministic, composite tools that bundle multiple agent actions into a single invocation. Meta-tools bypass unnecessary intermediate LLM reasoning steps and reduce operational cost while also shortening execution paths, leading to fewer failures. Experiments on two agentic AI benchmarks show that AWO reduces the number of LLM calls up to 11.9% while also increasing the task success rate by up to 4.2 percent points.
Authors: Reza T. Batley, Sourav Saha
Abstract: Transformer scaling law analyses typically treat parameters as interchangeable; an abstraction that accurately predicts loss-compute relationships. Yet, in sub-billion-parameter small language models (SLMs), embedding matrices dominate the parameter budget. This work argues that this allocation is as suboptimal as it is counterintuitive. Leviathan is an architecture with a continuous embedding generator to replace the discrete lookup tables of canonical models. Evaluating on the Pile dataset under isoparametric settings, Leviathan consistently outperforms a standard, LLaMA-style architecture. By means of an empirical power-law fit, Leviathan exhibits a markedly superior effective parameter capacity. Across the regime studied, Leviathan behaves as a dense model with $1.47$ to $2.11 \times$ more parameters.
Authors: Naomi Pitzer, Daniela Mihai
Abstract: Emergent communication offers insight into how agents develop shared structured representations, yet most research assumes homogeneous modalities or aligned representational spaces, overlooking the perceptual heterogeneity of real-world settings. We study a heterogeneous multi-step binary communication game where agents differ in modality and lack perceptual grounding. Despite perceptual misalignment, multimodal systems converge to class-consistent messages grounded in perceptual input. Unimodal systems communicate more efficiently, using fewer bits and achieving lower classification entropy, while multimodal agents require greater information exchange and exhibit higher uncertainty. Bit perturbation experiments provide strong evidence that meaning is encoded in a distributional rather than compositional manner, as each bit's contribution depends on its surrounding pattern. Finally, interoperability analyses show that systems trained in different perceptual worlds fail to directly communicate, but limited fine-tuning enables successful cross-system communication. This work positions emergent communication as a framework for studying how agents adapt and transfer representations across heterogeneous modalities, opening new directions for both theory and experimentation.
Authors: Mahdi Nikdan, Amir Zandieh, Dan Alistarh, Vahab Mirrokni
Abstract: Quantization has significantly improved the compute and memory efficiency of Large Language Model (LLM) training. However, existing approaches still rely on accumulating their updates in high-precision: concretely, gradient updates must be applied to a high-precision weight buffer, known as $\textit{master weights}$. This buffer introduces substantial memory overhead, particularly for Sparse Mixture of Experts (SMoE) models, where model parameters and optimizer states dominate memory usage. To address this, we introduce the Error-Compensating Optimizer (ECO), which eliminates master weights by applying updates directly to quantized parameters. ECO quantizes weights after each step and carefully injects the resulting quantization error into the optimizer momentum, forming an error-feedback loop with no additional memory. We prove that, under standard assumptions and a decaying learning rate, ECO converges to a constant-radius neighborhood of the optimum, while naive master-weight removal can incur an error that is inversely proportional to the learning rate. We show empirical results for pretraining small Transformers (30-800M), a Gemma-3 1B model, and a 2.1B parameter Sparse MoE model with FP8 quantization, and fine-tuning DeepSeek-MoE-16B in INT4 precision. Throughout, ECO matches baselines with master weights up to near-lossless accuracy, significantly shifting the static memory vs validation loss Pareto frontier.
Authors: Robert de Witt, Mikko S. Pakkanen
Abstract: We present the first application of MAP-Elites, a quality-diversity algorithm, to trade execution. Rather than searching for a single optimal policy, MAP-Elites generates a diverse portfolio of regime-specialist strategies indexed by liquidity and volatility conditions. Individual specialists achieve 8-10% performance improvements within their behavioural niches, while other cells show degradation, suggesting opportunities for ensemble approaches that combine improved specialists with the baseline PPO policy. Results indicate that quality-diversity methods offer promise for regime-adaptive execution, though substantial computational resources per behavioural cell may be required for robust specialist development across all market conditions. To ensure experimental integrity, we develop a calibrated Gymnasium environment focused on order scheduling rather than tactical placement decisions. The simulator features a transient impact model with exponential decay and square-root volume scaling, fit to 400+ U.S. equities with R^2>0.02 out-of-sample. Within this environment, two Proximal Policy Optimization architectures - both MLP and CNN feature extractors - demonstrate substantial improvements over industry baselines, with the CNN variant achieving 2.13 bps arrival slippage versus 5.23 bps for VWAP on 4,900 out-of-sample orders ($21B notional). These results validate both the simulation realism and provide strong single-policy baselines for quality-diversity methods.
Authors: Han Yang, Dong Hao, Zhuohan Wang, Qi Shi, Xingtong Li
Abstract: Automatically discovering formulaic alpha factors is a central problem in quantitative finance. Existing methods often ignore syntactic and semantic constraints, relying on exhaustive search over unstructured and unbounded spaces. We present AlphaCFG, a grammar-based framework for defining and discovering alpha factors that are syntactically valid, financially interpretable, and computationally efficient. AlphaCFG uses an alpha-oriented context-free grammar to define a tree-structured, size-controlled search space, and formulates alpha discovery as a tree-structured linguistic Markov decision process, which is then solved using a grammar-aware Monte Carlo Tree Search guided by syntax-sensitive value and policy networks. Experiments on Chinese and U.S. stock market datasets show that AlphaCFG outperforms state-of-the-art baselines in both search efficiency and trading profitability. Beyond trading strategies, AlphaCFG serves as a general framework for symbolic factor discovery and refinement across quantitative finance, including asset pricing and portfolio construction.
Authors: John Flynn, Wolfgang Paier, Dimitar Dinev, Sam Nhut Nguyen, Hayk Poghosyan, Manuel Toribio, Sandipan Banerjee, Guy Gafni
Abstract: Current generative video models excel at producing novel content from text and image prompts, but leave a critical gap in editing existing pre-recorded videos, where minor alterations to the spoken script require preserving motion, temporal coherence, speaker identity, and accurate lip synchronization. We introduce EditYourself, a DiT-based framework for audio-driven video-to-video (V2V) editing that enables transcript-based modification of talking head videos, including the seamless addition, removal, and retiming of visually spoken content. Building on a general-purpose video diffusion model, EditYourself augments its V2V capabilities with audio conditioning and region-aware, edit-focused training extensions. This enables precise lip synchronization and temporally coherent restructuring of existing performances via spatiotemporal inpainting, including the synthesis of realistic human motion in newly added segments, while maintaining visual fidelity and identity consistency over long durations. This work represents a foundational step toward generative video models as practical tools for professional video post-production.
Authors: Yifeng Ding, Lingming Zhang
Abstract: Test-time scaling has been widely adopted to enhance the capabilities of Large Language Model (LLM) agents in software engineering (SWE) tasks. However, the standard approach of repeatedly sampling trajectories from scratch is computationally expensive. While recent methods have attempted to mitigate costs using specialized value agents, they can suffer from model miscalibration and fail to generalize to modern agents that synthesize custom bash scripts as tools. In this paper, we introduce SWE-Replay, the first efficient and generalizable test-time scaling technique for modern agents without reliance on potentially noisy value estimates. SWE-Replay optimizes the scaling process by recycling trajectories from prior trials, dynamically choosing to either explore from scratch or exploit archived experience by branching at critical intermediate steps. This selection of intermediate steps is driven by the potential and reasoning significance of repository exploration, rather than external LLM-based quality estimates. Our evaluation shows that, on SWE-Bench Verified, SWE-Replay consistently outperforms naive scaling, reducing costs by up to 17.4% while maintaining or even improving performance by up to 3.8%. Further evaluation on SWE-Bench Pro and Multilingual validates the generalizability of SWE-Replay, establishing it as a robust foundation for efficient test-time scaling of software engineering agents.
Authors: Grzegorz Stefanski, Alberto Presta, Michal Byra
Abstract: In pruning, the Lottery Ticket Hypothesis posits that large networks contain sparse subnetworks, or winning tickets, that can be trained in isolation to match the performance of their dense counterparts. However, most existing approaches assume a single universal winning ticket shared across all inputs, ignoring the inherent heterogeneity of real-world data. In this work, we propose Routing the Lottery (RTL), an adaptive pruning framework that discovers multiple specialized subnetworks, called adaptive tickets, each tailored to a class, semantic cluster, or environmental condition. Across diverse datasets and tasks, RTL consistently outperforms single- and multi-model baselines in balanced accuracy and recall, while using up to 10 times fewer parameters than independent models and exhibiting semantically aligned. Furthermore, we identify subnetwork collapse, a performance drop under aggressive pruning, and introduce a subnetwork similarity score that enables label-free diagnosis of oversparsification. Overall, our results recast pruning as a mechanism for aligning model structure with data heterogeneity, paving the way toward more modular and context-aware deep learning.
Authors: Ajay Patel, Colin Raffel, Chris Callison-Burch
Abstract: Due to limited supervised training data, large language models (LLMs) are typically pre-trained via a self-supervised "predict the next word" objective on a vast amount of unstructured text data. To make the resulting model useful to users, it is further trained on a far smaller amount of "instruction-tuning" data comprised of supervised training examples of instructions and responses. To overcome the limited amount of supervised data, we propose a procedure that can transform the knowledge in internet-scale pre-training documents into billions of synthetic instruction and answer training pairs. The resulting dataset, called FineInstructions, uses ~18M instruction templates created from real user-written queries and prompts. These instruction templates are matched to and instantiated with human-written source documents from unstructured pre-training corpora. With "supervised" synthetic training data generated at this scale, an LLM can be pre-trained from scratch solely with the instruction-tuning objective, which is far more in-distribution with the expected downstream usage of LLMs (responding to user prompts). We conduct controlled token-for-token training experiments and find pre-training on FineInstructions outperforms standard pre-training and other proposed synthetic pre-training techniques on standard benchmarks measuring free-form response quality. Our resources can be found at https://huggingface.co/fineinstructions .
Authors: Yingfa Chen, Zhen Leng Thai, Zihan Zhou, Zhu Zhang, Xingyu Shen, Shuo Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu
Abstract: Hybrid Transformer architectures, which combine softmax attention blocks and recurrent neural networks (RNNs), have shown a desirable performance-throughput tradeoff for long-context modeling, but their adoption and studies are hindered by the prohibitive cost of large-scale pre-training from scratch. Some recent studies have shown that pre-trained softmax attention blocks can be converted into RNN blocks through parameter transfer and knowledge distillation. However, these transfer methods require substantial amounts of training data (more than 10B tokens), and the resulting hybrid models also exhibit poor long-context performance, which is the scenario where hybrid models enjoy significant inference speedups over Transformer-based models. In this paper, we present HALO (Hybrid Attention via Layer Optimization), a pipeline for distilling Transformer models into RNN-attention hybrid models. We then present HypeNet, a hybrid architecture with superior length generalization enabled by a novel position encoding scheme (named HyPE) and various architectural modifications. We convert the Qwen3 series into HypeNet using HALO, achieving performance comparable to the original Transformer models while enjoying superior long-context performance and efficiency. The conversion requires just 2.3B tokens, less than 0.01% of their pre-training data
Authors: Gwendal Debaussart-Joniec, Harry Sevi, Matthieu Jonckheere, Argyris Kalogeratos
Abstract: Vertex-level clustering for directed graphs (digraphs) remains challenging as edge directionality breaks the key assumptions underlying popular spectral methods, which also incur the overhead of eigen-decomposition. This paper proposes Parametrized Power-Iteration Clustering (ParPIC), a random-walk-based clustering method for weakly connected digraphs. This builds over the Power-Iteration Clustering paradigm, which uses the rows of the iterated diffusion operator as a data embedding. ParPIC has three important features: the use of parametrized reversible random walk operators, the automatic tuning of the diffusion time, and the efficient truncation of the final embedding, which produces low-dimensional data representations and reduces complexity. Empirical results on synthetic and real-world graphs demonstrate that ParPIC achieves competitive clustering accuracy with improved scalability relative to spectral and teleportation-based methods.
Authors: Ruiqi Wang, Diego Klabjan
Abstract: Stochastic optimization algorithms using exponential moving averages of the past gradients, such as ADAM, RMSProp and AdaGrad, have been having great successes in many applications, especially in training deep neural networks. ADAM in particular stands out as efficient and robust. Despite of its outstanding performance, ADAM has been proved to be divergent for some specific problems. We revisit the divergent question and provide divergent examples under stronger conditions such as in expectation or high probability. Under a variance reduction assumption, we show that an ADAM-type algorithm converges, which means that it is the variance of gradients that causes the divergence of original ADAM. To this end, we propose a variance reduced version of ADAM and provide a convergent analysis of the algorithm. Numerical experiments show that the proposed algorithm has as good performance as ADAM. Our work suggests a new direction for fixing the convergence issues.
Authors: RuiQi Liu, Boyu Diao, Libo Huang, Zijia An, Hangda Liu, Zhulin An, Yongjun Xu
Abstract: Continual learning (CL) aims to learn new tasks without erasing previous knowledge. However, current CL methods primarily emphasize improving accuracy while often neglecting training efficiency, which consequently restricts their practical application. Drawing inspiration from the brain's contextual gating mechanism, which selectively filters neural information and continuously updates past memories, we propose Low-redundancy Distillation (LoRD), a novel CL method that enhances model performance while maintaining training efficiency. This is achieved by eliminating redundancy in three aspects of CL: student model redundancy, teacher model redundancy, and rehearsal sample redundancy. By compressing the learnable parameters of the student model and pruning the teacher model, LoRD facilitates the retention and optimization of prior knowledge, effectively decoupling task-specific knowledge without manually assigning isolated parameters for each task. Furthermore, we optimize the selection of rehearsal samples and refine rehearsal frequency to improve training efficiency. Through a meticulous design of distillation and rehearsal strategies, LoRD effectively balances training efficiency and model precision. Extensive experimentation across various benchmark datasets and environments demonstrates LoRD's superiority, achieving the highest accuracy with the lowest training FLOPs.
Authors: Debolina Halder Lina, Arlei Silva
Abstract: Graph Neural Networks (GNNs) have achieved state-of-the-art results in many relevant tasks where decisions might disproportionately impact specific communities. However, existing work on fair GNNs often assumes that either sensitive attributes are fully observed or they are missing completely at random. We show that an adversarial missingness process can inadvertently disguise a fair model through the imputation, leading the model to overestimate the fairness of its predictions. We address this challenge by proposing Better Fair than Sorry (BFtS), a fair missing data imputation model for sensitive attributes. The key principle behind BFtS is that imputations should approximate the worst-case scenario for fairness -- i.e. when optimizing fairness is the hardest. We implement this idea using a 3-player adversarial scheme where two adversaries collaborate against a GNN classifier, and the classifier minimizes the maximum bias. Experiments using synthetic and real datasets show that BFtS often achieves a better fairness x accuracy trade-off than existing alternatives under an adversarial missingness process.
Authors: Ariel Neufeld, Philipp Schmocker
Abstract: We introduce a Banach space-valued extension of random feature learning, a data-driven supervised machine learning technique for large-scale kernel approximation. By randomly initializing the feature maps, only the linear readout needs to be trained, which reduces the computational complexity substantially. Viewing random feature models as Banach space-valued random variables, we prove a universal approximation result in the corresponding Bochner space. Moreover, we derive approximation rates and an explicit algorithm to learn an element of the given Banach space by such models. The framework of this paper includes random trigonometric/Fourier regression and in particular random neural networks which are single-hidden-layer feedforward neural networks whose weights and biases are randomly initialized, whence only the linear readout needs to be trained. For the latter, we can then lift the universal approximation property of deterministic neural networks to random neural networks, even within function spaces over non-compact domains, e.g., weighted spaces, $L^p$-spaces, and (weighted) Sobolev spaces, where the latter includes the approximation of the (weak) derivatives. In addition, we analyze when the training costs for approximating a given function grow polynomially in both the input/output dimension and the reciprocal of a pre-specified tolerated approximation error. Furthermore, we demonstrate in a numerical example the empirical advantages of random feature models over their deterministic counterparts.
Authors: Yuhang Liu, Zhen Zhang, Dong Gong, Erdun Gao, Biwei Huang, Mingming Gong, Anton van den Hengel, Kun Zhang, Javen Qinfeng Shi
Abstract: Causal representation learning (CRL) offers the promise of uncovering the underlying causal model by which observed data was generated, but the practical applicability of existing methods remains limited by the strong assumptions required for identifiability and by challenges in applying them to real-world settings. Most current approaches are applicable only to relatively restrictive model classes, such as linear or polynomial models, which limits their flexibility and robustness in practice. One promising approach to this problem seeks to address these issues by leveraging changes in causal influences among latent variables. In this vein we propose a more general and relaxed framework than typically applied, formulated by imposing constraints on the function classes applied. Within this framework, we establish partial identifiability results under weaker conditions, including scenarios where only a subset of causal influences change. We then extend our analysis to a broader class of latent post-nonlinear models. Building on these theoretical insights, we develop a flexible method for learning latent causal representations. We demonstrate the effectiveness of our approach on synthetic and semi-synthetic datasets, and further showcase its applicability in a case study on human motion analysis, a complex real-world domain that also highlights the potential to broaden the practical reach of identifiable CRL models.
Authors: Jiyoun Kim, Junu Kim, Kyunghoon Hur, Edward Choi
Abstract: The increasing volume of electronic health records (EHRs) presents the opportunity to improve the accuracy and robustness of models in clinical prediction tasks. Unlike traditional centralized approaches, federated learning enables training on data from multiple institutions while preserving patient privacy and complying with regulatory constraints. In practice, healthcare institutions (i.e., hosts) often need to build predictive models tailored to their specific needs (e.g., creatinine-level prediction, N-day readmission prediction) using federated learning. When building a federated learning model for a single healthcare institution, two key challenges arise: (1) ensuring compatibility across heterogeneous EHR systems, and (2) managing federated learning costs within budget constraints. Specifically, heterogeneity in EHR systems across institutions hinders compatible modeling, while the computational costs of federated learning can exceed practical budget limits for healthcare institutions. To address these challenges, we propose EHRFL, a federated learning framework designed for building a cost-effective, host-specific predictive model using patient EHR data. EHRFL consists of two components: (1) text-based EHR modeling, which facilitates cross-institution compatibility without costly data standardization, and (2) a participant selection strategy based on averaged patient embedding similarity to reduce the number of participants without degrading performance. Our participant selection strategy sharing averaged patient embeddings is differentially private, ensuring patient privacy. Experiments on multiple open-source EHR datasets demonstrate the effectiveness of both components. With our framework, healthcare institutions can build institution-specific predictive models under budgetary constraints with reduced costs and time.
Authors: Niklas Freymuth, Philipp Dahlinger, Tobias W\"urth, Simon Reisch, Luise K\"arger, Gerhard Neumann
Abstract: Simulating physical systems is essential in engineering, but analytical solutions are limited to straightforward problems. Consequently, numerical methods like the Finite Element Method (FEM) are widely used. However, the FEM becomes computationally expensive as problem complexity and accuracy demands increase. Adaptive Mesh Refinement (AMR) improves the FEM by dynamically placing mesh elements on the domain, balancing computational speed and accuracy. Classical AMR depends on heuristics or expensive error estimators, which may lead to suboptimal performance for complex simulations. While AMR methods based on machine learning are promising, they currently only scale to simple problems. In this work, we formulate AMR as a system of collaborating, homogeneous agents that iteratively split into multiple new agents. This agent-wise perspective enables a spatial reward formulation focused on reducing the maximum mesh element error. Our approach, Adaptive Swarm Mesh Refinement++ (ASMR++), offers efficient, stable optimization and generates highly adaptive meshes at user-defined resolution at inference time. Extensive experiments demonstrate that ASMR++ outperforms heuristic approaches and learned baselines, matching the performance of expensive error-based oracle AMR strategies. ASMR additionally generalizes to different domains during inference, and produces meshes that simulate up to 2 orders of magnitude faster than uniform refinements in more demanding settings.
Authors: Mohammad Zhalechian, Soroush Saghafian, Omar Robles
Abstract: The United States Food and Drug Administration's (FDA's) 510(k) pathway allows manufacturers to gain medical device approval by demonstrating substantial equivalence to a legally marketed device. However, the inherent ambiguity of this regulatory procedure has been associated with high recall among many devices cleared through this pathway, raising significant safety concerns. In this paper, we develop a combined human-algorithm approach to assist the FDA in improving its 510(k) medical device clearance process by reducing recall risk and regulatory workload. We first develop machine learning methods to estimate the risk of recall of 510(k) medical devices based on the information available at the time of submission. We then propose a data-driven clearance policy that recommends acceptance, rejection, or deferral to FDA's committees for in-depth evaluation. We conduct an empirical study using a unique dataset of over 31,000 submissions that we assembled based on data sources from the FDA and Centers for Medicare and Medicaid Service (CMS). Compared to the FDA's current practice, which has a recall rate of 10.3% and a normalized workload measure of 100%, a conservative evaluation of our policy shows a 32.9% improvement in the recall rate and a 40.5% reduction in the workload. Our analyses further suggest annual cost savings of approximately $1.7 billion for the healthcare system driven by avoided replacement costs, which is equivalent to 1.1% of the entire United States annual medical device expenditure. Our findings highlight the value of a holistic and data-driven approach to improve the FDA's current 510(k) pathway.
Authors: Krzysztof Olejniczak, Xingyue Huang, Mikhail Galkin, \.Ismail \.Ilkan Ceylan
Abstract: Motivated by the incompleteness of modern knowledge graphs, a new setup for query answering has emerged, where the goal is to predict answers that do not necessarily appear in the knowledge graph, but are present in its completion. In this paper, we formally introduce and study two query answering problems, namely, query answer classification and query answer retrieval. To solve these problems, we propose AnyCQ, a model that can classify answers to any conjunctive query on any knowledge graph. At the core of our framework lies a graph neural network trained using a reinforcement learning objective to answer Boolean queries. Trained only on simple, small instances, AnyCQ generalizes to large queries of arbitrary structure, reliably classifying and retrieving answers to queries that existing approaches fail to handle. This is empirically validated through our newly proposed, challenging benchmarks. Finally, we empirically show that AnyCQ can effectively transfer to completely novel knowledge graphs when equipped with an appropriate link prediction model, highlighting its potential for querying incomplete data.
Authors: Jie Cheng, Ruixi Qiao, Yingwei Ma, Binhua Li, Gang Xiong, Qinghai Miao, Yongbin Li, Yisheng Lv
Abstract: A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization. We will release codes and model weights at https://github.com/CJReinforce/JOWA
Authors: Guanyu Cui, Yuhe Guo, Zhewei Wei, Hsin-Hao Su
Abstract: The expressive power of graph neural networks (GNNs) is often studied through their relationship to the Weisfeiler-Lehman (WL) tests. Despite its influence, this perspective leaves two gaps: (i) it is unclear whether WL tests are sufficiently primitive for understanding GNN expressivity, and (ii) WL-induced equivalence does not align well with characterizing the function classes that GNNs can approximate or compute. We attempt to address both gaps. First, we strengthen hardness results for the vanilla WL test, showing that in many settings it is not primitive enough to be implemented by constant-depth GNNs. Second, we propose an alternative framework for studying GNN expressivity based on an extended CONGEST model with an explicit preprocessing phase. Within this framework, we identify implicit shortcuts introduced in prior analyses and establish further results for WL tests in settings where graphs are augmented with virtual nodes and virtual edges.
Authors: Jie Xiao, Qianyi Huang, Xu Chen, Chen Tian
Abstract: As large language models (LLMs) increasingly integrate into every aspect of our work and daily lives, there are growing concerns about user privacy, which push the trend toward local deployment of these models. There are a number of lightweight LLMs (e.g., Gemini Nano, LLAMA2 7B) that can run locally on smartphones, providing users with greater control over their personal data. As a rapidly emerging application, we are concerned about their performance on commercial-off-the-shelf mobile devices. To fully understand the current landscape of LLM deployment on mobile platforms, we conduct a comprehensive measurement study on mobile devices. While user experience is the primary concern for end-users, developers focus more on the underlying implementations. Therefore, we evaluate both user-centric metrics-such as token throughput, latency, and response quality-and developer-critical factors, including resource utilization, OS strategies, battery consumption, and launch time. We also provide comprehensive comparisons across the mobile system-on-chips (SoCs) from major vendors, highlighting their performance differences in handling LLM workloads, which may help developers identify and address bottlenecks for mobile LLM applications. We hope that this study can provide insights for both the development of on-device LLMs and the design for future mobile system architecture.
Authors: Loris Gaven, Clement Romac, Thomas Carta, Sylvain Lamprier, Olivier Sigaud, Pierre-Yves Oudeyer
Abstract: The past years have seen Large Language Models (LLMs) strive not only as generative models but also as agents solving textual sequential decision-making tasks. When facing complex environments where their zero-shot abilities are insufficient, recent work showed online Reinforcement Learning (RL) could be used for the LLM agent to discover and learn efficient strategies interactively. However, most prior work sticks to on-policy algorithms, which greatly reduces the scope of methods such agents could use for both exploration and exploitation, such as experience replay and hindsight relabeling. Yet, such methods may be key for LLM learning agents, and in particular when designing autonomous intrinsically motivated agents sampling and pursuing their own goals (i.e. autotelic agents). This paper presents and studies an adaptation of Soft Actor-Critic and hindsight relabeling to LLM agents. Our method not only paves the path towards autotelic LLM agents that learn online but can also outperform on-policy methods in more classic multi-goal RL environments.
Authors: Tom Pelletreau-Duris, Ruud van Bakel, Michael Cochez
Abstract: Deep neural networks (DNNs) achieve state-of-the-art performance on many tasks, but this often requires increasingly larger model sizes, which in turn leads to more complex internal representations. Explainability techniques (XAI) have made remarkable progress in the interpretability of ML models. However, the non-euclidean nature of Graph Neural Networks (GNNs) makes it difficult to reuse already existing XAI methods. While other works have focused on instance-based explanation methods for GNNs, very few have investigated model-based methods and, to our knowledge, none have tried to probe the embedding of the GNNs for structural graph properties. In this paper we present a model agnostic explainability pipeline for Graph Neural Networks (GNNs) employing diagnostic classifiers. We propose to consider graph-theoretic properties as the features of choice for studying the emergence of representations in GNNs. This pipeline aims to probe and interpret the learned representations in GNNs across various architectures and datasets, refining our understanding and trust in these models.
Authors: Parikshit Bansal, Ali Kavis, Sujay Sanghavi
Abstract: Self-supervised learning attempts to learn representations from un-labeled data; it does so via a loss function that encourages the embedding of a point to be close to that of its augmentations. This simple idea performs remarkably well, yet it is not precisely theoretically understood why this is the case. In this paper we analyze self-supervised learning in a natural context: dimensionality reduction in Gaussian Mixture Models. Crucially, we define an augmentation of a data point as being another independent draw from the same underlying mixture component. We show that vanilla contrastive learning (specifically, the InfoNCE loss) is able to find the optimal lower-dimensional subspace even when the Gaussians are not isotropic -- something that vanilla spectral techniques cannot do. We also prove a similar result for "non-contrastive" self-supervised learning (i.e., SimSiam loss). We further extend our analyses to multi-modal contrastive learning algorithms (e.g., CLIP). In this setting we show that contrastive learning learns the subset of fisher-optimal subspace, effectively filtering out all the noise from the learnt representations. Finally, we corroborate our theoretical finding through synthetic data experiments.
Authors: Agnish Kumar Behera, Alexandra Lamtyugina, Aditya Nandy, Daiki Goto, Carlos Floyd, Suriyanarayanan Vaikuntanathan
Abstract: Generative diffusion models have emerged as powerful tools for sampling high-dimensional distributions, yet they typically rely on white gaussian noise and noise schedules to destroy and reconstruct information. Here, we demonstrate that driving the generative process out of equilibrium using active, temporally correlated noise sources fundamentally alters the information thermodynamics of the system. We show that coupling the data to an active non-Markovian bath creates a `memory effect' where high-level semantic information (such as class identity or molecular metastability) is stored in the temporal correlations of auxiliary degrees of freedom. Using Fisher information analysis, we prove that this active mechanism significantly retards the rate of information decay compared to passive Brownian motion. Crucially, this memory effect facilitates an earlier and more robust symmetry breaking (speciation) during the reverse generative process, allowing the system to resolve multi-scale structures, reminiscent of metastable states in molecular configurations that are washed out in the typical noising processes. Our results suggest that non-equilibrium protocols, inspired by active matter physics, offer a thermodynamically distinct and potentially advantageous pathway for recovering high-dimensional energy landscapes using generative diffusion.
Authors: Dae-Young Yun, Hee-Youl Kwak, Yongjune Kim, Sang-Hyo Kim, Jong-Seon No
Abstract: In this paper, we propose a neural window decoder (NWD) for spatially coupled low-density parity-check (SC-LDPC) codes. The proposed NWD retains the conventional window decoder (WD) process but incorporates trainable neural weights. To train the weights of NWD, we introduce two novel training strategies. First, we restrict the loss function to target variable nodes (VNs) of the window, which prunes the neural network and accordingly enhances training efficiency. Second, we employ the active learning technique with a normalized loss term to prevent the training process from biasing toward specific training regions. Next, we develop a systematic method to derive non-uniform schedules for the NWD based on the training results. We introduce trainable damping factors that reflect the relative importance of check node (CN) updates. By skipping updates with less importance, we can omit $\mathbf{41\%}$ of CN updates without performance degradation compared to the conventional WD. Lastly, we address the error propagation problem inherent in SC-LDPC codes by deploying a complementary weight set, which is activated when an error is detected in the previous window. This adaptive decoding strategy effectively mitigates error propagation without requiring modifications to the code and decoder structures.
Authors: Christopher Homan, Flip Korn, Deepak Pandita, Chris Welty
Abstract: A cornerstone of machine learning evaluation is the (often hidden) assumption that model and human responses are reliable enough to evaluate models against unitary, authoritative, ``gold standard'' data, via simple metrics such as accuracy, precision, and recall. The generative AI revolution would seem to explode this assumption, given the critical role stochastic inference plays. Yet, in spite of public demand for more transparency in AI -- along with strong evidence that humans are unreliable judges -- estimates of model reliability are conventionally based on, at most, a few output responses per input item. We adapt a method, previously used to evaluate the reliability of various metrics and estimators for machine learning evaluation, to determine whether an (existing or planned) dataset has enough responses per item to assure reliable null hypothesis statistical testing. We show that, for many common metrics, collecting even 5-10 responses per item (from each model and team of human evaluators) is not sufficient. We apply our methods to several of the very few extant gold standard test sets with multiple disaggregated responses per item and show that even these datasets lack enough responses per item. We show how our methods can help AI researchers make better decisions about how to collect data for AI evaluation.
Authors: Pedram Khorsandi, Rushil Gupta, Mehrnaz Mofakhami, Simon Lacoste-Julien, Gauthier Gidel
Abstract: Performative prediction is a framework accounting for the shift in the data distribution induced by the prediction of a model deployed in the real world. Ensuring rapid convergence to a stable solution where the data distribution remains the same after the model deployment is crucial, especially in evolving environments. This paper extends the Repeated Risk Minimization (RRM) framework by utilizing historical datasets from previous retraining snapshots, yielding a class of algorithms that we call Affine Risk Minimizers and enabling convergence to a performatively stable point for a broader class of problems. We introduce a new upper bound for methods that use only the final iteration of the dataset and prove for the first time the tightness of both this new bound and the previous existing bounds within the same regime. We also prove that utilizing historical datasets can surpass the lower bound for last iterate RRM, and empirically observe faster convergence to the stable point on various performative prediction benchmarks. We offer at the same time the first lower bound analysis for RRM within the class of Affine Risk Minimizers, quantifying the potential improvements in convergence speed that could be achieved with other variants in our framework.
Authors: Yuzong Chen, Xilai Dai, Jake Hyun, Chi-Chih Chang, Wonsuk Jang, Yuheng Wu, Thierry Tambe, Jae-sun Seo, Mohamed S. Abdelfattah
Abstract: The recently introduced NVFP4 format demonstrates remarkable performance and memory benefits for quantized large language model (LLM) inference. However, we observe two types of redundancy in NVFP4 encoding: (1) The FP4 element format naturally exposes an unused quantization value due to its sign-magnitude representation that contains both positive and negative zeros. (2) The FP8 block scaling factor has an unused sign bit because it is always positive. Additionally, we find that LLM weights are more tolerant to a lower-precision block scaling factor. Based on these observations, we propose Redundant Zero Remapping (RaZeR), an enhanced numerical format that pushes the limits of NVFP4 for more accurate LLM quantization under the same memory footprint. RaZeR leverages the redundant bits of the block scaling factor to adaptively remap the redundant FP4 zero to additional quantization values with improved accuracy. To demonstrate the practicality of RaZeR, we design efficient GPU kernels for RaZeR-quantized LLM inference and propose novel hardware to natively support this. Extensive experiments validate RaZeR's superior performance for 4-bit LLM quantization. For example, relative to native NVFP4, RaZeR reduces the average perplexity loss by 34.6% and 31.2% under weight-only and weight-activation quantization, respectively.
Authors: Han Fang, Paul Weng, Yutong Ban
Abstract: Deep Reinforcement Learning (DRL) has emerged as a promising approach for solving Combinatorial Optimization (CO) problems, such as the 3D Bin Packing Problem (3D-BPP), Traveling Salesman Problem (TSP), or Vehicle Routing Problem (VRP), but these neural solvers often exhibit brittleness when facing distribution shifts. To address this issue, we uncover the Satisficing Generalization Edge, which we validate both theoretically and experimentally: identifying a set of promising actions is inherently more generalizable than selecting the single optimal action. To exploit this property, we propose Adaptive Selection After Proposal (ASAP), a generic framework that decomposes the decision-making process into two distinct phases: a proposal policy that acts as a robust filter, and a selection policy as an adaptable decision maker. This architecture enables a highly effective online adaptation strategy where the selection policy can be rapidly fine-tuned on a new distribution. Concretely, we introduce a two-phase training framework enhanced by Model-Agnostic Meta-Learning (MAML) to prime the model for fast adaptation. Extensive experiments on 3D-BPP, TSP, and CVRP demonstrate that ASAP improves the generalization capability of state-of-the-art baselines and achieves superior online adaptation on out-of-distribution instances.
Authors: Siddhartha Parupudi, Gourab Ghatak
Abstract: We consider the best arm identification (BAI) problem in the $K-$armed bandit framework with a modification - the agent is allowed to play a subset of arms at each time slot instead of one arm. Consequently, the agent observes the sample average of the rewards of the arms that constitute the probed subset. Several trade-offs arise here - e.g., sampling a larger number of arms together results in a wider view of the environment, while sampling fewer arms enhances the information about individual reward distributions. Furthermore, grouping a large number of suboptimal arms together albeit reduces the variance of the reward of the group, it may enhance the group mean to make it close to that containing the optimal arm. To solve this problem, we propose an algorithm that constructs $\log_2 K$ groups and performs a likelihood ratio test to detect the presence of the best arm in each of these groups. Then a Hamming decoding procedure determines the unique best arm. We derive an upper bound for the error probability of the proposed algorithm based on a new hardness parameter $H_4$. Finally, we demonstrate cases under which it outperforms the state-of-the-art algorithms for the single play case.
Authors: Yousef Koka, David Selby, Gerrit Gro{\ss}mann, Sebastian Vollmer, Kathan Pandya
Abstract: Data preprocessing is a critical yet frequently neglected aspect of machine learning, often paid little attention despite its potentially significant impact on model performance. While automated machine learning pipelines are starting to recognize and integrate data preprocessing into their solutions for classification and regression tasks, this integration is lacking for more specialized tasks like survival or time-to-event models. As a result, survival analysis not only faces the general challenges of data preprocessing but also suffers from the lack of tailored, automated solutions in this area. To address this gap, this paper presents 'CleanSurvival', a reinforcement-learning-based solution for optimizing preprocessing pipelines, extended specifically for survival analysis. The framework can handle continuous and categorical variables, using Q-learning to select which combination of data imputation, outlier detection and feature extraction techniques achieves optimal performance for a Cox, random forest, neural network or user-supplied time-to-event model. The package is available on GitHub: https://github.com/datasciapps/CleanSurvival Experimental benchmarks on real-world datasets show that the Q-learning-based data preprocessing results in superior predictive performance to standard approaches, finding such a model up to 10 times faster than undirected random grid search. Furthermore, a simulation study demonstrates the effectiveness in different types and levels of missingness and noise in the data.
Authors: Obaidullah Zaland, Shutong Jin, Florian T. Pokorny, Monowar Bhuyan
Abstract: Federated learning (FL) enables collaborative learning without data centralization but introduces significant communication costs due to multiple communication rounds between clients and the server. One-shot federated learning (OSFL) addresses this by forming a global model with a single communication round, often relying on the server's model distillation or auxiliary dataset generation - mostly through pre-trained diffusion models (DMs). Existing DM-assisted OSFL methods, however, typically employ classifier-guided DMs, which require training auxiliary classifier models at each client, introducing additional computation overhead. This work introduces OSCAR (One-Shot Federated Learning with Classifier-Free Diffusion Models), a novel OSFL approach that eliminates the need for auxiliary models. OSCAR uses foundation models to devise category-specific data representations at each client which are integrated into a classifier-free diffusion model pipeline for server-side data generation. In our experiments, OSCAR outperforms the state-of-the-art on four benchmark datasets while reducing the communication load by at least 99%.
Authors: Shiqian Li, Ruihong Shen, Yaoyu Tao, Chi Zhang, Yixin Zhu
Abstract: Physical reasoning is a remarkable human ability that enables rapid learning and generalization from limited experience. Current AI models, despite extensive training, still struggle to achieve similar generalization, especially in Out-of-distribution (OOD) settings. This limitation stems from their inability to abstract core physical principles from observations. A key challenge is developing representations that can efficiently learn and generalize physical dynamics from minimal data. Here we present Neural Force Field (NFF), a framework extending Neural Ordinary Differential Equation (NODE) to learn complex object interactions through force field representations, which can be efficiently integrated through an Ordinary Differential Equation (ODE) solver to predict object trajectories. Unlike existing approaches that rely on discrete latent spaces, NFF captures fundamental physical concepts such as gravity, support, and collision in continuous explicit force fields. Experiments on three challenging physical reasoning tasks demonstrate that NFF, trained with only a few examples, achieves strong generalization to unseen scenarios. This physics-grounded representation enables efficient forward-backward planning and rapid adaptation through interactive refinement. Our work suggests that incorporating physics-inspired representations into learning systems can help bridge the gap between artificial and human physical reasoning capabilities.
Authors: Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert M\"uller, Stefan Gugler
Abstract: Most current neural networks for molecular dynamics (MD) include physical inductive biases, resulting in specialized and complex architectures. This is in contrast to most other machine learning domains, where specialist approaches are increasingly replaced by general-purpose architectures trained on vast datasets. In line with this trend, several recent studies have questioned the necessity of architectural features commonly found in MD models, such as built-in rotational equivariance or energy conservation. In this work, we contribute to the ongoing discussion by evaluating the performance of an MD model with as few specialized architectural features as possible. We present a recipe for MD using an Edge Transformer, an ``off-the-shelf'' transformer architecture that has been minimally modified for the MD domain, termed MD-ET. Our model implements neither built-in equivariance nor energy conservation. We use a simple supervised pre-training scheme on $\sim$30 million molecular structures from the QCML database. Using this ``off-the-shelf'' approach, we show state-of-the-art results on several benchmarks after fine-tuning for a small number of steps. Additionally, we examine the effects of being only approximately equivariant and energy conserving for MD simulations, proposing a novel method for distinguishing the errors resulting from non-equivariance from other sources of inaccuracies like numerical rounding errors. While our model exhibits runaway energy increases on larger structures, we show approximately energy-conserving NVE simulations for a range of small structures.
Authors: Yongwoo Kim, Sungmin Cha, Donghyun Kim
Abstract: Machine unlearning is a process to remove specific data points from a trained model while maintaining the performance on the retain data, addressing privacy or legal requirements. Despite its importance, existing unlearning evaluations tend to focus on logit-based metrics under small-scale scenarios. We observe that this could lead to a false sense of security in unlearning approaches under real-world scenarios. In this paper, we conduct a comprehensive evaluation that employs representation-based evaluations of the unlearned model under large-scale scenarios to verify whether the unlearning approaches truly eliminate the targeted data from the model's representation perspective. Our analysis reveals that current state-of-the-art unlearning approaches either completely degrade the representational quality of the unlearned model or merely modify the classifier, thereby achieving superior logit-based performance while maintaining representational similarity to the original model. Furthermore, we introduce a novel unlearning evaluation scenario in which the forgetting classes exhibit semantic similarity to downstream task classes, necessitating that feature representations diverge significantly from those of the original model, thus enabling a more thorough evaluation from a representation perspective. We hope our benchmark will serve as a standardized protocol for evaluating unlearning algorithms under realistic conditions.
Authors: Liheng Ma, Soumyasundar Pal, Yingxue Zhang, Philip H. S. Torr, Mark Coates
Abstract: Transformers have attained outstanding performance across various modalities, owing to their simple but powerful scaled-dot-product (SDP) attention mechanisms. Researchers have attempted to migrate Transformers to graph learning, but most advanced Graph Transformers (GTs) have strayed far from plain Transformers, exhibiting major architectural differences either by integrating message-passing or incorporating sophisticated attention mechanisms. These divergences hinder the easy adoption of training advances for Transformers developed in other domains. Contrary to previous GTs, this work demonstrates that the plain Transformer architecture can be a powerful graph learner. To achieve this, we propose to incorporate three simple, minimal, and easy-to-implement modifications to the plain Transformer architecture to construct our Powerful Plain Graph Transformers (PPGT): (1) simplified $L_2$ attention for measuring the magnitude closeness among tokens; (2) adaptive root-mean-square normalization to preserve token magnitude information; and (3) a simple MLP-based stem for graph positional encoding. Consistent with its theoretical expressivity, PPGT demonstrates noteworthy realized expressivity on the empirical graph expressivity benchmark, comparing favorably to more complicated alternatives such as subgraph GNNs and higher-order GNNs. Its empirical performance across various graph datasets also justifies the effectiveness of PPGT. This finding underscores the versatility of plain Transformer architectures and highlights their strong potential as a unified backbone for multimodal learning across language, vision, and graph domains.
Authors: Yunhui Liu, Jiashun Cheng, Yiqing Lin, Qizhuo Xie, Jia Li, Fugee Tsung, Hongzhi Yin, Tao Zheng, Jianhua Zhao, Tieke He
Abstract: Graph anomaly detection (GAD) has garnered increasing attention in recent years, yet remains challenging due to two key factors: (1) label scarcity stemming from the high cost of annotations and (2) homophily disparity at node and class levels. In this paper, we introduce Anomaly-Aware Pre-Training and Fine-Tuning (APF), a targeted and effective framework to mitigate the above challenges in GAD. In the pre-training stage, APF incorporates node-specific subgraphs selected via the Rayleigh Quotient, a label-free anomaly metric, into the learning objective to enhance anomaly awareness. It further introduces two learnable spectral polynomial filters to jointly learn dual representations that capture both general semantics and subtle anomaly cues. During fine-tuning, a gated fusion mechanism adaptively integrates pre-trained representations across nodes and dimensions, while an anomaly-aware regularization loss encourages abnormal nodes to preserve more anomaly-relevant information. Furthermore, we theoretically show that APF tends to achieve linear separability under mild conditions. Comprehensive experiments on 10 benchmark datasets validate the superior performance of APF in comparison to state-of-the-art baselines.
Authors: Xiao Zhang, Yaoyao Ding, Bolin Sun, Yang Hu, Tatiana Shpeisman, Gennady Pekhimenko
Abstract: Efficient GPU programming is crucial for achieving high performance in deep learning (DL) applications. The performance of GPU programs depends on how data is parallelized across threads and arranged within memory subsystems. The mapping functions describing tensors on GPUs are known as \emph{tensor layouts}. Low-level programming frameworks, such as CUTLASS and Hidet, provide expressive layout abstractions but often require \emph{considerable programming effort} to manually specify optimal layouts. High-level GPU programming languages, such as Triton, rely on compiler heuristics to generate dataflow, layouts, and pipelining strategies in GPU programs. However, the heuristics for dataflow and pipelining strategies are not generalizable to complex operators. To balance expressiveness and programmability, we propose Hexcute, a compiler framework that automates layout synthesis while providing explicit control over dataflow and pipelining. Hexcute formalizes layout synthesis as a constraint programming problem and solves it with a type-inference-based algorithm. This approach enables systematic exploration of optimal layouts and instructions. Our evaluation shows that Hexcute matches the performance of libraries like cuBLAS and FlashAttention on GEMM, Attention, and their variants, while reducing the amount of code by 1.27$\times$-7.94$\times$ compared to CUTLASS. For mixed-type mixture-of-experts (MoE) operators, Hexcute achieves an average speedup of 6.46$\times$ over Triton. In the end-to-end evaluations of vLLM, Hexcute delivers up to 2.60$\times$ speedup on DeepSeek-R1-AWQ and 2.04$\times$ on a Mamba-based model.
Authors: Andrew Millard, Joshua Murphy, Simon Maskell, Zheng Zhao
Abstract: Partial Bayesian neural networks (pBNNs) have been shown to perform competitively with fully Bayesian neural networks while only having a subset of the parameters be stochastic. Using sequential Monte Carlo (SMC) samplers as the inference method for pBNNs gives a non-parametric probabilistic estimation of the stochastic parameters, and has shown improved performance over parametric methods. In this paper we introduce a new SMC-based training method for pBNNs by utilising a guided proposal and incorporating gradient-based Markov kernels, which gives us better scalability on high dimensional problems. We show that our new method outperforms the state-of-the-art in terms of predictive performance and optimal loss. We also show that pBNNs scale well with larger batch sizes, resulting in significantly reduced training times and often better performance.
Authors: Haoze Song, Zhihao Li, Xiaobo Zhang, Zecheng Gan, Zhilu Lai, Wei Wang
Abstract: Neural Operators (NOs) have emerged as powerful tools for learning mappings between function spaces. Among them, the kernel integral operator has been widely used in universally approximating architectures. Following the original formulation, most advancements focus on designing better parameterizations for the kernel over the original physical domain (with $d$ spatial dimensions, $d\in{1,2,3,\ldots}$). In contrast, embedding evolution remains largely unexplored, which often drives models toward brute-force embedding lengthening to improve approximation, but at the cost of substantially increased computation. In this paper, we introduce an auxiliary dimension that explicitly models embedding evolution in operator form, thereby redefining the NO framework in $d+1$ dimensions (the original $d$ dimensions plus one auxiliary dimension). Under this formulation, we develop a Schr\"odingerised Kernel Neural Operator (SKNO), which leverages Fourier-based operators to model the $d+1$ dimensional evolution. Across more than ten increasingly challenging benchmarks, ranging from the 1D heat equation to the highly nonlinear 3D Rayleigh-Taylor instability, SKNO consistently outperforms other baselines. We further validate its resolution invariance under mixed-resolution training and super-resolution inference, and evaluate zero-shot generalization to unseen temporal regimes. In addition, we present a broader set of design choices for the lifting and recovery operators, demonstrating their impact on SKNO's predictive performance.
Authors: Yueqi Cao, Anthea Monod
Abstract: We introduce the first graph kernels for metric graphs via tropical algebraic geometry. In contrast to conventional graph kernels based on graph combinatorics such as nodes, edges, and subgraphs, our metric graph kernels are purely based on the geometry and topology of the underlying metric space. A key characterizing property of our construction is its invariance under edge subdivision, making the kernels intrinsically well-suited for comparing graphs representing different underlying metric spaces. We develop efficient algorithms to compute our kernels and analyze their complexity, which depends primarily on the genus of the input graphs rather than their size. Through experiments on synthetic data and selected real-world datasets, we demonstrate that our kernels capture complementary geometric and topological information overseen by standard combinatorial approaches, particularly in label-free settings. We further showcase their practical utility with an urban road network classification task.
Authors: Brian B. Moser, Arundhati S. Shanbhag, Stanislav Frolov, Federico Raue, Joachim Folz, Andreas Dengel
Abstract: Coreset selection targets the challenge of finding a small, representative subset of a large dataset that preserves essential patterns for effective machine learning. Although several surveys have examined data reduction strategies before, most focus narrowly on either classical geometry-based methods or active learning techniques. In contrast, this survey presents a more comprehensive view by unifying three major lines of coreset research, namely, training-free, training-oriented, and label-free approaches, into a single taxonomy. We present subfields often overlooked by existing work, including submodular formulations, bilevel optimization, and recent progress in pseudo-labeling for unlabeled datasets. Additionally, we examine how pruning strategies influence generalization and neural scaling laws, offering new insights that are absent from prior reviews. Finally, we compare these methods under varying computational, robustness, and performance demands and highlight open challenges, such as robustness, outlier filtering, and adapting coreset selection to foundation models, for future research.
Authors: Quan Zhou, Mark Kozdoba, Shie Mannor
Abstract: We study the problem of selecting a subset from a large action space shared by a family of bandits, with the goal of achieving performance nearly matching that of using the full action space. Indeed, in many natural situations, while the nominal set of actions may be large, there also exist significant correlations between the rewards of different actions. In this paper we propose an algorithm that can significantly reduce the action space when such correlations are present, without the need to a-priori know the correlation structure. We provide theoretical guarantees on the performance of the algorithm and demonstrate its practical effectiveness through empirical comparisons with Thompson Sampling and Upper Confidence Bound methods.
Authors: Yiqiao Liao, Farinaz Koushanfar, Parinaz Naghizadeh
Abstract: We introduce DyCO-GNN, a novel unsupervised learning framework for Dynamic Combinatorial Optimization that requires no training data beyond the problem instance itself. DyCO-GNN leverages structural similarities across time-evolving graph snapshots to accelerate optimization while maintaining solution quality. We evaluate DyCO-GNN on dynamic maximum cut, maximum independent set, and the traveling salesman problem across diverse datasets of varying sizes, demonstrating its superior performance under tight and moderate time budgets. DyCO-GNN consistently outperforms the baseline methods, achieving high-quality solutions up to 3-60x faster, highlighting its practical effectiveness in rapidly evolving resource-constrained settings.
Authors: Rui Cai, Bangzheng Li, Xiaofei Wen, Muhao Chen, Zhe Zhao
Abstract: Multimodal Large Language Models demonstrate strong performance on multimodal benchmarks, yet often exhibit poor robustness when exposed to spurious modality interference, such as irrelevant text in vision understanding, or irrelevant visual content in question answering. At its core, modality interference refers to cases where spurious signals from non-essential modalities distort model decisions, which we systematically analyze through causal, perturbation-based diagnostic experiments. To address this problem, we propose a unified finetuning framework that combines heuristic and adversarial perturbation-based data augmentation with output-level consistency regularization between original and perturbed inputs. Extensive experiments across image-heavy, text-heavy, and multimodal benchmarks, spanning multiple MLLM architectures and model scales, demonstrate consistent improvements in unimodal robustness and generalization, while improving standard multimodal performance.
Authors: Ziqi Liu, Pei Zeng, Yi Ding
Abstract: Real-world multichannel time series prediction faces growing demands for efficiency across edge and cloud environments, making channel compression a timely and essential problem. Motivated by the success of Multiple-Input Multiple-Output (MIMO) methods in signal processing, we propose a predictability-aware compression-decompression framework to reduce runtime, decrease communication cost, and maintain prediction accuracy across diverse predictors. The core idea involves using a circular seasonal key matrix with orthogonality to capture underlying time series predictability during compression and to mitigate reconstruction errors during decompression by introducing more realistic data assumptions. Theoretical analyses show that the proposed framework is both time-efficient and accuracy-preserving under a large number of channels. Extensive experiments on six datasets across various predictors demonstrate that the proposed method achieves superior overall performance by jointly considering prediction accuracy and runtime, while maintaining strong compatibility with diverse predictors.
Authors: Davide D'Ascenzo, Sebastiano Cultrera di Montesano
Abstract: Training deep learning models on single-cell datasets with hundreds of millions of cells requires loading data from disk, as these datasets exceed available memory. While random sampling provides the data diversity needed for effective training, it is prohibitively slow due to the random access pattern overhead, whereas sequential streaming achieves high throughput but introduces biases that degrade model performance. We present scDataset, a PyTorch data loader that enables efficient training from on-disk data with seamless integration across diverse storage formats. Our approach combines block sampling and batched fetching to achieve quasi-random sampling that balances I/O efficiency with minibatch diversity. On Tahoe-100M, a dataset of 100 million cells, scDataset achieves more than two orders of magnitude speedup compared to true random sampling while working directly with AnnData files. We provide theoretical bounds on minibatch diversity and empirically show that scDataset matches the performance of true random sampling across multiple classification tasks.
Authors: Antonio Almud\'evar, Jos\'e Miguel Hern\'andez-Lobato, Alfonso Ortega
Abstract: Deep learning representations are often difficult to interpret, which can hinder their deployment in sensitive applications. Concept Bottleneck Models (CBMs) have emerged as a promising approach to mitigate this issue by learning representations that support target task performance while ensuring that each component predicts a concrete concept from a predefined set. In this work, we argue that CBMs do not impose a true bottleneck: the fact that a component can predict a concept does not guarantee that it encodes only information about that concept. This shortcoming raises concerns regarding interpretability and the validity of intervention procedures. To overcome this limitation, we propose Minimal Concept Bottleneck Models (MCBMs), which incorporate an Information Bottleneck (IB) objective to constrain each representation component to retain only the information relevant to its corresponding concept. This IB is implemented via a variational regularization term added to the training loss. As a result, MCBMs yield more interpretable representations, support principled concept-level interventions, and remain consistent with probability-theoretic foundations.
Authors: Shaoang Li, Jian Li
Abstract: The rapid pace at which new large language models (LLMs) appear, and older ones become obsolete, forces providers to manage a streaming inventory under a strict concurrency cap and per-query cost budgets. We cast this as an online decision problem that couples stage-wise deployment (at fixed maintenance windows) with per-query routing among live models. We introduce StageRoute, a hierarchical algorithm that (i) optimistically selects up to $M_{\max}$ models for the next stage using reward upper-confidence and cost lower-confidence bounds, and (ii) routes each incoming query by solving a budget- and throughput-constrained bandit subproblem over the deployed set. We prove a regret of $\tilde{\mathcal{O}}(T^{2/3})$ with a matching lower bound, establishing near-optimality, and validate the theory empirically: StageRoute tracks a strong oracle under tight budgets across diverse workloads.
Authors: Zijian Ye, Wei Huang, Yifei Yu, Tianhe Ren, Zhongrui Wang, Xiaojuan Qi
Abstract: Large language models (LLMs) demonstrate remarkable performance but face substantial computational and memory challenges that limit their practical deployment. Quantization has emerged as a promising solution; however, its effectiveness is often limited by quantization errors arising from weight distributions that are not quantization-friendly and the presence of activation outliers. To address these challenges, we introduce DBellQuant, an innovative post-training quantization (PTQ) framework that achieves nearly 1-bit weight compression and 6-bit activation quantization with minimal performance degradation. DBellQuant uses Learnable Transformation for Dual-Bell (LTDB) algorithm, which transforms single-bell weight distributions into dual-bell forms to reduce binarization errors and applies inverse transformations to smooth activations. DBellQuant sets a new state-of-the-art by preserving superior model performance under aggressive weight and activation quantization. For example, on the Wikitext2 dataset, DBellQuant achieves a perplexity of 14.39 on LLaMA2-13B with 6-bit activation quantization, significantly outperforming BiLLM's 21.35 without activation quantization, underscoring its potential in compressing LLMs for real-world applications.
Authors: Luca Bindini, Lorenzo Perini, Stefano Nistri, Jesse Davis, Paolo Frasconi
Abstract: Contextual anomaly detection (CAD) aims to identify anomalies in a target (behavioral) variable conditioned on a set of contextual variables that influence the normalcy of the target variable but are not themselves indicators of anomaly. In many anomaly detection tasks, there exist contextual variables that influence the normalcy of the target variable but are not themselves indicators of anomaly. In this work, we propose a novel framework for CAD, normalcy score (NS), that explicitly models both the aleatoric and epistemic uncertainties. Built on heteroscedastic Gaussian process regression, our method regards the Z-score as a random variable, providing confidence intervals that reflect the reliability of the anomaly assessment. Through experiments on benchmark datasets and a real-world application in cardiology, we demonstrate that NS outperforms state-of-the-art CAD methods in both detection accuracy and interpretability. Moreover, confidence intervals enable an adaptive, uncertainty-driven decision-making process, which may be very important in domains such as healthcare.
Authors: Junda Wu, Yuxin Xiong, Xintong Li, Sheldon Yu, Zhengmian Hu, Tong Yu, Rui Wang, Xiang Chen, Jingbo Shang, Julian McAuley
Abstract: Chain-of-thought (CoT) reasoning enables large language models (LLMs) to break down complex problems into interpretable intermediate steps, significantly enhancing model transparency and performance in reasoning tasks. However, conventional CoT methods rely on heuristic sampling without structured modeling of reasoning transitions, constraining their ability to systematically explore and discover diverse and effective reasoning trajectories. In this work, we introduce CTRLS, a framework that formulates CoT reasoning as a Markov decision process (MDP) with latent state transitions, enabling principled and state-aware exploration via distributional reinforcement learning. By modelling reasoning actions as explicit probability distributions in latent space, our approach explicitly models epistemic uncertainty, facilitating robust exploration of the reasoning space. As part of our framework, we introduce an on-policy reinforcement learning strategy incorporating epsilon-greedy exploration and entropy-based regularization to iteratively refine latent state transitions without requiring additional fine-tuning of the underlying LLM. Theoretical analyses provide evidence lower bounds (ELBO), theoretically grounding our transition-aware modeling of latent reasoning dynamics. Further experiments demonstrate improvements in reasoning accuracy, diversity, and exploration efficiency across benchmark reasoning tasks.
Authors: Jeyoung Lee, Hochul Kang
Abstract: We propose the Signal Dice Similarity Coefficient (SDSC), a structure-aware metric function for time series self-supervised representation learning. Most Self-Supervised Learning (SSL) methods for signals commonly adopt distance-based objectives such as mean squared error (MSE), which are sensitive to amplitude, invariant to waveform polarity, and unbounded in scale. These properties hinder semantic alignment and reduce interpretability. SDSC addresses this by quantifying structural agreement between temporal signals based on the intersection of signed amplitudes, derived from the Dice Similarity Coefficient (DSC).Although SDSC is defined as a structure-aware metric, it can be used as a loss by subtracting from 1 and applying a differentiable approximation of the Heaviside function for gradient-based optimization. A hybrid loss formulation is also proposed to combine SDSC with MSE, improving stability and preserving amplitude where necessary. Experiments on forecasting and classification benchmarks demonstrate that SDSC-based pre-training achieves comparable or improved performance over MSE, particularly in in-domain and low-resource scenarios. The results suggest that structural fidelity in signal representations enhances the semantic representation quality, supporting the consideration of structure-aware metrics as viable alternatives to conventional distance-based methods.
Authors: Joanna Komorniczak
Abstract: The research community continues to seek increasingly more advanced synthetic data generators to reliably evaluate the strengths and limitations of machine learning methods. This work aims to increase the availability of datasets encompassing a diverse range of problem complexities by proposing a genetic algorithm that optimizes a set of problem complexity measures for classification and regression tasks towards specific targets. For classification, a set of 10 complexity measures was used, while for regression tasks, 4 measures demonstrating promising optimization capabilities were selected. Experiments confirmed that the proposed genetic algorithm can generate datasets with varying levels of difficulty by transforming synthetically created datasets to achieve target complexity values through linear feature projections. Evaluations involving state-of-the-art classifiers and regressors revealed a correlation between the complexity of the generated data and the recognition quality.
Authors: Yutong Liu, Cairong Zhao, Guosheng Hu
Abstract: For large language models (LLMs), post-training quantization (PTQ) can significantly reduce memory footprint and computational overhead. Model quantization is rapidly evolving. Though many papers report breakthrough results, they are often evaluated under different settings because a method typically contains multiple components. Analyzing connections among existing methods is important for deeper understanding. To bridge these gaps, we conduct an extensive review of state-of-the-art methods and perform comprehensive evaluations under the same conditions for fair comparison. To our knowledge, such a fair and extensive investigation remains critically underexplored. To better understand connections, first, we decouple published quantization methods into two steps: pre-quantization transformation and quantization error mitigation. The former is a preprocessing step that reduces outlier impact by flattening the data distribution; the latter offsets quantization errors to improve performance. Second, we evaluate and analyze the impact of different settings, including granularity and symmetry. Third, we analyze and evaluate the latest MXFP4 and NVFP4 data formats and their performance. Our experiments first demonstrate that optimized rotation and scaling yield the best pre-quantization performance, and that combining low-rank compensation with GPTQ can occasionally outperform GPTQ alone for error mitigation. Second, finer granularity improves performance but increases storage overhead. Third, we find that scaling-factor format and precision greatly affect FP4 performance, and that rotation-based strategies effective for INT4 offer limited gains for MXFP4 and NVFP4, motivating further study.
Authors: Ruiqi Wang, Diego Klabjan
Abstract: The concept of learning to optimize involves utilizing a trainable optimization strategy rather than relying on manually defined full gradient estimations such as ADAM. We present a framework that jointly trains the full gradient estimator and the trainable weights of the model. Specifically, we prove that pseudo-linear TO (Trainable Optimizer), a linear approximation of the full gradient, matches SGD's convergence rate while effectively reducing variance. Pseudo-linear TO incurs negligible computational overhead, requiring only minimal additional tensor multiplications. To further improve computational efficiency, we introduce two simplified variants of Pseudo-linear TO. Experiments demonstrate that TO methods converge faster than benchmark algorithms (e.g., ADAM) in both strongly convex and non-convex settings, and fine tuning of an LLM.
Authors: Fernando Martinez, Tao Li, Yingdong Lu, Juntao Chen
Abstract: Deep Q-learning jointly learns representations and values within monolithic networks, promising beneficial co-adaptation between features and value estimates. Although this architecture has attained substantial success, the coupling between representation and value learning creates instability as representations must constantly adapt to non-stationary value targets, while value estimates depend on these shifting representations. This is compounded by high variance in bootstrapped targets, which causes bias in value estimation in off-policy methods. We introduce Stackelberg Coupled Representation and Reinforcement Learning (SCORER), a framework for value-based RL that views representation and Q-learning as two strategic agents in a hierarchical game. SCORER models the Q-function as the leader, which commits to its strategy by updating less frequently, while the perception network (encoder) acts as the follower, adapting more frequently to learn representations that minimize Bellman error variance given the leader's committed strategy. Through this division of labor, the Q-function minimizes MSBE while perception minimizes its variance, thereby reducing bias accordingly, with asymmetric updates allowing stable co-adaptation, unlike simultaneous parameter updates in monolithic solutions. Our proposed SCORER framework leads to a bi-level optimization problem whose solution is approximated by a two-timescale algorithm that creates an asymmetric learning dynamic between the two players. Extensive experiments on DQN and its variants demonstrate that gains stem from algorithmic insight rather than model complexity.
Authors: Matteo Pinna, Andrea Ceni, Claudio Gallicchio
Abstract: We introduce a novel class of untrained Recurrent Neural Networks (RNNs) within the Reservoir Computing (RC) paradigm, called Residual Reservoir Memory Networks (ResRMNs). ResRMN combines a linear memory reservoir with a non-linear reservoir, where the latter is based on residual orthogonal connections along the temporal dimension for enhanced long-term propagation of the input. The resulting reservoir state dynamics are studied through the lens of linear stability analysis, and we investigate diverse configurations for the temporal residual connections. The proposed approach is empirically assessed on time-series and pixel-level 1-D classification tasks. Our experimental results highlight the advantages of the proposed approach over other conventional RC models.
Authors: Bryan E. Tuck, Rakesh M. Verma
Abstract: Adversarial text attacks remain a persistent threat to transformer models, yet existing defenses are typically attack-specific or require costly model retraining, leaving a gap for attack-agnostic detection. We introduce Guided Perturbation Sensitivity (GPS), a detection framework that identifies adversarial examples by measuring how embedding representations change when important words are masked. GPS first ranks words using importance heuristics, then measures embedding sensitivity to masking top-k critical words, and processes the resulting patterns with a BiLSTM detector. Experiments show that adversarially perturbed words exhibit disproportionately high masking sensitivity compared to naturally important words. Across three datasets, three attack types, and two victim models, GPS achieves over 85% detection accuracy and demonstrates competitive performance compared to existing state-of-the-art methods, often at lower computational cost. Using Normalized Discounted Cumulative Gain (NDCG) to measure perturbation identification quality, we demonstrate that gradient-based ranking significantly outperforms attention, hybrid, and random selection approaches, with identification quality strongly correlating with detection performance for word-level attacks ($\rho = 0.65$). GPS generalizes to unseen datasets, attacks, and models without retraining, providing a practical solution for adversarial text detection.
Authors: Matteo Pinna, Andrea Ceni, Claudio Gallicchio
Abstract: Echo State Networks (ESNs) are a particular type of untrained Recurrent Neural Networks (RNNs) within the Reservoir Computing (RC) framework, popular for their fast and efficient learning. However, traditional ESNs often struggle with long-term information processing. In this paper, we introduce a novel class of deep untrained RNNs based on temporal residual connections, called Deep Residual Echo State Networks (DeepResESNs). We show that leveraging a hierarchy of untrained residual recurrent layers significantly boosts memory capacity and long-term temporal modeling. For the temporal residual connections, we consider different orthogonal configurations, including randomly generated and fixed-structure configurations, and we study their effect on network dynamics. A thorough mathematical analysis outlines necessary and sufficient conditions to ensure stable dynamics within DeepResESN. Our experiments on a variety of time series tasks showcase the advantages of the proposed approach over traditional shallow and deep RC.
Authors: Kaoru Otsuka, Yuki Takezawa, Makoto Yamada
Abstract: Partial participation is essential for communication-efficient federated learning at scale, yet existing Byzantine-robust methods typically assume full client participation. In the partial participation setting, a majority of the sampled clients may be Byzantine, once Byzantine clients dominate, existing methods break down immediately. We introduce delayed momentum aggregation, a principle where the central server aggregates cached momentum from non-sampled clients along with fresh momentum from sampled clients. This principle ensures Byzantine clients remain a minority from the server's perspective even when they dominate the sampled set. We instantiate this principle in our optimizer DeMoA. We analyze the convergence rate of DeMoA, showing that DeMoA is Byzantine-robust under partial participation. Experiments show that, with 20% Byzantine ratio and only 10% partial participation rate, DeMoA achieves the best accuracy even when existing methods fail empirically.
Authors: Peter Racioppo
Abstract: We introduce Robust Filter Attention (RFA), an attention mechanism that reformulates self-attention as parallel robust filtering under a latent stochastic differential equation (SDE) prior, where analytically propagated uncertainty defines a time-dependent precision prior over attention weights. This formulation integrates key advantages of existing positional encodings: it preserves RoPE-style rotational structure while achieving long-context stability through explicit modeling of dissipation and diffusion. By imposing isotropic constraints on the dynamics and noise, RFA matches the $O(N^2 d)$ time and $O(N^2 + Nd)$ memory complexity of standard attention. Empirically, we find that uncertainty-aware weighting induces specialization into distinct filtering regimes across heads, improving temporal consistency and extrapolation across varying context lengths.
Authors: Ya Zhou, Yujie Yang, Xiaohan Fan, Wei Zhao
Abstract: ECG foundation models are increasingly popular due to their adaptability across various tasks. However, their clinical applicability is often limited by performance gaps compared to task-specific models, even after pre-training on large ECG datasets and fine-tuning on target data. This limitation is likely due to the lack of an effective post-training strategy. In this paper, we propose a simple yet effective post-training approach to enhance ECG foundation models. We evaluate it on a publicly available Transformer-based foundation model. Experiments across multiple ECG tasks show that our method consistently outperforms baseline fine-tuning. On the PTB-XL benchmarks, it improves macro AUROC by 0.7%-8.9% and macro AUPRC by 23.3%-77.9%, also outperforming several recent state-of-the-art approaches, including task-specific and advanced architectures. Further analyses demonstrate improved training dynamics and data efficiency, with only 30% of the training data outperforming the baseline trained on the full dataset. Ablation studies highlight the importance of stochastic depth and preview linear probing. These findings underscore the potential of post-training strategies to improve ECG foundation models, and we hope this work will contribute to the continued development of foundation models in the ECG domain.
Authors: Xinyu He, Xianghui Cao
Abstract: Post-training fundamentally alters the behavior of large language models (LLMs), yet its impact on the internal parameter space remains poorly understood. In this work, we conduct a systematic singular value decomposition (SVD) analysis of principal linear layers in pretrained LLMs, focusing on two widely adopted post-training methods: instruction tuning and long-chain-of-thought (Long-CoT) distillation. Our analysis reveals two unexpected and robust structural changes: (1) a near-uniform geometric scaling of singular values across layers; and (2) highly consistent orthogonal transformations are applied to the left and right singular vectors of each matrix. Based on these findings, We propose a simple yet effective framework to describe the coordinated dynamics of parameters in LLMs, which elucidates why post-training inherently relies on the foundational capabilities developed during pre-training. Further experiments demonstrate that singular value scaling underpins the temperature-controlled regulatory mechanisms of post-training, while the coordinated rotation of singular vectors encodes the essential semantic alignment. These results challenge the prevailing view of the parameter space in large models as a black box, uncovering the first clear regularities in how parameters evolve during training, and providing a new perspective for deeper investigation into model parameter changes.
Authors: Samuel Yoon, Jongwon Kim, Juyoung Ha, Young Myoung Ko
Abstract: Recently reconstruction-based deep models have been widely used for time series anomaly detection, but as their capacity and generalization capability increase, these models tend to over-generalize, often reconstructing unseen anomalies accurately. Prior works have attempted to mitigate this by incorporating a memory architecture that stores prototypes of normal patterns. Nevertheless, these approaches suffer from high training costs and have yet to be effectively integrated with time series foundation models (TFMs). To address these challenges, we propose MOMEMTO, an improved variant of TFM for anomaly detection, enhanced with a patch-based memory module to mitigate over-generalization. The memory module is designed to capture representative normal patterns from multiple domains and enables a single model to be jointly fine-tuned across multiple datasets through a multi-domain training strategy. MOMEMTO initializes memory items with latent representations from a pre-trained encoder, organizes them into patch-level units, and updates them via an attention mechanism. We evaluate our method using 23 univariate benchmark datasets. Experimental results demonstrate that MOMEMTO, as a single model, achieves higher scores on AUC and VUS metrics compared to baseline methods, and further enhances the performance of its backbone TFM, particularly in few-shot learning scenarios.
Authors: Alex Schutz, Victor-Alexandru Darvariu, Efimia Panagiotaki, Bruno Lacerda, Nick Hawes
Abstract: Neural Algorithmic Reasoning (NAR) is a paradigm that trains neural networks to execute classic algorithms by supervised learning. Despite its successes, important limitations remain: inability to construct valid solutions without post-processing and to reason about multiple correct ones, poor performance on combinatorial NP-hard problems, and inapplicability to problems for which strong algorithms are not yet known. To address these limitations, we reframe the problem of learning algorithm trajectories as a Markov Decision Process, which imposes structure on the solution construction procedure and unlocks the powerful tools of imitation and reinforcement learning (RL). We propose the GNARL framework, encompassing the methodology to translate problem formulations from NAR to RL and a learning architecture suitable for a wide range of graph-based problems. We achieve very high graph accuracy results on several CLRS-30 problems, performance matching or exceeding much narrower NAR approaches for NP-hard problems and, remarkably, applicability even when lacking an expert algorithm.
Authors: Dongming Huang, Zhifan Li, Yicheng Li, Qian Lin
Abstract: We study spectral algorithms in the setting where kernels are learned from data. We introduce the effective span dimension (ESD), an alignment-sensitive complexity measure that depends jointly on the signal, spectrum, and noise level $\sigma^2$. The ESD is well-defined for arbitrary kernels and signals without requiring eigen-decay conditions or source conditions. We prove that for sequence models whose ESD is at most $K$, the minimax excess risk scales as $\sigma^2 K$. Furthermore, we analyze over-parameterized gradient flow and prove that it can reduce the ESD. This finding establishes a connection between adaptive feature learning and provable improvements in generalization of spectral algorithms. We demonstrate the generality of the ESD framework by extending it to linear models and RKHS regression, and we support the theory with numerical experiments. This framework provides a novel perspective on generalization beyond traditional fixed-kernel theories.
Authors: Hideyuki Suzuki, Wataru Kurebayashi, Hiroshi Yamashita
Abstract: We propose a deterministic denoising algorithm for discrete-state diffusion models. The key idea is to derandomize the generative reverse Markov chain by introducing a variant of the herding algorithm, which induces deterministic state transitions driven by weakly chaotic dynamics. It serves as a direct replacement for the stochastic denoising process, without requiring retraining or continuous state embeddings. We demonstrate consistent improvements in both efficiency and sample quality on text and image generation tasks. In addition, the proposed algorithm yields improved solutions for diffusion-based combinatorial optimization. Thus, herding-based denoising is a simple yet promising approach for enhancing the generative process of discrete diffusion models. Furthermore, our results reveal that deterministic reverse processes, well established in continuous diffusion, can also be effective in discrete state spaces.
Authors: Yevgeny Seldin
Abstract: Learning, whether natural or artificial, is a process of selection. It starts with a set of candidate options and selects the more successful ones. In the case of machine learning the selection is done based on empirical estimates of prediction accuracy of candidate prediction rules on some data. Due to randomness of data sampling the empirical estimates are inherently noisy, leading to selection under uncertainty. The book provides statistical tools to obtain theoretical guarantees on the outcome of selection under uncertainty. We start with concentration of measure inequalities, which are the main statistical instrument for controlling how much an empirical estimate of expectation of a function deviates from the true expectation. The book covers a broad range of inequalities, including Markov's, Chebyshev's, Hoeffding's, Bernstein's, Empirical Bernstein's, Unexpected Bernstein's, kl, and split-kl. We then study the classical (offline) supervised learning and provide a range of tools for deriving generalization bounds, including Occam's razor, Vapnik-Chervonenkis analysis, and PAC-Bayesian analysis. The latter is further applied to derive generalization guarantees for weighted majority votes. After covering the offline setting, we turn our attention to online learning. We present the space of online learning problems characterized by environmental feedback, environmental resistance, and structural complexity. A common performance measure in online learning is regret, which compares performance of an algorithm to performance of the best prediction rule in hindsight, out of a restricted set of prediction rules. We present tools for deriving regret bounds in stochastic and adversarial environments, and under full information and bandit feedback.
Authors: Takashi Morita
Abstract: Vector quantization, which discretizes a continuous vector space into a finite set of representative vectors (a codebook), has been widely adopted in modern machine learning. Despite its effectiveness, vector quantization poses a fundamental challenge: the non-differentiable quantization step blocks gradient backpropagation. Smoothed vector quantization addresses this issue by relaxing the hard assignment of a codebook vector into a weighted combination of codebook entries, represented as the matrix product of a simplex vector and the codebook. Effective smoothing requires two properties: (1) smoothed quantizers should remain close to a onehot vector, ensuring tight approximation, and (2) all codebook entries should be utilized, preventing code collapse. Existing methods typically address these desiderata separately. By contrast, the present study introduces a simple and intuitive regularization that promotes both simultaneously by minimizing the distance between each simplex vertex and its $K$-nearest smoothed quantizers. Experiments on representative benchmarks, including discrete image autoencoding and contrastive speech representation learning, demonstrate that the proposed method achieves more reliable codebook utilization and improves performance compared to prior approaches.
Authors: Isaac Reid, Arijit Sehanobish, Cederik H\"ofs, Bruno Mlodozeniec, Leonhard Vulpius, Federico Barbero, Adrian Weller, Krzysztof Choromanski, Richard E. Turner, Petar Veli\v{c}kovi\'c
Abstract: We study the extent to which rotary position encodings (RoPE), a recent transformer position encoding algorithm broadly adopted in large language models (LLMs) and vision transformers (ViTs), can be applied to graph-structured data. We find that rotating tokens depending on the spectrum of the graph Laplacian efficiently injects structural information into the attention mechanism, boosting performance in synthetic and real-world graph learning tasks. This approach, coined _Wave-Induced Rotary Encodings_ (WIRE), enjoys intriguing theoretical properties: it recovers regular RoPE on grids, and depends asymptotically on the graph effective resistance. Unlike bias-based relative position encodings, WIRE is compatible with linear attention.
Authors: Guillem Capellera, Luis Ferraz, Antonio Rubio, Alexandre Alahi, Antonio Agudo
Abstract: Generative models often treat continuous data and discrete events as separate processes, creating a gap in modeling complex systems where they interact synchronously. To bridge this gap, we introduce JointDiff, a novel diffusion framework designed to unify these two processes by simultaneously generating continuous spatio-temporal data and synchronous discrete events. We demonstrate its efficacy in the sports domain by simultaneously modeling multi-agent trajectories and key possession events. This joint modeling is validated with non-controllable generation and two novel controllable generation scenarios: weak-possessor-guidance, which offers flexible semantic control over game dynamics through a simple list of intended ball possessors, and text-guidance, which enables fine-grained, language-driven generation. To enable the conditioning with these guidance signals, we introduce CrossGuid, an effective conditioning operation for multi-agent domains. We also share a new unified sports benchmark enhanced with textual descriptions for soccer and football datasets. JointDiff achieves state-of-the-art performance, demonstrating that joint modeling is crucial for building realistic and controllable generative models for interactive systems. https://guillem-cf.github.io/JointDiff/
Authors: Lorenz K. M\"uller, Philippe Bich, Jiawei Zhuang, Ahmet \c{C}elik, Luca Benfenati, Lukas Cavigelli
Abstract: Post-training quantization has emerged as the most widely used strategy for deploying large language models at low precision. Still, current methods show perplexity degradation at bit-widths less than or equal to 4, partly because representing outliers causes precision issues in parameters that share the same scales as these outliers. This problem is especially pronounced for calibration-free, uniform quantization methods. We introduce SINQ to augment existing post-training quantizers with an additional second-axis scale factor and a fast Sinkhorn-Knopp-style algorithm that finds scales to normalize per-row and per-column variances. We show that this approximates activation-aware quantization by recovering column scales from the weight matrix structure that are predictive of the typical activation magnitudes the matrix received during training. Our method has no interactions between layers and can be trivially applied to new architectures to quantize any linear layer. We evaluate our method on the Qwen3 model family, among others. SINQ reduces the perplexity gap on WikiText2 and C4 by over 50% against uncalibrated uniform quantization baselines, incurs zero to negligible compute overhead, and can be further enhanced by combining it with calibration and non-uniform quantization levels. Code is available at https://github.com/huawei-csl/SINQ.
Authors: Wenzhi Fang, Dong-Jun Han, Liangqi Yuan, Evan Chen, Christopher Brinton
Abstract: Device-cloud collaboration holds promise for deploying large language models (LLMs), leveraging lightweight on-device models for efficiency while relying on powerful cloud models for superior reasoning. A central challenge in this setting is determining, for each incoming query, whether it should be processed locally or offloaded to the cloud. Existing approaches typically rely on external routers, which often struggle to determine difficulty from the prompt itself, especially for tasks involving complex reasoning. Motivated by this limitation, we propose enabling on-device LLMs to decide internally whether to invoke cloud assistance at inference time, with this capability instilled through reinforcement learning based post-training. Casting on-device LLM post-training as a reward maximization problem, we design hierarchical rewards to encourage local problem solving and judicious cloud offloading. To solve the resulting problem, we develop an algorithm featuring a group-level policy gradient that stabilizes optimization, together with adaptive prompt filtering that provides complementary learning signals to mitigate policy collapse (i.e., exclusive local execution or exclusive cloud offloading). Extensive experiments on on-device-scale LLaMA and Qwen models across multiple reasoning benchmarks show that our method consistently outperforms baselines and significantly narrows the gap to full cloud LLMs.
Authors: Ran Elbaz, Guy Bar-Shalom, Yam Eitan, Fabrizio Frasca, Haggai Maron
Abstract: Permutation equivariant neural networks employing parameter-sharing schemes have emerged as powerful models for leveraging a wide range of data symmetries, significantly enhancing the generalization and computational efficiency of the resulting models. Recently, Kolmogorov-Arnold Networks (KANs) have demonstrated promise through their improved interpretability and expressivity compared to traditional architectures based on MLPs. While equivariant KANs have been explored in recent literature for a few specific data types, a principled framework for applying them to data with permutation symmetries in a general context remains absent. This paper introduces Function Sharing KAN (FS-KAN), a principled approach to constructing equivariant and invariant KA layers for arbitrary permutation symmetry groups, unifying and significantly extending previous work in this domain. We derive the basic construction of these FS-KAN layers by generalizing parameter-sharing schemes to the Kolmogorov-Arnold setup and provide a theoretical analysis demonstrating that FS-KANs have the same expressive power as networks that use standard parameter-sharing layers, allowing us to transfer well-known and important expressivity results from parameter-sharing networks to FS-KANs. Empirical evaluations on multiple data types and symmetry groups show that FS-KANs exhibit superior data efficiency compared to standard parameter-sharing layers, by a wide margin in certain cases, while preserving the interpretability and adaptability of KANs, making them an excellent architecture choice in low-data regimes.
Authors: Rongguang Ye, Ming Tang, Edith C. H. Ngai
Abstract: As increasingly large pre-trained models are released, deploying them on edge devices for privacy-preserving applications requires effective compression. Recent works combine quantization with the fine-tuning of high-precision LoRA adapters, which can substantially reduce model size while mitigating the accuracy loss from quantization. However, edge devices have inherently heterogeneous capabilities, while performing configuration-wise fine-tuning for every quantization setting is computationally prohibitive. In this paper, we propose CoA-LoRA, a method that dynamically adjusts the LoRA adapter to arbitrary quantization configurations (i.e., the per-layer bit-width choices of a pre-trained model) without requiring repeated fine-tuning. This is accomplished via a configuration-aware model that maps each configuration to its low-rank adjustments. The effectiveness of this model critically depends on the training configuration set, a collection of configurations chosen to cover different total bit-width budgets. However, constructing a high-quality configuration set is non-trivial. We therefore design a Pareto-based configuration search that iteratively optimizes the training configuration set, yielding more precise low-rank adjustments. Our experiments demonstrate that, unlike the state-of-the-art methods that require fine-tuning a separate LoRA adapter for each configuration, CoA-LoRA incurs no additional time cost while achieving comparable or even superior performance to those methods.
Authors: Olivier Goudet, Quentin Suire, Adrien Go\"effon, Fr\'ed\'eric Saubion, Sylvain Lamprier
Abstract: We introduce an order-invariant reinforcement learning framework for black-box combinatorial optimization. Classical estimation-of-distribution algorithms (EDAs) often rely on learning explicit variable dependency graphs, which can be costly and fail to capture complex interactions efficiently. In contrast, we parameterize a multivariate autoregressive generative model trained without a fixed variable ordering. By sampling random generation orders during training, a form of information-preserving dropout, the model is encouraged to be invariant to variable order, promoting search-space diversity, and shaping the model to focus on the most relevant variable dependencies, improving sample efficiency. We adapt Group Relative Policy Optimization (GRPO) to this setting, providing stable policy-gradient updates from scale-invariant advantages. Across a wide range of benchmark algorithms and problem instances of varying sizes, our method frequently achieves the best performance and consistently avoids catastrophic failures.
Authors: Zhaoyi Li, Jingtao Ding, Yong Li, Shihua Li
Abstract: Flow Matching (FM) models achieve remarkable results in generative tasks. Building upon diffusion models, FM's simulation-free training paradigm enables simplicity and efficiency but introduces a train-inference gap: model outputs cannot be assessed during training. Moreover, the straight flow assumption suffers from some inherent limitations. To address this, we propose to fine-tune FM via Maximum Likelihood Estimation (MLE) of reconstructions -- enabled by FM's smooth ODE formulation, unlike the stochastic differential equations (SDEs) in diffusion models. We first theoretically analyze the relationship between training loss and inference error in FM under numerical precision constraints. We then propose an easy-to-implement fine-tuning framework based on MLE of reconstructions, with flexibility for sophisticated extensions. Building on this, we incorporate a generalized artificial viscosity term that enhances flow stability and robustness, accompanied by a direct parameterization method and rigorous theoretical guarantees. Experiments demonstrate our method's effectiveness across diverse settings: a toy example provides mechanistic insights into the fine-tuning process, while large-scale evaluations on meteorological forecasting and robotic manipulation policies validate reliable performance improvements.
Authors: Mariona Jaramillo-Civill, Peng Wu, Pau Closas
Abstract: Clustered Federated Learning (CFL) improves performance under non-IID client heterogeneity by clustering clients and training one model per cluster, thereby balancing between a global model and fully personalized models. However, most CFL methods require the number of clusters K to be fixed a priori, which is impractical when the latent structure is unknown. We propose DPMM-CFL, a CFL algorithm that places a Dirichlet Process (DP) prior over the distribution of cluster parameters. This enables nonparametric Bayesian inference to jointly infer both the number of clusters and client assignments, while optimizing per-cluster federated objectives. This results in a method where, at each round, federated updates and cluster inferences are coupled, as presented in this paper. The algorithm is validated on benchmark datasets under Dirichlet and class-split non-IID partitions.
Authors: Florine Hartwig, Josua Sassen, Juliane Braunsmann, Martin Rumpf, Benedikt Wirth
Abstract: Latent manifolds of autoencoders provide low-dimensional representations of data, which can be studied from a geometric perspective. We propose to describe these latent manifolds as implicit submanifolds of some ambient latent space. Based on this, we develop tools for a discrete Riemannian calculus approximating classical geometric operators. These tools are robust against inaccuracies of the implicit representation often occurring in practical examples. To obtain a suitable implicit representation, we propose to learn an approximate projection onto the latent manifold by minimizing a denoising objective. This approach is independent of the underlying autoencoder and supports the use of different Riemannian geometries on the latent manifolds. The framework in particular enables the computation of geodesic paths connecting given end points and shooting geodesics via the Riemannian exponential maps on latent manifolds. We evaluate our approach on various autoencoders trained on synthetic and real data.
Authors: Xinyu Luo, Jie Liu, Kecheng Chen, Junyi Yang, Bo Ding, Arindam Basu, Haoliang Li
Abstract: Real-world deployment often exposes models to distribution shifts, making test-time adaptation (TTA) critical for robustness. Yet most TTA methods are unfriendly to edge deployment, as they rely on backpropagation, activation buffering, or test-time mini-batches, leading to high latency and memory overhead. We propose $\textbf{ELaTTA}$ ($\textit{Efficient Latent Test-Time Adaptation}$), a gradient-free framework for single-instance TTA under strict on-device constraints. ELaTTA freezes model weights and adapts each test sample by optimizing a low-dimensional coefficient vector in a source-induced principal latent subspace, pre-computed offline via truncated SVD and stored with negligible overhead. At inference, ELaTTA encourages prediction confidence by optimizing the $k$-D coefficients with CMA-ES, effectively optimizing a Gaussian-smoothed objective and improving stability near decision boundaries. Across six benchmarks and multiple architectures, ELaTTA achieves state-of-the-art accuracy under both strict and continual single-instance protocols, while reducing compute by up to $\textit{63$\times$}$ and peak memory by $\textit{11$\times$}$. We further demonstrate on-device deployment on a ZYNQ-7020 platform. Code will be released upon acceptance.
Authors: Jegwang Ryu, Minkyu Kim, Seungjun Shin, Hee Min Choi, Dokwan Oh, Jaeho Lee
Abstract: Efficient storage and transmission of language model weights are increasingly critical as model scale and deployment grow. Yet, most existing compression methods rely on handcrafted transforms and heuristics, reflecting the limited understanding of weights as a data modality. This motivates a shift toward learning-based paradigm, where compression schemes are optimized directly from data rather than manually designed. In this work, we take a step in this direction by formulating weight compression as a neural codec learning. We propose Neural Weight Compression (NWC), a flexible framework for training neural codecs on pretrained weight datasets. NWC addresses challenges intrinsic to weight compression, such as tensor shape heterogeneity and the misalignment between training losses and downstream performance, through components such as chunk-and-normalize preprocessing and an importance-aware training objective. Experiments show that NWC achieves state-of-the-art accuracy-compression tradeoffs, particularly at 4--6 bit regime, without relying on rigid handcrafted components such as the Hadamard transform. These gains extend across diverse architectures, e.g., vision encoders. Our analysis further supports the learning-based perspective, highlighting the roles of entropy-constrained quantization in high rate regime and learned transforms in adapting to downstream tasks.
Authors: Ayush Sawarni, Jikai Jin, Justin Whitehouse, Vasilis Syrgkanis
Abstract: Policy learning algorithms are widely used in areas such as personalized medicine and advertising to develop individualized treatment regimes. However, most methods force a decision even when predictions are uncertain, which is risky in high-stakes settings. We study policy learning with abstention, where a policy may defer to a safe default or an expert. When a policy abstains, it receives a small additive reward on top of the value of a random guess. We propose a two-stage learner that first identifies a set of near-optimal policies and then constructs an abstention rule from their disagreements. We establish fast O(1/n)-type regret guarantees when propensities are known, and extend these guarantees to the unknown-propensity case via a doubly robust (DR) objective. We further show that abstention is a versatile tool with direct applications to other core problems in policy learning: it yields improved guarantees under margin conditions without the common realizability assumption, connects to distributionally robust policy learning by hedging against small data shifts, and supports safe policy improvement by ensuring improvement over a baseline policy with high probability.
Authors: Noah Oberweis, Semih Cayci
Abstract: Continuous-time models provide important insights into the training dynamics of optimization algorithms in deep learning. In this work, we establish a non-asymptotic convergence analysis of stochastic gradient Langevin dynamics (SGLD), which is an It\^o stochastic differential equation (SDE) approximation of stochastic gradient descent in continuous time, in the lazy training regime. We show that, under regularity conditions on the Hessian of the loss function, SGLD with multiplicative and state-dependent noise (i) yields a non-degenerate kernel throughout the training process with high probability, and (ii) achieves exponential convergence to the empirical risk minimizer in expectation, and we establish finite-time and finite-width bounds on the optimality gap. We corroborate our theoretical findings with numerical examples in the regression setting.
Authors: Sukanya Patra, Souhaib Ben Taieb
Abstract: Unsupervised anomaly detection (AD) methods typically assume clean training data, yet real-world datasets often contain undetected or mislabeled anomalies, leading to significant performance degradation. Existing solutions require access to the training pipelines, data or prior knowledge of the proportions of anomalies in the data, limiting their real-world applicability. To address this challenge, we propose EPHAD, a simple yet effective test-time adaptation framework that updates the outputs of AD models trained on contaminated datasets using evidence gathered at test time. Our approach integrates the prior knowledge captured by the AD model trained on contaminated datasets with evidence derived from multimodal foundation models like Contrastive Language-Image Pre-training (CLIP), classical AD methods like the Local Outlier Factor or domain-specific knowledge. We illustrate the intuition behind EPHAD using a synthetic toy example and validate its effectiveness through comprehensive experiments across eight visual AD datasets, twenty-six tabular AD datasets, and a real-world industrial AD dataset. Additionally, we conduct an ablation study to analyse hyperparameter influence and robustness to varying contamination levels, demonstrating the versatility and robustness of EPHAD across diverse AD models and evidence pairs. To ensure reproducibility, our code is publicly available at https://github.com/sukanyapatra1997/EPHAD.
Authors: Edward Berman, Jacob Ginesin, Marco Pacini, Robin Walters
Abstract: Data-sparse settings such as robotic manipulation, molecular physics, and galaxy morphology classification are some of the hardest domains for deep learning. For these problems, equivariant networks can help improve modeling across undersampled parts of the input space, and uncertainty estimation can guard against overconfidence. However, until now, the relationships between equivariance and model confidence, and more generally equivariance and model calibration, has yet to be studied. Since traditional classification and regression error terms show up in the definitions of calibration error, it is natural to suspect that previous work can be used to help understand the relationship between equivariance and calibration error. In this work, we present a theory relating equivariance to uncertainty estimation. By proving lower and upper bounds on uncertainty calibration errors (ECE and ENCE) under various equivariance conditions, we elucidate the generalization limits of equivariant models and illustrate how symmetry mismatch can result in miscalibration in both classification and regression. We complement our theoretical framework with numerical experiments that clarify the relationship between equivariance and uncertainty using a variety of real and simulated datasets, and we comment on trends with symmetry mismatch, group size, and aleatoric and epistemic uncertainties.
Authors: Xiandong Zou, Jia Li, Xiaotong Yuan, Pan Zhou
Abstract: Normalization is fundamental to deep learning, but existing approaches such as BatchNorm, LayerNorm, and RMSNorm are variance-centric by enforcing zero mean and unit variance, stabilizing training without controlling how representations capture task-relevant information. We propose IB-Inspired Normalization (IBNorm), a simple yet powerful family of methods grounded in the Information Bottleneck principle. IBNorm introduces bounded compression operations that encourage embeddings to preserve predictive information while suppressing nuisance variability, yielding more informative representations while retaining the stability and compatibility of standard normalization. Theoretically, we prove that IBNorm achieves a higher IB value and tighter generalization bounds than variance-centric methods. Empirically, IBNorm consistently outperforms BatchNorm, LayerNorm, and RMSNorm across large-scale language models (LLaMA, GPT-2) and vision models (ResNet, ViT), with mutual information analysis confirming superior information bottleneck behavior. Code will be released publicly.
Authors: Kang Chen, Zhihao Liu, Tonghe Zhang, Zhen Guo, Si Xu, Hao Lin, Hongzhi Zang, Xiang Li, Quanlu Zhang, Zhaofei Yu, Guoliang Fan, Tiejun Huang, Yu Wang, Chao Yu
Abstract: Vision-Language-Action (VLA) models enable robots to understand and perform complex tasks from multimodal input. Although recent work explores using reinforcement learning (RL) to automate the laborious data collection process in scaling supervised fine-tuning (SFT), applying RL to large-scale flow-based VLAs (\eg, $\pi_0$, $\pi_{0.5}$) remains challenging due to intractable action log-likelihoods raised from flow matching. We address this challenge with $\pi_{\texttt{RL}}$, featuring two technical approaches: (1) \textbf{Flow-Noise} models the denoising process as a discrete-time MDP with a learnable noise network for exact log-likelihood computation. (2) \textbf{Flow-SDE} integrates denoising with agent-environment interaction, formulating a two-layer MDP that employs ODE-to-SDE conversion for efficient RL exploration. We evaluate $\pi_{\texttt{RL}}$ across various benchmarks, with experiments demonstrating that RL yields significant performance improvements in both in-distribution and out-of-distribution settings.
Authors: Frederik M\"obius Rygaard, Shen Zhu, Yinzhu Jin, S{\o}ren Hauberg, Tom Fletcher
Abstract: The geometry of generative models serves as the basis for interpolation, model inspection, and more. Unfortunately, most generative models lack a principal notion of geometry without restrictive assumptions on either the model or the data dimension. In this paper, we construct a general geometry compatible with different metrics and probability distributions to analyze generative models that do not require additional training. We consider curves analogous to geodesics constrained to a suitable data distribution aimed at targeting high-density regions learned by generative models. We formulate this as a (pseudo)-metric and prove that this corresponds to a Newtonian system on a Riemannian manifold. We show that shortest paths in our framework can be characterized by a system of ordinary differential equations, which locally corresponds to geodesics under a suitable Riemannian metric. Numerically, we derive a novel algorithm to efficiently compute shortest paths and generalized Fr\'echet means. Quantitatively, we show that curves using our metric traverse regions of higher density than baselines across a range of models and datasets.
Authors: Ahmet Erdem Pamuk, Emir Kaan \"Ozdemir, \c{S}uayp Talha Kocabay
Abstract: Large language models (LLMs) are increasingly trained with classical optimization techniques like AdamW to improve convergence and generalization. However, the mechanisms by which quantum-inspired methods enhance classical training remain underexplored. We introduce Superpositional Gradient Descent (SGD), a novel optimizer linking gradient updates with quantum superposition by injecting quantum circuit perturbations. We present a mathematical framework and implement hybrid quantum-classical circuits in PyTorch and Qiskit. On synthetic sequence classification and large-scale LLM fine-tuning, SGD converges faster and yields lower final loss than AdamW. Despite promising results, scalability and hardware constraints limit adoption. Overall, this work provides new insights into the intersection of quantum computing and deep learning, suggesting practical pathways for leveraging quantum principles to control and enhance model behavior.
Authors: Martin Carrasco, Olga Zaghen, Kavir Sumaraj, Erik Bekkers, Bastian Rieck
Abstract: A large driver of the complexity of graph learning is the interplay between \emph{structure} and \emph{features}.When analyzing the expressivity of graph neural networks, however, existing approaches ignore features in favor of structure, making it nigh-impossible to assess to what extent two graphs with close features should be considered similar.We address this by developing a new \mbox{(pseudo-)metric} based on graph homomorphisms.Inspired by concepts from metric geometry, our \emph{graph homomorphism distortion} measures the minimal worst-case distortion that node features of one graph are subjected to when mapping one graph to another.We demonstrate the utility of our novel measure by showing that (i.) it can be efficiently calculated under some additional assumptions, (ii.) it complements existing expressivity measures like \mbox{$1$-WL}, and (iii.)it permits defining structural encodings, which improve the predictive capabilities of graph neural networks.
Authors: Tim Tomov, Dominik Fuchsgruber, Tom Wollschl\"ager, Stephan G\"unnemann
Abstract: Accurate uncertainty quantification (UQ) in Large Language Models (LLMs) is critical for trustworthy deployment. While real-world language is inherently ambiguous, reflecting aleatoric uncertainty, existing UQ methods are typically benchmarked against tasks with no ambiguity. In this work, we demonstrate that while current uncertainty estimators perform well under the restrictive assumption of no ambiguity, they degrade to close-to-random performance on ambiguous data. To this end, we introduce MAQA* and AmbigQA*, the first ambiguous question-answering (QA) datasets equipped with ground-truth answer distributions estimated from factual co-occurrence. We find this performance deterioration to be consistent across different estimation paradigms: using the predictive distribution itself, internal representations throughout the model, and an ensemble of models. We show that this phenomenon can be theoretically explained, revealing that predictive-distribution and ensemble-based estimators are fundamentally limited under ambiguity. Overall, our study reveals a key shortcoming of current UQ methods for LLMs and motivates a rethinking of current modeling paradigms.
Authors: Yong-Woon Kim, Chulung Kang, Yung-Cheol Byun
Abstract: Green hydrogen production via polymer electrolyte membrane (PEM) water electrolysis is pivotal for energy transition, yet hydrogen crossover through membranes threatens safety and economic viability-approaching explosive limits (4 mol% H$_2$ in O$_2$) while reducing Faradaic efficiency and accelerating membrane degradation. Current physics-based models require extensive calibration and computational resources that preclude real-time implementation, while purely data-driven approaches fail to extrapolate beyond training conditions-critical for dynamic electrolyzer operation. Here we present the first application of physics-informed neural networks (PINNs) for hydrogen crossover prediction, trained on 184 published measurements augmented to 1,114 points and constrained by a constitutive physics model (Henry's law, Fick's diffusion, and Faraday-based gas production) embedded in the loss function. Our compact architecture (17,793 parameters), validated across six membranes under industrially relevant conditions (0.05-5.0 A/cm$^2$, 1-200 bar, 25-85{\deg}C), achieves exceptional accuracy (R$^2$ = 99.84% $\pm$ 0.15%, RMSE = 0.0932% $\pm$ 0.0438%) based on five-fold cross-validation, with sub-millisecond inference enabling real-time control. Remarkably, the model maintains R$^2$ > 86% when predicting crossover at pressures 2.5x beyond training range-substantially outperforming pure neural networks (R$^2$ = 43.4%). The hardware-agnostic deployment, from desktop CPUs to edge devices (Raspberry Pi 4), enables distributed safety monitoring essential for gigawatt-scale installations. By bridging physical rigor and computational efficiency, this work establishes a new paradigm for real-time electrolyzer monitoring, accelerating deployment of safe, efficient green hydrogen infrastructure crucial for net-zero emissions targets.
Authors: Samyak Sanghvi, Nishant Ranjan, Tarak Karmakar
Abstract: Designing ligands that are both chemically valid and structurally compatible with protein binding pockets is a key bottleneck in computational drug discovery. Existing approaches either ignore structural context or rely on expensive, memory-intensive encoding that limits throughput and scalability. We present SiDGen (Structure-informed Diffusion Generator), a protein-conditioned diffusion framework that integrates masked SMILES generation with lightweight folding-derived features for pocket awareness. To balance expressivity with efficiency, SiDGen supports two conditioning pathways: a streamlined mode that pools coarse structural signals from protein embeddings and a full mode that injects localized pairwise biases for stronger coupling. A coarse-stride folding mechanism with nearest-neighbor upsampling alleviates the quadratic memory costs of pair tensors, enabling training on realistic sequence lengths. Learning stability is maintained through in-loop chemical validity checks and an invalidity penalty, while large-scale training efficiency is restored \textit{via} selective compilation, dataloader tuning, and gradient accumulation. In automated benchmarks, SiDGen generates ligands with high validity, uniqueness, and novelty, while achieving competitive performance in docking-based evaluations and maintaining reasonable molecular properties. These results demonstrate that SiDGen can deliver scalable, pocket-aware molecular design, providing a practical route to conditional generation for high-throughput drug discovery.
Authors: Kazi Ahmed Asif Fuad, Lizhong Chen
Abstract: Kolmogorov Arnold Networks (KANs) represent a new class of neural architectures that replace conventional linear transformations and node-based nonlinearities with spline-based function approximations distributed along network edges. Although KANs offer strong expressivity and interpretability, their heterogeneous spline and base branch parameters hinder efficient quantization, which remains unexamined compared to CNNs and Transformers. In this paper, we present QuantKAN, a unified framework for quantizing KANs across both quantization aware training (QAT) and post-training quantization (PTQ) regimes. QuantKAN extends modern quantization algorithms, such as LSQ, LSQ+, PACT, DoReFa, QIL, GPTQ, BRECQ, AdaRound, AWQ, and HAWQ-V2, to spline based layers with branch-specific quantizers for base, spline, and activation components. Through extensive experiments on MNIST, CIFAR 10, and CIFAR 100 across multiple KAN variants (EfficientKAN, FastKAN, PyKAN, and KAGN), we establish the first systematic benchmarks for low-bit spline networks. Our results show that KANs, particularly deeper KAGN variants, are compatible with low-bit quantization but exhibit strong method architecture interactions: LSQ, LSQ+, and PACT preserve near full precision accuracy at 4 bit for shallow KAN MLP and ConvNet models, while DoReFa provides the most stable behavior for deeper KAGN under aggressive low-bit settings. For PTQ, GPTQ and Uniform consistently deliver the strongest overall performance across datasets, with BRECQ highly competitive on simpler regimes such as MNIST. Our proposed QuantKAN framework thus unifies spline learning and quantization, and provides practical tools and guidelines for efficiently deploying KANs in real-world, resource-constrained environments.
Authors: Sanjeev Manivannan (Indian Institute of Technology Madras, Chennai, India), Chandrashekar Lakshminarayan (Indian Institute of Technology Madras, Chennai, India)
Abstract: Cross-subject motor-imagery decoding remains a major challenge in EEG-based brain-computer interfaces. To mitigate strong inter-subject variability, recent work has emphasized manifold-based approaches operating on covariance representations. Yet dispersion scaling and orientation alignment remain largely unaddressed in existing methods. In this paper, we address both issues through congruence transforms and introduce three complementary geometry-aware models: (i) Discriminative Congruence Transform (DCT), (ii) Deep Linear DCT (DLDCT), and (iii) Deep DCT-UNet (DDCT-UNet). These models are evaluated both as pre-alignment modules for downstream classifiers and as end-to-end discriminative systems trained via cross-entropy backpropagation with a custom logistic-regression head. Across challenging motor-imagery benchmarks, the proposed framework improves transductive cross-subject accuracy by 2-3%, demonstrating the value of geometry-aware congruence learning.
Authors: Ihab Ahmed, Denis Krompa{\ss}, Cheng Feng, Volker Tresp
Abstract: We investigate input normalization methods for Time-Series Foundation Models (TSFMs). While normalization is well-studied in dataset-specific time-series models, it remains overlooked in TSFMs where generalization is critical. Time-series data, unlike text or images, exhibits significant scale variation across domains and channels, coupled with non-stationarity, can undermine TSFM performance regardless of architectural complexity. Through systematic evaluation across four architecturally diverse TSFMs, we empirically establish REVIN as the most efficient approach, reducing zero-shot MASE by 89\% relative to an un-normalized baseline and by 44\% versus other normalization methods, while matching the best in-domain accuracy (0.84 MASE) without any dataset-level preprocessing -- yielding the highest accuracy-efficiency trade-off. Yet its effect utilization depends on architectural design choices and optimization objective, particularly with respect to training loss scale sensitivity and model type (probabilistic, point-forecast, or LLM-based models).
Authors: Feiyu Wang, Xinyu Tan, Bokai Huang, Yihao Zhang, Guoan Wang, Peizhuang Cong, Tong Yang
Abstract: Large language models (LLMs) have revolutionized artificial intelligence, yet their massive memory and computational demands necessitate aggressive quantization, increasingly pushing representations toward the theoretical limit of a single bit. While complex-valued LLMs, such as iFairy, offer a superior chance for low-bit representation compared to real-valued counterparts, they require training from scratch, preventing the utilization of the vast ecosystem of pre-trained real-valued foundation models. Here we present Fairy2i, a universal framework that transforms pre-trained real-valued layers into an equivalent widely-linear complex form, enabling extremely low-bit quantization while reusing existing checkpoints. By proving a lossless mathematical equivalence between real and widely-linear maps, we convert standard Transformers into the complex domain and employ a phase-aware quantization scheme with a highly efficient codebook of fourth roots of unity. Furthermore, we introduce a recursive residual quantization mechanism that iteratively minimizes quantization error, allowing inference to proceed via efficient multiplication-free accumulation. We demonstrate that Fairy2i restores the performance of LLaMA-2 7B at an effective 2-bit precision to levels nearly comparable with full-precision baselines, significantly outperforming state-of-the-art real-valued binary and ternary quantization methods. This work bridges the gap between the representational efficiency of complex-valued arithmetic and the practical utility of pre-trained models, paving a new way for efficient inference on commodity hardware. We open-source the Fairy2i model and code at https://huggingface.co/PKU-DS-LAB/Fairy2i-W2 and https://github.com/PKULab1806/Fairy2i-W2.
URLs: https://huggingface.co/PKU-DS-LAB/Fairy2i-W2, https://github.com/PKULab1806/Fairy2i-W2.
Authors: Zhiming Xu, Baile Xu, Jian Zhao, Furao Shen, Suorong Yang
Abstract: Existing replay and distillation-based class-incremental learning (CIL) methods are effective at retaining past knowledge but are still constrained by the stability-plasticity dilemma. Since their resulting models are learned over a sequence of incremental tasks, they encode rich representations and can be regarded as pre-trained bases. Building on this view, we propose a plug-in extension paradigm termed Deployment of LoRA Components (DLC) to enhance them. For each task, we use Low-Rank Adaptation (LoRA) to inject task-specific residuals into the base model's deep layers. During inference, representations with task-specific residuals are aggregated to produce classification predictions. To mitigate interference from non-target LoRA plugins, we introduce a lightweight weighting unit. This unit learns to assign importance scores to different LoRA-tuned representations. Like downloadable content in software, DLC serves as a plug-and-play enhancement that efficiently extends the base methods. Remarkably, on the large-scale ImageNet-100, with merely 4\% of the parameters of a standard ResNet-18, our DLC model achieves a significant 8\% improvement in accuracy, demonstrating exceptional efficiency. Under a fixed memory budget, methods equipped with DLC surpass state-of-the-art expansion-based methods.
Authors: Stefan Nielsen, Edoardo Cetin, Peter Schwendeman, Qi Sun, Jinglue Xu, Yujin Tang
Abstract: Powerful large language models (LLMs) from different providers have been expensively trained and finetuned to specialize across varying domains. In this work, we introduce a new kind of Conductor model trained with reinforcement learning to automatically discover powerful coordination strategies among LLMs. Our Conductor learns not only to design targeted communication topologies for effective agent-to-agent collaboration, but also to prompt engineer focused instructions to the LLMs to maximally leverage their individual capabilities. We show that, by learning optimal coordination strategies over pools of powerful worker LLMs, a 7B Conductor achieves significant performance gains beyond any individual worker, attaining state-of-the-art results in challenging reasoning benchmarks, such as LiveCodeBench and GPQA. By training with randomized agent pools, our conductor effectively adapts to arbitrary sets of open- and closed-source agents, meeting any user requirements. Furthermore, allowing the Conductor to select itself as a worker gives rise to recursive topologies, elevating performance with a new form of dynamic test-time scaling through online iterative adaptation. More broadly, ours is among the early work demonstrating language model coordination can be unlocked through RL, where powerful coordination strategies emerge naturally in LLMs through pure end-to-end reward maximization.
Authors: Yiming Tang, Harshvardhan Saini, Zhaoqian Yao, Zheng Lin, Yizhen Liao, Qianxiao Li, Mengnan Du, Dianbo Liu
Abstract: As AI models achieve remarkable capabilities across diverse domains, understanding what representations they learn and how they encode concepts has become increasingly important for both scientific progress and trustworthy deployment. Recent works in mechanistic interpretability have widely reported that neural networks represent meaningful concepts as linear directions in their representation spaces and often encode diverse concepts in superposition. Various sparse dictionary learning (SDL) methods, including sparse autoencoders, transcoders, and crosscoders, are utilized to address this by training auxiliary models with sparsity constraints to disentangle these superposed concepts into monosemantic features. These methods are the backbone of modern mechanistic interpretability, yet in practice they consistently produce polysemantic features, feature absorption, and dead neurons, with very limited theoretical understanding of why these phenomena occur. Existing theoretical work is limited to tied-weight sparse autoencoders, leaving the broader family of SDL methods without formal grounding. We develop the first unified theoretical framework that casts all major SDL variants as a single piecewise biconvex optimization problem, and characterize its global solution set, non-identifiability, and spurious optima. This analysis yields principled explanations for feature absorption and dead neurons. To expose these pathologies under full ground-truth access, we introduce the Linear Representation Bench. Guided by our theory, we propose feature anchoring, a novel technique that restores SDL identifiability, substantially improving feature recovery across synthetic benchmarks and real neural representations.
Authors: Daniel Rose, Roxane Axel Jacob, Johannes Kirchmair, Thierry Langer
Abstract: Transformer-based autoregressive models offer a promising alternative to diffusion- and flow-matching approaches for generating 3D molecular structures. However, standard transformer architectures require a sequential ordering of tokens, which is not uniquely defined for the atoms in a molecule. Prior work has addressed this by using canonical atom orderings, but these do not ensure permutation invariance of atoms, which is essential for tasks like prefix completion. We introduce NEAT, a Neighborhood-guided, Efficient, Autoregressive, Set Transformer that treats molecular graphs as sets of atoms and learns an order-agnostic distribution over admissible tokens at the graph boundary. NEAT achieves state-of-the-art performance in autoregressive 3D molecular generation whilst ensuring atom-level permutation invariance by design.
Authors: Jaerin Lee, Kyoung Mu Lee
Abstract: We lay the theoretical foundation for automating optimizer design in gradient-based learning. Based on the greedy principle, we formulate the problem of designing optimizers and their hyperparameters as maximizing the instantaneous decrease in loss. By treating an optimizer as a function that translates loss gradient signals into parameter motions, the problem reduces to a family of convex optimization problems over the space of optimizers. Solving these problems under various constraints not only recovers a wide range of popular optimizers as closed-form solutions, but also produces the optimal hyperparameters of these optimizers with respect to the problems at hand. This enables a systematic approach to design optimizers and tune their hyperparameters dynamically according to the gradient statistics that are collected during the training process. Furthermore, this optimization of optimization can be performed dynamically during training.
Authors: Aoxiang Ma, Salah Ghamizi, Jun Cao, Pedro Rodriguez
Abstract: Battery energy storage systems (BESS) have become increasingly vital in three-phase unbalanced distribution grids for maintaining voltage stability and enabling optimal dispatch. However, existing deep learning approaches often lack explicit three-phase representation, making it difficult to accurately model phase-specific dynamics and enforce operational constraints--leading to infeasible dispatch solutions. This paper demonstrates that by embedding detailed three-phase grid information--including phase voltages, unbalanced loads, and BESS states--into heterogeneous graph nodes, diverse GNN architectures (GCN, GAT, GraphSAGE, GPS) can jointly predict network state variables with high accuracy. Moreover, a physics-informed loss function incorporates critical battery constraints--SoC and C-rate limits--via soft penalties during training. Experimental validation on the CIGRE 18-bus distribution system shows that this embedding-loss approach achieves low prediction errors, with bus voltage MSEs of 6.92e-07 (GCN), 1.21e-06 (GAT), 3.29e-05 (GPS), and 9.04e-07 (SAGE). Importantly, the physics-informed method ensures nearly zero SoC and C-rate constraint violations, confirming its effectiveness for reliable, constraint-compliant dispatch.
Authors: Shuibai Zhang, Fred Zhangzhi Peng, Yiheng Zhang, Jin Pan, Grigorios G. Chrysos
Abstract: While Diffusion Language Models (DLMs) are theoretically well-suited for iterative refinement due to their non-causal structure, they often fail to reliably revise incorrect tokens in practice. The key challenge lies in the model's inability to distinguish between correct and erroneous tokens in a visible sequence. Standard masked diffusion language model (MDLM) training is restricted to the objective of unmasking, undermining the effectiveness of refinement guided by confidence. Based on this observation, we study corrective behavior in DLMs, defined as the ability to assign lower confidence to incorrect tokens and iteratively refine them while preserving correct content. We show that this capability is not induced by conventional masked diffusion objectives and propose a post-training principle oriented by correction that explicitly supervises visible incorrect tokens, enabling discriminative confidence and targeted refinement. To evaluate corrective behavior, we introduce the Code Revision Benchmark, a controllable and executable benchmark for assessing error localization and in-place correction. Experiments on code revision tasks and parallel decoding scenarios demonstrate that models trained with our approach substantially outperform standard MDLMs, with gains that are most pronounced when parallel decoding introduces substantial uncertainty and iterative refinement becomes essential. Our code is publicly available at https://github.com/zhangshuibai/CDLM.
Authors: Mathieu Blondel, Michael E. Sander, Germain Vivier-Ardisson, Tianlin Liu, Vincent Roulet
Abstract: Autoregressive models (ARMs) currently constitute the dominant paradigm for large language models (LLMs). Energy-based models (EBMs) represent another class of models, which have historically been less prevalent in LLM development, yet naturally characterize the optimal policy in post-training alignment. In this paper, we provide a unified view of these two model classes. Taking the chain rule of probability as a starting point, we establish an explicit bijection between ARMs and EBMs in function space, which we show to correspond to a special case of the soft Bellman equation in maximum entropy reinforcement learning. Building upon this bijection, we derive the equivalence between supervised learning of ARMs and EBMs. Furthermore, we analyze the distillation of EBMs into ARMs by providing theoretical error bounds. Our results provide insights into the ability of ARMs to plan ahead, despite being based on the next-token prediction paradigm.
Authors: Yuanchao Xu, Fengyi Li, Masahiro Fujisawa, Xiaoyuan Cheng, Youssef Marzouk, Isao Ishikawa
Abstract: We propose Koopman Spectral Wasserstein Gradient Descent (KSWGD), a particle-based generative modeling framework that learns the Langevin generator via Koopman theory and integrates it with Wasserstein gradient descent. Our key insight is that this spectral structure of the underlying distribution can be directly estimated from trajectory data via the Koopman operator, eliminating the need for explicit knowledge of the target potential. Additionally, we prove that KSWGD maintains an approximately constant dissipation rate, thereby establishing linear convergence and overcoming the vanishing-gradient phenomenon that hinders existing kernel-based particle methods. We further provide a Feynman--Kac interpretation that clarifies the method's probabilistic foundation. Experiments on compact manifolds, metastable multi-well systems, and high-dimensional stochastic partial differential equations demonstrate that KSWGD consistently outperforms baselines in both convergence speed and sample quality.
Authors: Jong-Ik Park, Shreyas Chaudhari, Srinivasa Pranav, Carlee Joe-Wong, Jos\'e M. F. Moura
Abstract: In many deployed systems (multilingual ASR, cross-hospital imaging, region-specific perception), multiple pretrained specialist models coexist. Yet, new target domains often require domain expansion: a generalized model that performs well beyond any single specialist's domain. Given a new target domain, existing methods obtain a single strong initialization prior for the model parameters by blending expert models to initialize a target model. However, heuristic blending -- using mixing coefficients based on data size or proxy metrics -- often yields lower target-domain test accuracy, and learning these coefficients on the target domain's loss function typically requires computationally-expensive full backpropagation through a neural network. We propose GLUE, Gradient-free Learning to Unify Experts, which initializes the target model as a convex combination of fixed experts and learns the mixture coefficients of this combination via gradient-free two-point SPSA (simultaneous perturbation stochastic approximation) updates, requiring only two forward passes per step. Across experiments on three datasets and three network architectures, GLUE produces model parameter priors that can be fine-tuned to outperform baselines. GLUE improves test accuracy by up to 8.5% over data-size weighting and by up to 9.1% over proxy-metric selection. GLUE either outperforms backpropagation-based full-gradient mixing or matches its performance within 1.4%.
Authors: Maryam Mirzaei, Farzaneh Shayegh, Hamed Narimani
Abstract: Accurate recognition of human emotional states is critical for effective human-machine interaction. Electroencephalography (EEG) offers a reliable source for emotion recognition due to its high temporal resolution and its direct reflection of neural activity. Nevertheless, variations across recording sessions present a major challenge for model generalization. To address this issue, we propose EGDA, a framework that reduces cross-session discrepancies by jointly aligning the global (marginal) and class-specific (conditional) distributions, while preserving the intrinsic structure of EEG data through graph regularization. Experimental results on the SEED-IV dataset demonstrate that EGDA achieves robust cross-session performance, obtaining accuracies of 81.22%, 80.15%, and 83.27% across three transfer tasks, and surpassing several baseline methods. Furthermore, the analysis highlights the Gamma frequency band as the most discriminative and identifies the central-parietal and prefrontal brain regions as critical for reliable emotion recognition.
Authors: Xiaoze Liu, Weichen Yu, Matt Fredrikson, Xiaoqian Wang, Jing Gao
Abstract: The open-weight language model ecosystem is increasingly defined by model composition techniques (such as weight merging, speculative decoding, and vocabulary expansion) that remix capabilities from diverse sources. A critical prerequisite for applying these methods across different model families is tokenizer transplant, which aligns incompatible vocabularies to a shared embedding space. We demonstrate that this essential interoperability step introduces a supply-chain vulnerability: we engineer a single breaker token that is functionally inert in a donor model yet reliably reconstructs into a high-salience malicious feature after transplant into a base model. By exploiting the geometry of coefficient reuse, our attack sabotages the base model's generation while leaving the donor's utility statistically indistinguishable from nominal behavior. We formalize this as a dual-objective optimization problem and instantiate the attack using a sparse solver. Empirically, the attack is training-free and evades outlier detection, while demonstrating structural persistence against fine-tuning and weight merging, highlighting a hidden risk in the pipeline of modular AI composition. Code is available at https://github.com/xz-liu/tokenforge
Authors: Dai Shi, Lequan Lin, Andi Han, Luke Thompson, Jos\'e Miguel Hern\'andez-Lobato, Zhiyong Wang, Junbin Gao
Abstract: Stochastic differential equations (SDEs) and stochastic partial differential equations (SPDEs) are fundamental for modeling stochastic dynamics across the natural sciences and modern machine learning. Learning their solution operators with deep learning models promises fast solvers and new perspectives on classical learning tasks. In this work, we build on Wiener-chaos expansions (WCE) to design neural operator (NO) architectures for SDEs and SPDEs: we project driving noise paths onto orthonormal Wick-Hermite features and use NOs to parameterize the resulting chaos coefficients, enabling reconstruction of full trajectories from noise in a single forward pass. We also make the underlying WCE structure explicit for multi-dimensional SDEs and semilinear SPDEs by showing the coupled deterministic ODE/PDE systems governing these coefficients. Empirically, we achieve competitive accuracy across several tasks, including standard SPDE benchmarks and SDE-based diffusion one-step image sampling, topological graph interpolation, financial extrapolation, parameter estimation, and manifold SDE flood forecasting. These results suggest WCE-based neural operators are a practical and scalable approach to learning SDE/SPDE solution operators across domains.
Authors: Lucas Schott, Elies Gherbi, Hatem Hajri, Sylvain Lamprier
Abstract: Adversarial training in reinforcement learning (RL) is challenging because perturbations cascade through trajectories and compound over time, making fixed-strength attacks either overly destructive or too conservative. We propose reward-preserving attacks, which adapt adversarial strength so that an $\alpha$ fraction of the nominal-to-worst-case return gap remains achievable at each state. In deep RL, perturbation magnitudes $\eta$ are selected dynamically, using a learned critic $Q((s,a),\eta)$ that estimates the expected return of $\alpha$-reward-preserving rollouts. For intermediate values of $\alpha$, this adaptive training yields policies that are robust across a wide range of perturbation magnitudes while preserving nominal performance, outperforming fixed-radius and uniformly sampled-radius adversarial training.
Authors: Yalan Tan, Yanyong Huang, Zongxin Shen, Dongjie Wang, Fengmao Lv, Tianrui Li
Abstract: Although multi-view unsupervised feature selection (MUFS) has demonstrated success in dimensionality reduction for unlabeled multi-view data, most existing methods reduce feature redundancy by focusing on linear correlations among features but often overlook complex nonlinear dependencies. This limits the effectiveness of feature selection. In addition, existing methods fuse similarity graphs from multiple views by employing sample-invariant weights to preserve local structure. However, this process fails to account for differences in local neighborhood clarity among samples within each view, thereby hindering accurate characterization of the intrinsic local structure of the data. In this paper, we propose a Kernel Alignment-based multi-view unsupervised FeatUre selection with Sample-level adaptive graph lEarning method (KAFUSE) to address these issues. Specifically, we first employ kernel alignment with an orthogonal constraint to reduce feature redundancy in both linear and nonlinear relationships. Then, a cross-view consistent similarity graph is learned by applying sample-level fusion to each slice of a tensor formed by stacking similarity graphs from different views, which automatically adjusts the view weights for each sample during fusion. These two steps are integrated into a unified model for feature selection, enabling mutual enhancement between them. Extensive experiments on real multi-view datasets demonstrate the superiority of KAFUSE over state-of-the-art methods.
Authors: Yu-Yang Qian, Junda Su, Lanxiang Hu, Peiyuan Zhang, Zhijie Deng, Peng Zhao, Hao Zhang
Abstract: Diffusion large language models (dLLMs) offer capabilities beyond those of autoregressive (AR) LLMs, such as parallel decoding and random-order generation. However, realizing these benefits in practice is non-trivial, as dLLMs inherently face an accuracy-parallelism trade-off. Despite increasing interest, existing methods typically focus on only one-side of the coin, targeting either efficiency or performance. To address this limitation, we propose d3LLM (Pseudo-Distilled Diffusion Large Language Model), striking a balance between accuracy and parallelism: (i) during training, we introduce pseudo-trajectory distillation to teach the model which tokens can be decoded confidently at early steps, thereby improving parallelism; (ii) during inference, we employ entropy-based multi-block decoding with a KV-cache refresh mechanism to achieve high parallelism while maintaining accuracy. To better evaluate dLLMs, we also introduce AUP (Accuracy Under Parallelism), a new metric that jointly measures accuracy and parallelism. Experiments demonstrate that our d3LLM achieves up to 10$\times$ speedup over vanilla LLaDA/Dream and 5$\times$ speedup over AR models without much accuracy drop. Our code is available at https://github.com/hao-ai-lab/d3LLM.
Authors: Nilin Abrahamsen
Abstract: This note introduces Projected Microbatch Accumulation (PROMA), a proximal policy method that modifies gradient accumulation across microbatches rather than relying on likelihood ratios relative to a reference policy. During accumulation, PROMA projects the partially accumulated gradient to be orthogonal to the sequence-wise gradients of the current microbatch. This projection is applied layer-wise during the backward pass, enabling efficient implementation. A within-microbatch variant (Intra-PROMA) acts independently across microbatches. Empirically, PROMA achieves proximal updates without entropy collapse while providing tighter local KL control than GRPO.
Authors: Zhikang Shen, Jianrong Lu, Haiyuan Wan, Jianhai Chen
Abstract: Federated learning (FL) for large language models (LLMs) has attracted increasing attention as a privacy-preserving approach for adapting models over distributed data, where parameter-efficient methods such as Low-Rank Adaptation (LoRA) are widely adopted to reduce communication and memory costs. However, practical deployments often exhibit rank and data heterogeneity: clients operate under different low-rank budgets and data distributions, making direct aggregation of LoRA updates biased and unstable. Existing approaches either enforce a unified rank or align heterogeneous updates into a single shared subspace, which tends to mix transferable and client-specific directions and consequently undermines personalization. Moreover, under differential privacy (DP), perturbing such structurally mixed updates injects noise into directions that should remain purely local, leading to unnecessary utility degradation. To address these issues, we propose Selective Decoupled Federated LoRA (SDFLoRA), a structure-aware LoRA framework that decouples each client update into a shared component for aggregation and a private component that preserves client-specific semantics. Only the shared component participates in subspace alignment, while the private component remains local and uncommunicated, making the training DP-compatible and stabilizing aggregation under rank heterogeneity. By injecting noise only into the aggregated shareable update, this approach avoids perturbations to local directions and improves the utility-privacy trade-off. Experiments on multiple benchmarks demonstrate that SDFLoRA outperforms federated LoRA baselines and achieves a strong utility-privacy trade-off.
Authors: James O'Neill, Robert Clancy, Mariia Matskevichus, Fergal Reid
Abstract: The key-value (KV) cache is a primary memory bottleneck in Transformers. We propose Low-Rank Key-Value (LRKV) attention, which reduces KV cache memory by exploiting redundancy across attention heads, while being compute efficient. Each layer uses a shared full-rank KV projection augmented with low-rank, head-specific residuals, providing a continuous trade-off between complete sharing and full independence. After pretraining models of size 128M to 6.3B parameters, LRKV consistently achieves the lowest test loss among standard MHA, MQA/GQA, and MLA while using only 45-53\% of MHA's KV cache. LRKV reaches equivalent baseline quality 18-25\% faster (measured in training steps). After supervised midtraining, LRKV achieves the highest downstream task performance across ARC-Easy, ARC-Challenge, MMLU, GSM8K, and HumanEval benchmarks.
Authors: Qingyu Meng, Yangshuai Wang
Abstract: Quantum neural networks (QNNs) have attracted growing interest for scientific machine learning, yet in regression settings they often suffer from limited trainability under noisy gradients and ill-conditioned optimization. We propose a hybrid quantum-classical regression framework designed to mitigate these bottlenecks. Our model prepends a lightweight classical embedding that acts as a learnable geometric preconditioner, reshaping the input representation to better condition a downstream variational quantum circuit. Building on this architecture, we introduce a curriculum optimization protocol that progressively increases circuit depth and transitions from SPSA-based stochastic exploration to Adam-based gradient fine-tuning. We evaluate the approach on PDE-informed regression benchmarks and standard regression datasets under a fixed training budget in a simulator setting. Empirically, the proposed framework consistently improves over pure QNN baselines and yields more stable convergence in data-limited regimes. We further observe reduced structured errors that are visually correlated with oscillatory components on several scientific benchmarks, suggesting that geometric preconditioning combined with curriculum training is a practical approach for stabilizing quantum regression.
Authors: Dirk Tasche
Abstract: Factorizable joint shift (FJS) represents a type of distribution shift (or dataset shift) that comprises both covariate and label shift. Recently, it has been observed that FJS actually arises from consecutive label and covariate (or vice versa) shifts. Research into FJS so far has been confined to the case of categorical labels. We propose a framework for analysing distribution shift in the case of a general label space, thus covering both classification and regression models. Based on the framework, we generalise existing results on FJS to general label spaces and present and analyse a related extension of the expectation maximisation (EM) algorithm for class prior probabilities. We also take a fresh look at generalized label shift (GLS) in the case of a general label space.
Authors: Yuval Ran-Milo, Yotam Alexander, Shahar Mendel, Nadav Cohen
Abstract: Transformers trained via Reinforcement Learning (RL) with outcome-based supervision can spontaneously develop the ability to generate intermediate reasoning steps (Chain-of-Thought). Yet the mechanism by which sparse rewards drive policy gradient to discover such systematic reasoning remains poorly understood. We address this by analyzing the policy gradient dynamics of single-layer Transformers on a synthetic graph traversal task that cannot be solved without Chain-of-Thought but admits a simple iterative solution. We prove that despite training solely on final-answer correctness, policy gradient drives the Transformer to converge to a structured, interpretable algorithm that iteratively traverses the graph vertex-by-vertex. We characterize the distributional properties required for this emergence, identifying the critical role of "simple examples": instances requiring fewer reasoning steps. When the training distribution places sufficient mass on these simpler examples, the Transformer learns a generalizable traversal strategy that extrapolates to longer chains; when this mass vanishes, policy gradient learning becomes infeasible. We corroborate our theoretical results through experiments on synthetic data and with real-world language models on mathematical reasoning tasks, validating that our theoretical findings carry over to practical settings.
Authors: Alejandro Ascarate, Leo Lebrat, Rodrigo Santa Cruz, Clinton Fookes, Olivier Salvado
Abstract: Variational autoencoders (VAE) encode data into lower-dimensional latent vectors before decoding those vectors back to data. Once trained, one can hope to detect out-of-distribution (abnormal) latent vectors, but several issues arise when the latent space is high dimensional. This includes an exponential growth of the hypervolume with the dimension, which severely affects the generative capacity of the VAE. In this paper, we draw insights from high dimensional statistics: in these regimes, the latent vectors of a standard VAE are distributed on the `equators' of a hypersphere, challenging the detection of anomalies. We propose to formulate the latent variables of a VAE using hyperspherical coordinates, which allows compressing the latent vectors towards a given direction on the hypersphere, thereby allowing for a more expressive approximate posterior. We show that this improves both the fully unsupervised and OOD anomaly detection ability of the VAE, achieving the best performance on the datasets we considered, outperforming existing methods. For the unsupervised and OOD modalities, respectively, these are: i) detecting unusual landscape from the Mars Rover camera and unusual Galaxies from ground based imagery (complex, real world datasets); ii) standard benchmarks like Cifar10 and subsets of ImageNet as the in-distribution (ID) class.
Authors: Nima Leclerc, Chris Miller, Nicholas Brawand
Abstract: Quantum hardware suffers from intrinsic device heterogeneity and environmental drift, forcing practitioners to choose between suboptimal non-adaptive controllers or costly per-device recalibration. We derive a scaling law lower bound for meta-learning showing that the adaptation gain (expected fidelity improvement from task-specific gradient steps) saturates exponentially with gradient steps and scales linearly with task variance, providing a quantitative criterion for when adaptation justifies its overhead. Validation on quantum gate calibration shows negligible benefits for low-variance tasks but $>40\%$ fidelity gains on two-qubit gates under extreme out-of-distribution conditions (10$\times$ the training noise), with implications for reducing per-device calibration time on cloud quantum processors. Further validation on classical linear-quadratic control confirms these laws emerge from general optimization geometry rather than quantum-specific physics. Together, these results offer a transferable framework for decision-making in adaptive control.
Authors: Joshua V. Dillon
Abstract: Biological neural systems must be fast but are energy-constrained. Evolution's solution: act on the first signal. Winner-take-all circuits and time-to-first-spike coding implicitly treat when a neuron fires as an expression of confidence. We apply this principle to ensembles of Tiny Recursive Models (TRM) [Jolicoeur-Martineau et al., 2025]. On Sudoku-Extreme, halt-first selection achieves 97% accuracy vs. 91% for probability averaging--while requiring 10x fewer reasoning steps. A single baseline model achieves 85.5% +/- 1.3%. Can we internalize this as a training-only cost? Yes: by maintaining K=4 parallel latent states but backpropping only through the lowest-loss "winner," we achieve 96.9% +/- 0.6% accuracy--matching ensemble performance at 1x inference cost, with less than half the variance of the baseline. A key diagnostic: 89% of baseline failures are selection problems, revealing a 99% accuracy ceiling. As in nature, this work was also resource constrained: all experiments used a single RTX 5090. A modified SwiGLU [Shazeer, 2020] made Muon [Jordan et al., 2024] and high LR viable, enabling baseline training in 48 minutes and full WTA (K=4) in 6 hours on consumer hardware.
Authors: Vijay Janapa Reddi
Abstract: Machine learning education faces a fundamental gap: students learn algorithms without understanding the systems that execute them. They study gradient descent without measuring memory, attention mechanisms without analyzing O(N^2) scaling, optimizer theory without knowing why Adam requires 3x the memory of SGD. This "algorithm-systems divide" produces practitioners who can train models but cannot debug memory failures, optimize inference latency, or reason about deployment trade-offs--the very skills industry demands as "ML systems engineering." We present TinyTorch, a 20-module curriculum that closes this gap through "implementation-based systems pedagogy": students construct PyTorch's core components (tensors, autograd, optimizers, CNNs, transformers) in pure Python, building a complete framework where every operation they invoke is code they wrote. The design employs three patterns: "progressive disclosure" of complexity, "systems-first integration" of profiling from the first module, and "build-to-validate milestones" recreating 67 years of ML breakthroughs--from Perceptron (1958) through Transformers (2017) to MLPerf-style benchmarking. Requiring only 4GB RAM and no GPU, TinyTorch demonstrates that deep ML systems understanding is achievable without specialized hardware. The curriculum is available open-source at mlsysbook.ai/tinytorch.
Authors: Wayner Barrios
Abstract: The growing adoption of Apple Silicon for machine learning development has created demand for efficient inference solutions that leverage its unique unified memory architecture. However, existing tools either lack native optimization (PyTorch MPS) or focus solely on text models, leaving multimodal workloads underserved. We present vllm-mlx, a framework for efficient LLM and MLLM inference on Apple Silicon built natively on MLX. For text models, we achieve 21\% to 87\% higher throughput than llama-cpp across models ranging from Qwen3-0.6B to Nemotron-30B, while providing continuous batching that scales to 4.3x aggregate throughput at 16 concurrent requests. For multimodal models, we introduce content-based prefix caching that eliminates redundant vision encoding by identifying identical images through content hashing, regardless of input format. Our evaluation on Apple M4 Max demonstrates throughput of up to 525 tokens per second on text models and 28x speedup on repeated image queries, reducing multimodal latency from 21.7 seconds to under 1 second. Video analysis with up to 64 frames achieves 24.7x cache speedup. We release our implementation as open source to support efficient inference on consumer Apple hardware.
Authors: Yiying Sheng, Wenhao Ding, Dylan Roi, Leonard Leong Litt Yeo, Hwa Liang Leo, Choon Hwai Yap
Abstract: Extensive studies suggested that fluid mechanical markers of intracranial aneurysms (IAs) derived from Computational Fluid Dynamics (CFD) can indicate disease progression risks, but to date this has not been translated clinically. This is because CFD requires specialized expertise and is time-consuming and low throughput, making it difficult to support clinical trials. A deep learning model that maps IA morphology to biomechanical markers can address this, enabling physicians to obtain these markers in real time without performing CFD. Here, we show that a Graph Transformer model that incorporates temporal information, which is supervised by large CFD data, can accurately predict Wall Shear Stress (WSS) across the cardiac cycle from IA surface meshes. The model effectively captures the temporal variations of the WSS pattern, achieving a Structural Similarity Index (SSIM) of up to 0.981 and a maximum-based relative L2 error of 2.8%. Ablation studies and SOTA comparison confirmed its optimality. Further, as pulsatile CFD data is computationally expensive to generate and sample sizes are limited, we engaged a strategy of injecting a large amount of steady-state CFD data, which are extremely low-cost to generate, as augmentation. This approach enhances network performance substantially when pulsatile CFD data sample size is small. Our study provides a proof of concept that temporal sequences cardiovascular fluid mechanical parameters can be computed in real time using a deep learning model from the geometric mesh, and this is achievable even with small pulsatile CFD sample size. Our approach is likely applicable to other cardiovascular scenarios.
Authors: Mingxuan Luo, Guipeng Xv, Sishuo Chen, Xinyu Li, Li Zhang, Zhangming Chan, Xiang-Rong Sheng, Han Zhu, Jian Xu, Bo Zheng, Chen Lin
Abstract: In industrial recommender systems, conversion rate (CVR) is widely used for traffic allocation, but it fails to fully reflect recommendation effectiveness because it ignores refund behavior. To better capture true user satisfaction and business value, net conversion rate (NetCVR), defined as the probability that a clicked item is purchased and not refunded, has been proposed.Unlike CVR, NetCVR prediction involves a more complex multi-stage cascaded delayed feedback process. The two cascaded delays from click to conversion and from conversion to refund have opposite effects, making traditional CVR modeling methods inapplicable. Moreover, the lack of open-source datasets and online continuous training schemes further hinders progress in this area.To address these challenges, we introduce CASCADE (Cascaded Sequences of Conversion and Delayed Refund), the first large-scale open dataset derived from the Taobao app for online continuous NetCVR prediction. Through an in-depth analysis of CASCADE, we identify three key insights: (1) NetCVR exhibits strong temporal dynamics, necessitating online continuous modeling; (2) cascaded modeling of CVR and refund rate outperforms direct NetCVR modeling; and (3) delay time, which correlates with both CVR and refund rate, is an important feature for NetCVR prediction.Based on these insights, we propose TESLA, a continuous NetCVR modeling framework featuring a CVR-refund-rate cascaded architecture, stage-wise debiasing, and a delay-time-aware ranking loss. Extensive experiments demonstrate that TESLA consistently outperforms state-of-the-art methods on CASCADE, achieving absolute improvements of 12.41 percent in RI-AUC and 14.94 percent in RI-PRAUC on NetCVR prediction. The code and dataset are publicly available at https://github.com/alimama-tech/NetCVR.
Authors: Jinlin Liu, Wei Chen, Xiaojin Zhang
Abstract: Collecting web data to train deep models has become increasingly common, raising concerns about unauthorized data usage. To mitigate this issue, unlearnable examples introduce imperceptible perturbations into data, preventing models from learning effectively. However, existing methods typically rely on deep neural networks as surrogate models for perturbation generation, resulting in significant computational costs. In this work, we propose Perturbation-Induced Linearization (PIL), a computationally efficient yet effective method that generates perturbations using only linear surrogate models. PIL achieves comparable or better performance than existing surrogate-based methods while reducing computational time dramatically. We further reveal a key mechanism underlying unlearnable examples: inducing linearization to deep models, which explains why PIL can achieve competitive results in a very short time. Beyond this, we provide an analysis about the property of unlearnable examples under percentage-based partial perturbation. Our work not only provides a practical approach for data protection but also offers insights into what makes unlearnable examples effective.
Authors: Richard Csaky
Abstract: We present a large autoregressive model for source-space MEG that scales next-token prediction to long context across datasets and scanners: handling a corpus of over 500 hours and thousands of sessions across the three largest MEG datasets. A modified SEANet-style vector-quantizer reduces multichannel MEG into a flattened token stream on which we train a Qwen2.5-VL backbone from scratch to predict the next brain token and to recursively generate minutes of MEG from up to a minute of context. To evaluate long-horizon generation, we introduce task-matched tests: (i) on-manifold stability via generated-only drift compared to the time-resolved distribution of real sliding windows, and (ii) conditional specificity via correct context versus prompt-swap controls using a neurophysiologically grounded metric set. We train on CamCAN and Omega and run all analyses on held-out MOUS, establishing cross-dataset generalization. Across metrics, generations remain relatively stable over long rollouts and are closer to the correct continuation than swapped controls. Code available at: https://github.com/ricsinaruto/brain-gen.
Authors: Alexander Benanti, Xi Han, Hong Qin
Abstract: Numerical techniques for solving partial differential equations (PDEs) are integral for many fields across science and engineering. Such techniques usually involve solving large, sparse linear systems, where preconditioning methods are critical. In recent years, neural methods, particularly graph neural networks (GNNs), have demonstrated their potential through accelerated convergence. Nonetheless, to extract connective structures, existing techniques aggregate discretized system matrices into graphs, and suffer from rank inflation and a suboptimal convergence rate. In this paper, we articulate NeuraLSP, a novel neural preconditioner combined with a novel loss metric that leverages the left singular subspace of the system matrix's near-nullspace vectors. By compressing spectral information into a fixed low-rank operator, our method exhibits both theoretical guarantees and empirical robustness to rank inflation, affording up to a 53% speedup. Besides the theoretical guarantees for our newly-formulated loss function, our comprehensive experimental results across diverse families of PDEs also substantiate the aforementioned theoretical advances.
Authors: Chen Cheng, Ang Li
Abstract: Out-of-distribution (OOD) generalization remains a fundamental challenge in real-world classification, where test distributions often differ substantially from training data. Most existing approaches pursue domain-invariant representations, implicitly assuming that invariance implies reliability. However, features that are invariant across domains are not necessarily causally effective for prediction. In this work, we revisit OOD classification from a causal perspective and propose to evaluate learned representations based on their necessity and sufficiency under distribution shift. We introduce an explicit segment-level framework that directly measures causal effectiveness across domains, providing a more faithful criterion than invariance alone. Experiments on multi-domain benchmarks demonstrate consistent improvements in OOD performance, particularly under challenging domain shifts, highlighting the value of causal evaluation for robust generalization.
Authors: Jinren Ding, Xuejian Xu, Shen Jiang, Zhitong Hao, Jinhui Yang, Peng Jiang
Abstract: Decision Transformer (DT) shows promise for generative auto-bidding by capturing temporal dependencies, but suffers from two critical limitations: insufficient cross-correlation modeling among state, action, and return-to-go (RTG) sequences, and indiscriminate learning of optimal/suboptimal behaviors. To address these, we propose C2, a novel framework enhancing DT with two core innovations: (1) a Cross Learning Block (CLB) via cross-attention to strengthen inter-sequence correlation modeling; (2) a Constraint-aware Loss (CL) incorporating budget and Cost-Per-Acquisition (CPA) constraints for selective learning of optimal trajectories. Extensive offline evaluations on the AuctionNet dataset demonstrate consistent performance gains (up to 3.2% over state-of-the-art method) across diverse budget settings; ablation studies verify the complementary synergy of CLB and CL, confirming C2's superiority in auto-bidding. The code for reproducing our results is available at: https://github.com/Dingjinren/C2.
Authors: Yuhang Liu, Erdun Gao, Dong Gong, Anton van den Hengel, Javen Qinfeng Shi
Abstract: Developing human understandable interpretation of large language models (LLMs) becomes increasingly critical for their deployment in essential domains. Mechanistic interpretability seeks to mitigate the issues through extracts human-interpretable process and concepts from LLMs' activations. Sparse autoencoders (SAEs) have emerged as a popular approach for extracting interpretable and monosemantic concepts by decomposing the LLM internal representations into a dictionary. Despite their empirical progress, SAEs suffer from a fundamental theoretical ambiguity: the well-defined correspondence between LLM representations and human-interpretable concepts remains unclear. This lack of theoretical grounding gives rise to several methodological challenges, including difficulties in principled method design and evaluation criteria. In this work, we show that, under mild assumptions, LLM representations can be approximated as a {linear mixture} of the log-posteriors over concepts given the input context, through the lens of a latent variable model where concepts are treated as latent variables. This motivates a principled framework for concept extraction, namely Concept Component Analysis (ConCA), which aims to recover the log-posterior of each concept from LLM representations through a {unsupervised} linear unmixing process. We explore a specific variant, termed sparse ConCA, which leverages a sparsity prior to address the inherent ill-posedness of the unmixing problem. We implement 12 sparse ConCA variants and demonstrate their ability to extract meaningful concepts across multiple LLMs, offering theory-backed advantages over SAEs.
Authors: Ziwei Liu, Borui Kang, Hangjie Yuan, Zixiang Zhao, Wei Li, Yifan Zhu, Tao Feng
Abstract: As digital environments (data distribution) are in flux, with new GUI data arriving over time-introducing new domains or resolutions-agents trained on static environments deteriorate in performance. In this work, we introduce Continual GUI Agents, a new task that requires GUI agents to perform continual learning under shifted domains and resolutions. We find existing methods fail to maintain stable grounding as GUI distributions shift over time, due to the diversity of UI interaction points and regions in fluxing scenarios. To address this, we introduce GUI-Anchoring in Flux (GUI-AiF), a new reinforcement fine-tuning framework that stabilizes continual learning through two novel rewards: Anchoring Point Reward in Flux (APR-iF) and Anchoring Region Reward in Flux (ARR-iF). These rewards guide the agents to align with shifting interaction points and regions, mitigating the tendency of existing reward strategies to over-adapt to static grounding cues (e.g., fixed coordinates or element scales). Extensive experiments show GUI-AiF surpasses state-of-the-art baselines. Our work establishes the first continual learning framework for GUI agents, revealing the untapped potential of reinforcement fine-tuning for continual GUI Agents.
Authors: Zihao Wang, Hang Yin, Lihui Liu, Hanghang Tong, Yangqiu Song, Ginny Wong, Simon See
Abstract: This paper studies the minimal dimension required to embed subset memberships ($m$ elements and ${m\choose k}$ subsets of at most $k$ elements) into vector spaces, denoted as Minimal Embeddable Dimension (MED). The tight bounds of MED are derived theoretically and supported empirically for various notions of "distances" or "similarities," including the $\ell_2$ metric, inner product, and cosine similarity. In addition, we conduct numerical simulation in a more achievable setting, where the ${m\choose k}$ subset embeddings are chosen as the centroid of the embeddings of the contained elements. Our simulation easily realizes a logarithmic dependency between the MED and the number of elements to embed. These findings imply that embedding-based retrieval limitations stem primarily from learnability challenges, not geometric constraints, guiding future algorithm design.
Authors: Serena Della Corte, Laurens Van Mieghem, Antonis Papapantoleon, Jonas Papazoglou-Hennig
Abstract: We consider the supervised learning problem of learning the price of an option or the implied volatility given appropriate input data (model parameters) and corresponding output data (option prices or implied volatilities). The majority of articles in this literature considers a (plain) feed forward neural network architecture in order to connect the neurons used for learning the function mapping inputs to outputs. In this article, motivated by methods in image classification and recent advances in machine learning methods for PDEs, we investigate empirically whether and how the choice of network architecture affects the accuracy and training time of a machine learning algorithm. We find that the generalized highway network architecture achieves the best performance, when considering the mean squared error and the training time as criteria, within the considered parameter budgets for the Black-Scholes and Heston option pricing problems. Considering the transformed implied volatility problem, a simplified DGM variant achieves the lowest error among the tested architectures. We also carry out a capacity-normalised comparison for completeness, where all architectures are evaluated with an equal number of parameters. Finally, for the implied volatility problem, we additionally include experiments using real market data.
Authors: Berivan Isik, Natalia Ponomareva, Hussein Hazimeh, Dimitris Paparas, Sergei Vassilvitskii, Sanmi Koyejo
Abstract: Scaling laws provide important insights that can guide the design of large language models (LLMs). Existing work has primarily focused on studying scaling laws for pretraining (upstream) loss. However, in transfer learning settings, in which LLMs are pretrained on an unsupervised dataset and then finetuned on a downstream task, we often also care about the downstream performance. In this work, we study the scaling behavior in a transfer learning setting, where LLMs are finetuned for machine translation tasks. Specifically, we investigate how the choice of the pretraining data and its size affect downstream performance (translation quality) as judged by: downstream cross-entropy and translation quality metrics such as BLEU and COMET scores. Our experiments indicate that the size of the finetuning dataset and the distribution alignment between the pretraining and downstream data significantly influence the scaling behavior. With sufficient alignment, both downstream cross-entropy and translation quality scores improve monotonically with more pretraining data. In such cases, we show that it is possible to predict the downstream translation quality metrics with good accuracy using a log-law. However, there are cases where moderate misalignment causes the downstream translation scores to fluctuate or get worse with more pretraining, whereas downstream cross-entropy monotonically improves. By analyzing these, we provide new practical insights for choosing appropriate pretraining data.
Authors: Aswini Kumar Patra, Ankit Varshney, Lingaraj Sahoo
Abstract: Early detection of drought stress is critical for taking timely measures for reducing crop loss before the drought impact becomes irreversible. The subtle phenotypical and physiological changes in response to drought stress are captured by non-invasive imaging techniques and these imaging data serve as valuable resource for machine learning methods to identify drought stress. While convolutional neural networks (CNNs) are in wide use, vision transformers (ViTs) present a promising alternative in capturing long-range dependencies and intricate spatial relationships, thereby enhancing the detection of subtle indicators of drought stress. We propose an explainable deep learning pipeline that leverages the power of ViTs for drought stress detection in potato crops using aerial imagery. We applied two distinct approaches: a synergistic combination of ViT and support vector machine (SVM), where ViT extracts intricate spatial features from aerial images, and SVM classifies the crops as stressed or healthy and an end-to-end approach using a dedicated classification layer within ViT to directly detect drought stress. Our key findings explain the ViT model's decision-making process by visualizing attention maps. These maps highlight the specific spatial features within the aerial images that the ViT model focuses as the drought stress signature. Our findings demonstrate that the proposed methods not only achieve high accuracy in drought stress identification but also shedding light on the diverse subtle plant features associated with drought stress. This offers a robust and interpretable solution for drought stress monitoring for farmers to undertake informed decisions for improved crop management.
Authors: Yu-Shin Huang, Peter Just, Hanyun Yin, Krishna Narayanan, Ruihong Huang, Chao Tian
Abstract: We consider coverless steganography where a Large Language Model (LLM) is used to generate stego-texts in combination with arithmetic coding. An efficient method should embed secret bits in as few language tokens as possible while keeping the stego-text as natural as possible. We show that this problem is equivalent to maximizing the entropy of a replacement probability distribution of the next token generation, subject to a constraint on the divergence between the new distribution and the original one produced by the LLM. A closed-form solution is provided under either the KL divergence or the total variation constraint. Several important practical issues are also tackled: 1) An often-overlooked tokenization mismatch issue is resolved with a simple prompt selection approach, 2) The combination of the optimized distribution and the vocabulary truncation technique is considered, and 3) The incorporation of the proposed approach with existing (potentially non arithmetic coding based) techniques, e.g., the Discop technique.
Authors: Zhendong Liu, Yi Nian, Yuehan Qin, Henry Peng Zou, Li Li, Xiyang Hu, Yue Zhao
Abstract: How can models effectively detect out-of-distribution (OOD) samples in complex, multi-label settings without extensive retraining? Existing OOD detection methods struggle to capture the intricate semantic relationships and label co-occurrences inherent in multi-label settings, often requiring large amounts of training data and failing to generalize to unseen label combinations. While large language models have revolutionized zero-shot OOD detection, they primarily focus on single-label scenarios, leaving a critical gap in handling real-world tasks where samples can be associated with multiple interdependent labels. To address these challenges, we introduce COOD, a novel zero-shot multi-label OOD detection framework. COOD leverages pre-trained vision-language models, enhancing them with a concept-based label expansion strategy and a new scoring function. By enriching the semantic space with both positive and negative concepts for each label, our approach models complex label dependencies, precisely differentiating OOD samples without the need for additional training. Extensive experiments demonstrate that our method significantly outperforms existing approaches, achieving approximately 95% average AUROC on both VOC and COCO datasets, while maintaining robust performance across varying numbers of labels and different types of OOD samples.
Authors: Yubo Wang, Haoyang Li, Fei Teng, Lei Chen
Abstract: Text classification is vital for Web for Good applications like hate speech and misinformation detection. However, traditional models (e.g., BERT) often fail in dynamic few-shot settings where labeled data are scarce, and target labels frequently evolve. While Large Language Models (LLMs) show promise in few-shot settings, their performance is often hindered by increased input size in dynamic evolving scenarios. To address these issues, we propose GORAG, a Graph-based Online Retrieval-Augmented Generation framework for dynamic few-shot text classification. GORAG constructs and maintains a weighted graph of keywords and text labels, representing their correlations as edges. To model these correlations, GORAG employs an edge weighting mechanism to prioritize the importance and reliability of extracted information and dynamically retrieves relevant context using a tailored minimum-cost spanning tree for each input. Empirical evaluations show GORAG outperforms existing approaches by providing more comprehensive and precise contextual information. Our code is released at: https://github.com/Wyb0627/GORAG.
Authors: Jiarong Wu, Pavel Perezhogin, David John Gagne, Brandon Reichl, Aneesh C. Subramanian, Elizabeth Thompson, Laure Zanna
Abstract: Accurately quantifying air-sea fluxes is important for understanding air-sea interactions and improving coupled weather and climate systems. This study introduces a probabilistic framework to represent the highly variable nature of air-sea fluxes, which is missing in deterministic bulk algorithms. Assuming Gaussian distributions conditioned on the input variables, we use artificial neural networks and eddy-covariance measurement data to estimate the mean and variance by minimizing negative log-likelihood loss. The trained neural networks provide alternative mean flux estimates to existing bulk algorithms, and quantify the uncertainty around the mean estimates. Stochastic parameterization of air-sea turbulent fluxes can be constructed by sampling from the predicted distributions. Tests in a single-column forced upper-ocean model suggest that changes in flux algorithms influence sea surface temperature and mixed layer depth seasonally. The ensemble spread in stochastic runs is most pronounced during spring restratification.
Authors: Andrew Millard, Joshua Murphy, Daniel Frisch, Simon Maskell
Abstract: Markov chain Monte Carlo (MCMC) methods are a powerful but computationally expensive way of performing non-parametric Bayesian inference. MCMC proposals which utilise gradients, such as Hamiltonian Monte Carlo (HMC), can better explore the parameter space of interest if the additional hyper-parameters are chosen well. The No-U-Turn Sampler (NUTS) is a variant of HMC which is extremely effective at selecting these hyper-parameters but is slow to run and is not suited to GPU architectures. An alternative to NUTS, Change in the Estimator of the Expected Square HMC (ChEES-HMC) was shown not only to run faster than NUTS on GPU but also sample from posteriors more efficiently. Sequential Monte Carlo (SMC) samplers are another sampling method which instead output weighted samples from the posterior. They are very amenable to parallelisation and therefore being run on GPUs while having additional flexibility in their choice of proposal over MCMC. We incorporate (ChEEs-HMC) as a proposal into SMC samplers and demonstrate competitive but faster performance than NUTS on a number of tasks.
Authors: Rijul Magu, Arka Dutta, Sean Kim, Ashiqur R. KhudaBukhsh, Munmun De Choudhury
Abstract: Large Language Models (LLMs) have been shown to demonstrate imbalanced biases against certain groups. However, the study of unprovoked targeted attacks by LLMs towards at-risk populations remains underexplored. Our paper presents three novel contributions: (1) the explicit evaluation of LLM-generated attacks on highly vulnerable mental health groups; (2) a network-based framework to study the propagation of relative biases; and (3) an assessment of the relative degree of stigmatization that emerges from these attacks. Our analysis of a recently released large-scale bias audit dataset reveals that mental health entities occupy central positions within attack narrative networks, as revealed by a significantly higher mean centrality of closeness (p-value = 4.06e-10) and dense clustering (Gini coefficient = 0.7). Drawing from an established stigmatization framework, our analysis indicates increased labeling components for mental health disorder-related targets relative to initial targets in generation chains. Taken together, these insights shed light on the structural predilections of large language models to heighten harmful discourse and highlight the need for suitable approaches for mitigation.
Authors: Rebecca Pelke, Nils Bosbach, Lennart M. Reimann, Rainer Leupers
Abstract: Accelerating Machine Learning (ML) workloads requires efficient methods due to their large optimization space. Autotuning has emerged as an effective approach for systematically evaluating variations of implementations. Traditionally, autotuning requires the workloads to be executed on the target hardware (HW). We present an interface that allows executing autotuning workloads on simulators. This approach offers high scalability when the availability of the target HW is limited, as many simulations can be run in parallel on any accessible HW. Additionally, we evaluate the feasibility of using fast instruction-accurate simulators for autotuning. We train various predictors to forecast the performance of ML workload implementations on the target HW based on simulation statistics. Our results demonstrate that the tuned predictors are highly effective. The best workload implementation in terms of actual run time on the target HW is always within the top 3 % of predictions for the tested x86, ARM, and RISC-V-based architectures. In the best case, this approach outperforms native execution on the target HW for embedded architectures when running as few as three samples on three simulators in parallel.
Authors: Puwei Lian, Yujun Cai, Songze Li, Bingkun Bao
Abstract: Diffusion models have achieved tremendous success in image generation, but they also raise significant concerns regarding privacy and copyright issues. Membership Inference Attacks (MIAs) are designed to ascertain whether specific data were utilized during a model's training phase. As current MIAs for diffusion models typically exploit the model's image prediction ability, we formalize them into a unified general paradigm which computes the membership score for membership identification. Under this paradigm, we empirically find that existing attacks overlook the inherent deficiency in how diffusion models process high-frequency information. Consequently, this deficiency leads to member data with more high-frequency content being misclassified as hold-out data, and hold-out data with less high-frequency content tend to be misclassified as member data. Moreover, we theoretically demonstrate that this deficiency reduces the membership advantage of attacks, thereby interfering with the effective discrimination of member data and hold-out data. Based on this insight, we propose a plug-and-play high-frequency filter module to mitigate the adverse effects of the deficiency, which can be seamlessly integrated into any attacks within this general paradigm without additional time costs. Extensive experiments corroborate that this module significantly improves the performance of baseline attacks across different datasets and models.
Authors: Ander Artola Velasco, Stratis Tsirtsis, Nastaran Okati, Manuel Gomez-Rodriguez
Abstract: State-of-the-art large language models require specialized hardware and substantial energy to operate. As a consequence, cloud-based services that provide access to large language models have become very popular. In these services, the price users pay for an output provided by a model depends on the number of tokens the model uses to generate it: they pay a fixed price per token. In this work, we show that this pricing mechanism creates a financial incentive for providers to strategize and misreport the (number of) tokens a model used to generate an output, and users cannot prove, or even know, whether a provider is overcharging them. However, we also show that, if an unfaithful provider is obliged to be transparent about the generative process used by the model, misreporting optimally without raising suspicion is hard. Nevertheless, as a proof-of-concept, we develop an efficient heuristic algorithm that allows providers to significantly overcharge users without raising suspicion. Crucially, we demonstrate that the cost of running the algorithm is lower than the additional revenue from overcharging users, highlighting the vulnerability of users under the current pay-per-token pricing mechanism. Further, we show that, to eliminate the financial incentive to strategize, a pricing mechanism must price tokens linearly on their character count. While this makes a provider's profit margin vary across tokens, we introduce a simple prescription under which the provider who adopts such an incentive-compatible pricing mechanism can maintain the average profit margin they had under the pay-per-token pricing mechanism. Along the way, to illustrate and complement our theoretical results, we conduct experiments with several large language models from the $\texttt{Llama}$, $\texttt{Gemma}$ and $\texttt{Ministral}$ families, and input prompts from the LMSYS Chatbot Arena platform.
Authors: Guillaume Levy, Cedric Colas, Pierre-Yves Oudeyer, Thomas Carta, Clement Romac
Abstract: Large Language Models (LLMs) possess general world knowledge but often struggle to generate precise predictions in structured, domain-specific contexts such as simulations. These limitations arise from their inability to ground their broad, unstructured understanding in specific environments. To address this, we present WorldLLM, a framework that enhances LLM-based world modeling by combining Bayesian inference and autonomous active exploration with reinforcement learning. WorldLLM leverages the in-context learning abilities of LLMs to guide an LLM-based world model's predictions using natural language hypotheses given in its prompt. These hypotheses are iteratively refined through a Bayesian inference framework that leverages a second LLM as the proposal distribution given collected evidence. This evidence is collected using a curiosity-driven reinforcement learning policy that explores the environment to find transitions with a low log-likelihood under our LLM-based predictive model using the current hypotheses. By alternating between refining hypotheses and collecting new evidence, our framework autonomously drives continual improvement of the predictions. Our experiments demonstrate the effectiveness of WorldLLM in a textual game environment that requires agents to manipulate and combine objects. The framework not only enhances predictive accuracy, but also generates human-interpretable theories of environment dynamics.
Authors: Md Aminul Islam, Kathryn Vasilaky, Elena Zheleva
Abstract: Implicit feedback data, such as user clicks, is commonly used in learning-to-rank (LTR) systems because it is easy to collect and it often reflects user preferences. However, this data is prone to various biases, and training an LTR algorithm directly on biased data can result in suboptimal ranking performance. One of the most prominent and well-studied biases in implicit feedback data is position bias, which occurs because users are more likely to interact with higher-ranked items regardless of their true relevance. In this paper, we propose a novel control function-based method that accounts for position bias in a two-stage process. The first stage uses exogenous variation from the residuals of the ranking process to correct for position bias in the second stage click equation. Unlike previous position bias correction methods, our method does not require knowledge of the click or propensity model and allows for nonlinearity in the underlying ranking model. Moreover, our method is general and allows for debiasing any state-of-the-art ranking algorithm by plugging it into the second stage. We also introduce a new technique to debias validation clicks for hyperparameter tuning to select the optimal model in the absence of unbiased validation data. Experimental results show that our method outperforms state-of-the-art approaches in correcting for position bias.
Authors: Ruijia Zhang, Xiangyu Zhang, Zhengling Qi, Yue Wu, Yanxun Xu
Abstract: Dynamic treatment regimes (DTRs) provide a principled framework for optimizing sequential decision-making in domains where decisions must adapt over time in response to individual trajectories, such as healthcare, education, and digital interventions. However, existing statistical methods often rely on strong positivity assumptions and lack robustness under partial data coverage, while offline reinforcement learning approaches typically focus on average training performance, lack statistical guarantees, and require solving complex optimization problems. To address these challenges, we propose POLAR, a novel pessimistic model-based policy learning algorithm for offline DTR optimization. POLAR estimates the transition dynamics from offline data and quantifies uncertainty for each history-action pair. A pessimistic penalty is then incorporated into the reward function to discourage actions with high uncertainty. Unlike many existing methods that focus on average training performance or provide guarantees only for an oracle policy, POLAR directly targets the suboptimality of the final learned policy and offers theoretical guarantees, without relying on computationally intensive minimax or constrained optimization procedures. To the best of our knowledge, POLAR is the first model-based DTR method to provide both statistical and computational guarantees, including finite-sample bounds on policy suboptimality. Empirical results on both synthetic data and the MIMIC-III dataset demonstrate that POLAR outperforms state-of-the-art methods and yields near-optimal, history-aware treatment strategies.
Authors: Peibo Li, Shuang Ao, Hao Xue, Yang Song, Maarten de Rijke, Johan Barth\'elemy, Tomasz Bednarz, Flora D. Salim
Abstract: Advancing large language models (LLMs) for the next point-of-interest (POI) recommendation task faces two fundamental challenges: (i) although existing methods produce semantic IDs that incorporate semantic information, their topology-blind indexing fails to preserve semantic continuity, meaning that proximity in ID values does not mirror the coherence of the underlying semantics; and (ii) supervised fine-tuning (SFT)-based methods restrict model outputs to top-1 predictions. These approaches suffer from "answer fixation" and neglect the need for top-k ranked lists and reasoning due to the scarcity of supervision. We propose Refine-POI, a framework that addresses these challenges through topology-aware ID generation and reinforcement fine-tuning. First, we introduce a hierarchical self-organizing map (SOM) quantization strategy to generate semantic IDs, ensuring that coordinate proximity in the codebook reflects semantic similarity in the latent space. Second, we employ a policy-gradient framework to optimize the generation of top-k recommendation lists, liberating the model from strict label matching. Extensive experiments on three real-world datasets demonstrate that Refine-POI significantly outperforms state-of-the-art baselines, effectively synthesizing the reasoning capabilities of LLMs with the representational fidelity required for accurate and explainable next-POI recommendation.
Authors: Thanh Hoang-Minh
Abstract: Along with the explosion of large language models, improvements in speech synthesis, advancements in hardware, and the evolution of computer graphics, the current bottleneck in creating digital humans lies in generating character movements that correspond naturally to text or speech inputs. In this work, we present DeepGesture, a diffusion-based gesture synthesis framework for generating expressive co-speech gestures conditioned on multimodal signals - text, speech, emotion, and seed motion. Built upon the DiffuseStyleGesture model, DeepGesture introduces novel architectural enhancements that improve semantic alignment and emotional expressiveness in generated gestures. Specifically, we integrate fast text transcriptions as semantic conditioning and implement emotion-guided classifier-free diffusion to support controllable gesture generation across affective states. To visualize results, we implement a full rendering pipeline in Unity based on BVH output from the model. Evaluation on the ZeroEGGS dataset shows that DeepGesture produces gestures with improved human-likeness and contextual appropriateness. Our system supports interpolation between emotional states and demonstrates generalization to out-of-distribution speech, including synthetic voices - marking a step forward toward fully multimodal, emotionally aware digital humans.
Authors: Jiaqi Zhang, Juntuo Wang, Zhixin Sun, John Zou, Randall Balestriero
Abstract: Large-scale vision foundation models such as DINOv2 boast impressive performances by leveraging massive architectures and training datasets. But numerous scenarios require practitioners to reproduce those pre-training solutions, such as on private data, new modalities, or simply for scientific questioning--which is currently extremely demanding computation-wise. We thus propose a novel pre-training strategy for DINOv2 that simultaneously accelerates convergence--and strengthens robustness to common corruptions as a by-product. Our approach involves a frequency filtering curriculum--low-frequency being seen first--and the Gaussian noise patching augmentation. Applied to a ViT-B/16 backbone trained on ImageNet-1K, while pre-training time and FLOPs are reduced by 1.6x and 2.25x, our method still achieves matching robustness in corruption benchmarks (ImageNet-C) and maintains competitive linear probing performance compared with baseline. This dual benefit of efficiency and robustness makes large-scale self-supervised foundation modeling more attainable, while opening the door to novel exploration around data curriculum and augmentation as means to improve self-supervised learning models robustness. The code is available at https://github.com/KevinZ0217/fast_dinov2
Authors: Tim Gyger, Reinhard Furrer, Fabio Sigrist
Abstract: Gaussian processes are flexible, probabilistic, non-parametric models widely used in machine learning and statistics. However, their scalability to large data sets is limited by computational constraints. To overcome these challenges, we propose Vecchia-inducing-points full-scale (VIF) approximations combining the strengths of global inducing points and local Vecchia approximations. Vecchia approximations excel in settings with low-dimensional inputs and moderately smooth covariance functions, while inducing point methods are better suited to high-dimensional inputs and smoother covariance functions. Our VIF approach bridges these two regimes by using an efficient correlation-based neighbor-finding strategy for the Vecchia approximation of the residual process, implemented via a modified cover tree algorithm. We further extend our framework to non-Gaussian likelihoods by introducing iterative methods that substantially reduce computational costs for training and prediction by several orders of magnitudes compared to Cholesky-based computations when using a Laplace approximation. In particular, we propose and compare novel preconditioners and provide theoretical convergence results. Extensive numerical experiments on simulated and real-world data sets show that VIF approximations are both computationally efficient as well as more accurate and numerically stable than state-of-the-art alternatives. All methods are implemented in the open source C++ library GPBoost with high-level Python and R interfaces.
Authors: Guojun Huang, Kai Zhou
Abstract: In this paper, we study complex-valued neural network (CVNNs) with tensor-valued hidden-to-output weights within the framework of neural-network quantum field theory (NN-QFT). For standard CVNNs with scalar weights, we derive the generating functional and identify the exact Gaussian process that arises in the infinite-width limit, together with its associated effective quantum state. When the last-layer weights are promoted to Clifford-algebra-valued tensors, the network output becomes complex matrix-valued, and a fermion-like sign structure in the large-width correlation functions of the network output is induced. We show that, in the infinite-width limit, correlators with equal numbers of $f^{\dag}$ and $f$ obey fermionic Wick rules and can be written as determinants built from a scalar Euclidean kernel $S(x,y)=\langle f^{\dag}(x)f(y)\rangle$. This provides a sign-structured extension of NN-QFT at the level of Euclidean correlators and Feynman rules, even though a microscopic Grassmann path integral representation for the network parameters has not yet been constructed. Our analysis thus pushes the NN-QFT correspondence beyond purely bosonic Gaussian fields and suggests a possible route to encoding fermion-like symmetries in neural architectures for QFT correspondence.
Authors: Samaneh Shafee, Alysson Bessani, Pedro M. Ferreira
Abstract: Cyber Threat Intelligence (CTI) has emerged as a vital complementary approach that operates in the early phases of the cyber threat lifecycle. CTI involves collecting, processing, and analyzing threat data to provide a more accurate and rapid understanding of cyber threats. Due to the large volume of data, automation through Machine Learning (ML) and Natural Language Processing (NLP) models is essential for effective CTI extraction. These automated systems leverage Open Source Intelligence (OSINT) from sources like social networks, forums, and blogs to identify Indicators of Compromise (IoCs). Although prior research has focused on adversarial attacks on specific ML models, this study expands the scope by investigating vulnerabilities within various components of the entire CTI pipeline and their susceptibility to adversarial attacks. These vulnerabilities arise because they ingest textual inputs from various open sources, including real and potentially fake content. We analyse three types of attacks against CTI pipelines, including evasion, flooding, and poisoning, and assess their impact on the system's information selection capabilities. Specifically, on fake text generation, the work demonstrates how adversarial text generation techniques can create fake cybersecurity and cybersecurity-like text that misleads classifiers, degrades performance, and disrupts system functionality. The focus is primarily on the evasion attack, as it precedes and enables flooding and poisoning attacks within the CTI pipeline.
Authors: Chengyuan Zhang, Zhengbing He, Cathy Wu, Lijun Sun
Abstract: Modeling car-following behavior is fundamental to microscopic traffic simulation, yet traditional deterministic models often fail to capture the full extent of variability and unpredictability in human driving. While many modern approaches incorporate context-aware inputs (e.g., spacing, speed, relative speed), they frequently overlook structured stochasticity that arises from latent driver intentions, perception errors, and memory effects -- factors that are not directly observable from context alone. To fill the gap, this study introduces an interpretable stochastic modeling framework that captures not only context-dependent dynamics but also residual variability beyond what context can explain. Leveraging deep neural networks integrated with nonstationary Gaussian processes (GPs), our model employs a scenario-adaptive Gibbs kernel to learn dynamic temporal correlations in acceleration decisions, where the strength and duration of correlations between acceleration decisions evolve with the driving context. This formulation enables a principled, data-driven quantification of uncertainty in acceleration, speed, and spacing, grounded in both observable context and latent behavioral variability. Comprehensive experiments on the naturalistic vehicle trajectory dataset collected from the German highway, i.e., the HighD dataset, demonstrate that the proposed stochastic simulation method within this framework surpasses conventional methods in both predictive performance and interpretable uncertainty quantification. The integration of interpretability and accuracy makes this framework a promising tool for traffic analysis and safety-critical applications.
Authors: Bonwoo Lee, Cheolwoo Park, Jeongyoun Ahn
Abstract: Differential privacy has become a cornerstone in the development of privacy-preserving learning algorithms. This work addresses optimizing differentially private kernel learning within the empirical risk minimization (ERM) framework. We propose a novel differentially private kernel ERM algorithm based on random projection in the reproducing kernel Hilbert space using Gaussian processes. Our method achieves minimax-optimal excess risk for both the squared loss and Lipschitz-smooth convex loss functions under a local strong convexity condition. We further show that existing approaches based on alternative dimension reduction techniques, such as random Fourier feature mappings or $\ell_2$ regularization, yield suboptimal generalization performance. Our key theoretical contribution also includes the derivation of dimension-free generalization bounds for objective perturbation-based private linear ERM -- marking the first such result that does not rely on noisy gradient-based mechanisms. Additionally, we obtain sharper generalization bounds for existing differentially private kernel ERM algorithms. Empirical evaluations support our theoretical claims, demonstrating that random projection enables statistically efficient and optimally private kernel learning. These findings provide new insights into the design of differentially private algorithms and highlight the central role of dimension reduction in balancing privacy and utility.
Authors: Xiaopeng Li, Shasha Li, Xi Wang, Shezheng Song, Bin Ji, Shangwen Wang, Jun Ma, Xiaodong Liu, Mina Liu, Jie Yu
Abstract: Large Language Models (LLMs) power numerous AI applications, yet updating their knowledge remains costly. Model editing provides a lightweight alternative through targeted parameter modifications, with meta-learning-based model editing (MLME) demonstrating strong effectiveness and efficiency. However, we find that MLME struggles in low-data regimes and incurs high training costs due to the use of KL divergence. To address these issues, we propose $\textbf{E}$fficient $\textbf{M}$ulti-$\textbf{S}$tep $\textbf{Edit (EMSEdit)}$, which leverages multi-step backpropagation (MSBP) to effectively capture gradient-activation mapping patterns within editing samples, performs multi-step edits per sample to enhance editing performance under limited data, and introduces norm-based regularization to preserve unedited knowledge while improving training efficiency. Experiments on two datasets and three LLMs show that EMSEdit consistently outperforms state-of-the-art methods in both sequential and batch editing. Moreover, MSBP can be seamlessly integrated into existing approaches to yield additional performance gains. Further experiments on a multi-hop reasoning editing task demonstrate EMSEdit's robustness in handling complex edits, while ablation studies validate the contribution of each design component. Our code is available at https://github.com/xpq-tech/emsedit.
Authors: \v{S}imon Kucharsk\'y, Aayush Mishra, Daniel Habermann, Stefan T. Radev, Paul-Christian B\"urkner
Abstract: Amortized Bayesian inference (ABI) offers fast, scalable approximations to posterior densities by training neural surrogates on data simulated from the statistical model. However, ABI methods are highly sensitive to model misspecification: when observed data fall outside the training distribution (generative scope of the statistical models), neural surrogates can behave unpredictably. This makes it a challenge in a model comparison setting, where multiple statistical models are considered, of which at least some are misspecified. Recent work on self-consistency (SC) provides a promising remedy to this issue, accessible even for empirical data (without ground-truth labels). In this work, we investigate how SC can improve amortized model comparison conceptualized in four different ways. Across two synthetic and two real-world case studies, we find that approaches for model comparison that estimate marginal likelihoods through approximate parameter posteriors consistently outperform methods that directly approximate model evidence or posterior model probabilities. SC training improves robustness when the likelihood is available, even under severe model misspecification. The benefits of SC for methods without access of analytic likelihoods are more limited and inconsistent. Our results suggest practical guidance for reliable amortized Bayesian model comparison: prefer parameter posterior-based methods and augment them with SC training on empirical datasets to mitigate extrapolation bias under model misspecification.
Authors: Yuzhi Yang, Omar Alhussein, M\'erouane Debbah
Abstract: We propose a novel diffusion model, termed the non-identical diffusion model, and investigate its application to wireless orthogonal frequency division multiplexing (OFDM) channel generation. Unlike the standard diffusion model that uses a scalar-valued time index to represent the global noise level, we extend this notion to an element-wise time indicator to capture local error variations more accurately. Non-identical diffusion enables us to characterize the reliability of each element (e.g., subcarriers in OFDM) within the noisy input, leading to improved generation results when the initialization is biased. Specifically, we focus on the recovery of wireless multi-input multi-output (MIMO) OFDM channel matrices, where the initial channel estimates exhibit highly uneven reliability across elements due to the pilot scheme. Conventional time embeddings, which assume uniform noise progression, fail to capture such variability across pilot schemes and noise levels. We introduce a matrix that matches the input size to control element-wise noise progression. Following a similar diffusion procedure to existing methods, we show the correctness and effectiveness of the proposed non-identical diffusion scheme both theoretically and numerically. For MIMO-OFDM channel generation, we propose a dimension-wise time embedding strategy. We also develop and evaluate multiple training and generation methods and compare them through numerical experiments.
Authors: Jung-hun Kim, Min-hwan Oh
Abstract: We introduce a bandit framework for stochastic matching under the multinomial logit (MNL) choice model. In our setting, $N$ agents on one side are assigned to $K$ arms on the other side, where each arm stochastically selects an agent from its assigned pool according to unknown preferences and yields a corresponding reward over a horizon $T$. The objective is to minimize regret by maximizing the cumulative revenue from successful matches. A naive approach requires solving an NP-hard combinatorial optimization problem at every round, resulting in a prohibitive computational cost. To address this challenge, we propose batched algorithms that strategically limit the number of times matching assignments are updated to $\Theta(\log\log T)$ over the entire horizon. By invoking expensive combinatorial optimization only on a vanishing fraction of rounds, our algorithms substantially reduce overall computational overhead while still achieving a regret bound of $\widetilde{\mathcal{O}}(\sqrt{T})$.
Authors: Nicolas Chatzikiriakos, Kevin Jamieson, Andrea Iannelli
Abstract: In this work, we consider the problem of identifying an unknown linear dynamical system given a finite hypothesis class. In particular, we analyze the effect of the excitation input on the sample complexity of identifying the true system with high probability. To this end, we present sample complexity lower bounds that capture the choice of the selected excitation input. The sample complexity lower bound gives rise to a system theoretic condition to determine the potential benefit of experiment design. Informed by the analysis of the sample complexity lower bound, we propose a persistent excitation (PE) condition tailored to the considered setting, which we then use to establish sample complexity upper bounds. Notably, the PE condition is weaker than in the case of an infinite hypothesis class and allows analyzing different excitation inputs modularly. Crucially, the lower and upper bounds share the same dependency on key problem parameters. Finally, we leverage these insights to propose an active learning algorithm that sequentially excites the system optimally with respect to the current estimate, and provide sample complexity guarantees for the presented algorithm. Concluding simulations showcase the effectiveness of the proposed algorithm.
Authors: Francisco Messina, Francesca Ronchini, Luca Comanducci, Paolo Bestagini, Fabio Antonacci
Abstract: A persistent challenge in generative audio models is data replication, where the model unintentionally generates parts of its training data during inference. In this work, we address this issue in text-to-audio diffusion models by exploring the use of anti-memorization strategies. We adopt Anti-Memorization Guidance (AMG), a technique that modifies the sampling process of pre-trained diffusion models to discourage memorization. Our study explores three types of guidance within AMG, each designed to reduce replication while preserving generation quality. We use Stable Audio Open as our backbone, leveraging its fully open-source architecture and training dataset. Our comprehensive experimental analysis suggests that AMG significantly mitigates memorization in diffusion-based text-to-audio generation without compromising audio fidelity or semantic alignment.
Authors: Mahindra Singh Rautela, Alexander Most, Siddharth Mansingh, Bradley C. Love, Alexander Scheinker, Diane Oyen, Nathan Debardeleben, Earl Lawrence, Ayan Biswas
Abstract: We introduce MORPH, a modality-agnostic, autoregressive foundation model for partial differential equations (PDEs). MORPH is built on a convolutional vision transformer backbone that seamlessly handles heterogeneous spatiotemporal datasets of varying data modality (1D--3D) at different resolutions, and multiple fields with mixed scalar and vector components. The architecture combines (i) component-wise convolution, which jointly processes scalar and vector channels to capture local interactions, (ii) inter-field cross-attention, which models and selectively propagates information between different physical fields, (iii) axial attentions, which factorize full spatiotemporal self-attention along individual spatial and temporal axes to reduce computational burden while retaining expressivity. We pretrain multiple model variants on a diverse collection of heterogeneous PDE datasets and evaluate transfer to a range of downstream prediction tasks. Using both full-model fine-tuning and parameter-efficient low-rank adapters, MORPH outperforms models trained from scratch. Across extensive evaluations, MORPH matches or surpasses strong baselines and recent state-of-the-art models. Collectively, these capabilities present a flexible and powerful backbone for learning from the heterogeneous and multimodal nature of scientific observations, charting a path toward scalable and data-efficient scientific machine learning. The source code, datasets, and models are publicly available at https://github.com/lanl/MORPH.
Authors: Kevin Xu, Issei Sato
Abstract: Chain of thought (CoT) elicits reasoning in large language models by explicitly generating intermediate tokens. In contrast, latent thought reasoning operates directly in the continuous latent space, enabling computation beyond discrete linguistic representations. While both approaches exploit iterative computation, their comparative capabilities remain underexplored. In this work, we present a formal analysis showing that latent thought admits more efficient parallel computation than inherently sequential CoT. In contrast, CoT enables approximate counting and sampling through stochastic decoding. These separations suggest the tasks for which depth-driven recursion is more suitable, thereby offering practical guidance for choosing between reasoning paradigms.
Authors: Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen
Abstract: Patch-based transformers have emerged as efficient and improved long-horizon modeling architectures for time series modeling. Yet, existing approaches rely on temporally-agnostic patch construction, where arbitrary starting positions and fixed lengths fracture temporal coherence by splitting natural transitions across boundaries. This naive segmentation often disrupts short-term dependencies and weakens representation learning. We propose a novel Entropy-Guided Dynamic Patch Encoder (EntroPE), as a temporally informed framework that dynamically detects transition points via conditional entropy and dynamically places patch boundaries. This preserves temporal structure while retaining the computational benefits of patching. EntroPE consists of two key modules, namely an Entropy-based Dynamic Patcher (EDP) that applies information-theoretic criteria to locate natural temporal shifts and determine patch boundaries, and an Adaptive Patch Encoder (APE) that employs pooling and cross-attention to capture intra-patch dependencies and produce fixed-size latent representations. Extensive experiments on long-term forecasting, classification, and anomaly detection demonstrate that the proposed method improves both accuracy and efficiency, establishing entropy-guided dynamic patching as a promising new paradigm for time series modeling. Code is available at https://github.com/Sachithx/EntroPE.
Authors: Daofu Zhang, Mehrdad Pournaderi, Hanne M. Clifford, Yu Xiang, Pramod K. Varshney
Abstract: This paper studies the adversarial robustness of conformal novelty detection. In particular, we focus on two powerful learning-based frameworks that come with finite-sample false discovery rate (FDR) control: one is AdaDetect (by Marandon et al., 2024) that is based on the positive-unlabeled classifier, and the other is a one-class classifier-based approach (by Bates et al., 2023). While they provide rigorous statistical guarantees under benign conditions, their behavior under adversarial perturbations remains underexplored. We first formulate an oracle attack setup, under the AdaDetect formulation, that quantifies the worst-case degradation of FDR, deriving an upper bound that characterizes the statistical cost of attacks. This idealized formulation directly motivates a practical and effective attack scheme that only requires query access to the output labels of both frameworks. Coupling these formulations with two popular and complementary black-box adversarial algorithms, we systematically evaluate the vulnerability of both frameworks on synthetic and real-world datasets. Our results show that adversarial perturbations can significantly increase the FDR while maintaining high detection power, exposing fundamental limitations of current error-controlled novelty detection methods and motivating the development of more robust alternatives.
Authors: Shahriar Kabir Nahin, Hadi Askari, Muhao Chen, Anshuman Chhabra
Abstract: Test-Time Scaling (TTS) improves LLM reasoning by exploring multiple candidate responses and then operating over this set to find the best output. A tacit premise behind TTS is that sufficiently diverse candidate pools enhance reliability. In this work, we show that this assumption in TTS introduces a previously unrecognized failure mode. When candidate diversity is curtailed, even by a modest amount, TTS becomes much more likely to produce unsafe outputs. We present a reference-guided diversity reduction protocol (RefDiv) that serves as a diagnostic attack to stress test TTS pipelines. Through extensive experiments across open-source models (e.g. Qwen3, Mistral, Llama3.1, Gemma3) and two widely used TTS strategies (Monte Carlo Tree Search and Best-of-N), constraining diversity consistently signifies the rate at which TTS produces unsafe results. The effect is often stronger than that produced by prompts directly with high adversarial intent scores. This observed phenomenon also transfers across TTS strategies and to closed-source models (e.g. OpenAI o3-mini and Gemini-2.5-Pro), thus indicating that this is a general and extant property of TTS rather than a model-specific artifact. Additionally, we find that numerous widely used safety guardrail classifiers (e.g. Llama-Guard), are unable to flag the adversarial input prompts generated by RefDiv, demonstrating that existing defenses offer limited protection against this diversity-driven failure mode.
Authors: Stephane Hatgis-Kessell, Logan Mondal Bhamidipaty, Emma Brunskill
Abstract: Human-designed reward functions for reinforcement learning (RL) agents are frequently misaligned with the humans' true, unobservable objectives, and thus act only as proxies. Optimizing for a misspecified proxy reward function often induces reward hacking, resulting in a policy misaligned with the human's true objectives. An alternative is to perform RL from human feedback, which involves learning a reward function from scratch by collecting human preferences over pairs of trajectories. However, building such datasets is costly. To address the limitations of both approaches, we propose Preference-Based Reward Repair (PBRR): an automated iterative framework that repairs a human-specified proxy reward function by learning an additive, transition-dependent correction term from preferences. A manually specified reward function can yield policies that are highly suboptimal under the ground-truth objective, yet corrections on only a few transitions may suffice to recover optimal performance. To identify and correct for those transitions, PBRR uses a targeted exploration strategy and a new preference-learning objective. We prove in tabular domains PBRR has a cumulative regret that matches, up to constants, that of prior preference-based RL methods. In addition, on a suite of reward-hacking benchmarks, PBRR consistently outperforms baselines that learn a reward function from scratch from preferences or modify the proxy reward function using other approaches, requiring substantially fewer preferences to learn high performing policies.
Authors: Yuxiang Huang, Pengjie Wang, Jicheng Han, Weilin Zhao, Zhou Su, Ao Sun, Hongya Lyu, Hengyu Zhao, Yudong Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu
Abstract: Decoding throughput improvements from larger inference batches are limited by GPU memory, which is largely consumed by the key-value (KV) cache. Prior training-free KV cache offloading alleviates this by keeping redundant context on the CPU and fetching only a sparse subset for attention, but it often degrades long-generation quality due to training-inference mismatch on sparse patterns. Meanwhile, trainable sparse attention is incompatible with efficient offloading, as unconstrained KV accesses may force large CPU-to-GPU transfers and erase throughput gains. To this end, we propose NOSA, a trainable sparse attention mechanism natively designed for KV cache offloading. NOSA explicitly constrains the volume of CPU-GPU KV transfers, thereby achieving low communication overhead and high decoding throughput. We further build NOSI, a KV cache offloading inference system that fully unlocks NOSA's efficiency. Empirical results on 1,3,8B LLMs demonstrate that NOSA outperforms KV cache offloading baselines on general, long-input, and long-generation tasks, while boosting decoding throughput by up to 5.04x, 1.92x, and 1.83x over FullAttn, InfLLMv2, and ShadowKV, respectively. We release our code at https://github.com/thunlp/NOSA.
Authors: Danqi Liao, Chen Liu, Xingzhi Sun, Di\'e Tang, Haochen Wang, Scott Youlten, Srikar Krishna Gopinath, Haejeong Lee, Ethan C. Strayer, Antonio J. Giraldez, Smita Krishnaswamy
Abstract: Generating property-optimized mRNA sequences is central to applications such as vaccine design and protein replacement therapy, but remains challenging due to limited data, complex sequence-function relationships, and the narrow space of biologically viable sequences. Generative methods that drift away from the data manifold can yield sequences that fail to fold, translate poorly, or are otherwise nonfunctional. We present RNAGenScape, a property-guided manifold Langevin dynamics framework for mRNA sequence generation that operates directly on a learned manifold of real data. By performing iterative local optimization constrained to this manifold, RNAGenScape preserves biological viability, accesses reliable guidance, and avoids excursions into nonfunctional regions of the ambient sequence space. The framework integrates three components: (1) an autoencoder jointly trained with a property predictor to learn a property-organized latent manifold, (2) a denoising autoencoder that projects updates back onto the manifold, and (3) a property-guided Langevin dynamics procedure that performs optimization along the manifold. Across three real-world mRNA datasets spanning two orders of magnitude in size, RNAGenScape increases median property gain by up to 148% and success rate by up to 30% while ensuring biological viability of generated sequences, and achieves competitive inference efficiency relative to existing generative approaches.
Authors: Sara P\'erez-Vieites, Sahel Iqbal, Simo S\"arkk\"a, Dominik Baumann
Abstract: Bayesian experimental design (BED) provides a principled framework for optimizing data collection by choosing experiments that are maximally informative about unknown parameters. However, existing methods cannot deal with the joint challenge of (a) partially observable dynamical systems, where only noisy and incomplete observations are available, and (b) fully online inference, which updates posterior distributions and selects designs sequentially in a computationally efficient manner. Under partial observability, dynamical systems are naturally modeled as state-space models (SSMs), where latent states mediate the link between parameters and data, making the likelihood -- and thus information-theoretic objectives like the expected information gain (EIG) -- intractable. We address these challenges by deriving new estimators of the EIG and its gradient that explicitly marginalize latent states, enabling scalable stochastic optimization in nonlinear SSMs. Our approach leverages nested particle filters for efficient online state-parameter inference with convergence guarantees. Applications to realistic models, such as the susceptible-infectious-recovered (SIR) and a moving source location task, show that our framework successfully handles both partial observability and online inference.
Authors: Md. Abdul Awal, Mrigank Rochan, Chanchal K. Roy
Abstract: Transformer-based language models of code have achieved state-of-the-art performance across a wide range of software analytics tasks, but their practical deployment remains limited due to high computational costs, slow inference speeds, and significant environmental impact. To address these challenges, recent research has increasingly explored knowledge distillation as a method for compressing a large language model of code (the teacher) into a smaller model (the student) while maintaining performance. However, the degree to which a student model deeply mimics the predictive behavior and internal representations of its teacher remains largely unexplored, as current accuracy-based evaluation provides only a surface-level view of model quality and often fails to capture more profound discrepancies in behavioral fidelity between the teacher and student models. To address this gap, we empirically show that the student model often fails to deeply mimic the teacher model, resulting in up to 285% greater performance drop under adversarial attacks, which is not captured by traditional accuracy-based evaluation. Therefore, we propose MetaCompress, a metamorphic testing framework that systematically evaluates behavioral fidelity by comparing the outputs of teacher and student models under a set of behavior-preserving metamorphic relations. We evaluate MetaCompress on two widely studied tasks, using compressed versions of popular language models of code, obtained via three different knowledge distillation techniques: Compressor, AVATAR, and MORPH. The results show that MetaCompress identifies up to 62% behavioral discrepancies in student models, underscoring the need for behavioral fidelity evaluation within the knowledge distillation pipeline and establishing MetaCompress as a practical framework for testing compressed language models of code derived through knowledge distillation.
Authors: Marcelo Arenas, Pablo Barcel\'o, Luis Cofr\'e, Alexander Kozachinskiy
Abstract: Kleinberg and Mullainathan showed that language generation in the limit is always possible at the level of computability: given enough positive examples, a learner can eventually generate data indistinguishable from a target language. However, such existence results do not address feasibility. We study the sample complexity of language generation in the limit for several canonical classes of formal languages. Our results show that infeasibility already appears for context-free and regular languages, and persists even for strict subclasses such as locally threshold testable languages, as well as for incomparable classes such as non-erasing pattern languages, a well-studied class in the theory of language identification. Overall, our results establish a clear gap between the theoretical possibility of language generation in the limit and its computational feasibility.
Authors: Zhi Zheng, Yu Gu, Wei Liu, Yee Whye Teh, Wee Sun Lee
Abstract: The soft-thinking paradigm for Large Language Model (LLM) reasoning can outperform the conventional discrete-token Chain-of-Thought (CoT) reasoning in some scenarios, underscoring its research and application value. However, while the discrete-token CoT reasoning pattern can be reinforced through policy optimization algorithms such as group relative policy optimization (GRPO), extending the soft-thinking pattern with Reinforcement Learning (RL) remains challenging. This difficulty stems from the complexities of injecting stochasticity into soft-thinking tokens and updating soft-thinking policies accordingly. As a result, previous attempts to combine soft-thinking with GRPO typically underperform their discrete-token GRPO counterparts. To fully unlock the potential of soft-thinking, this paper presents a novel policy optimization algorithm, SofT-GRPO, to reinforce LLMs under the soft-thinking reasoning pattern. SofT-GRPO injects the Gumbel noise into logits, employs the Gumbel-Softmax technique to avoid soft-thinking tokens outside the pre-trained embedding space, and leverages the reparameterization trick in policy gradient. We conduct experiments across base LLMs ranging from 1.5B to 7B parameters, and results demonstrate that SofT-GRPO enables soft-thinking LLMs to slightly outperform discrete-token GRPO on Pass@1 (+0.13% on average accuracy), while exhibiting a substantial uplift on Pass@32 (+2.19% on average accuracy). Codes and weights are available on https://github.com/zz1358m/SofT-GRPO-master
Authors: Jianghan Zhu, Yaoxin Wu, Zhuoyi Lin, Zhengyuan Zhang, Haiyan Yin, Zhiguang Cao, Senthilnath Jayavelu, Xiaoli Li
Abstract: Recent advances in Neural Combinatorial Optimization (NCO) methods have significantly improved the capability of neural solvers to handle synthetic routing instances. Nonetheless, existing neural solvers typically struggle to generalize effectively from synthetic, uniformly-distributed training data to real-world VRP scenarios, including widely recognized benchmark instances from TSPLib and CVRPLib. To bridge this generalization gap, we present Evolutionary Realistic Instance Synthesis (EvoReal), which leverages an evolutionary module guided by large language models (LLMs) to generate synthetic instances characterized by diverse and realistic structural patterns. Specifically, the evolutionary module produces synthetic instances whose structural attributes statistically mimics those observed in authentic real-world instances. Subsequently, pre-trained NCO models are progressively refined, firstly aligning them with these structurally enriched synthetic distributions and then further adapting them through direct fine-tuning on actual benchmark instances. Extensive experimental evaluations demonstrate that EvoReal markedly improves the generalization capabilities of state-of-the-art neural solvers, yielding a notable reduced performance gap compared to the optimal solutions on the TSPLib (1.05%) and CVRPLib (2.71%) benchmarks across a broad spectrum of problem scales.
Authors: Loukas Sfountouris, Giannis Daras, Paris Giampouras
Abstract: Enforcing alignment between the internal representations of diffusion or flow-based generative models and those of pretrained self-supervised encoders has recently been shown to provide a powerful inductive bias, improving both convergence and sample quality. In this work, we extend this idea to inverse problems, where pretrained generative models are employed as priors. We propose applying representation alignment (REPA) between diffusion or flow-based models and a DINOv2 visual encoder, to guide the reconstruction process at inference time. Although ground-truth signals are unavailable in inverse problems, we empirically show that aligning model representations of approximate target features can substantially enhance reconstruction quality and perceptual realism. We provide theoretical results showing (a) that REPA regularization can be viewed as a variational approach for minimizing a divergence measure in the DINOv2 embedding space, and (b) how under certain regularity assumptions REPA updates steer the latent diffusion states toward those of the clean image. These results offer insights into the role of REPA in improving perceptual fidelity. Finally, we demonstrate the generality of our approach by We integrate REPA into multiple state-of-the-art inverse problem solvers, and provide extensive experiments on super-resolution, box inpainting, Gaussian deblurring, and motion deblurring confirming that our method consistently improves reconstruction quality, while also providing efficiency gains reducing the number of required discretization steps.
Authors: Jos\'e Camacho
Abstract: The validation of a data-driven model is the process of assessing the model's ability to generalize to new, unseen data in the population of interest. This paper proposes a set of general rules for model validation. These rules are designed to help practitioners create reliable validation plans and report their results transparently. While no validation scheme is flawless, these rules can help practitioners ensure their strategy is sufficient for practical use, openly discuss any limitations of their validation strategy, and report clear, comparable performance metrics.
Authors: Alla Chepurova, Aydar Bulatov, Mikhail Burtsev, Yuri Kuratov
Abstract: Knowledge graphs (KGs) provide structured, verifiable grounding for large language models (LLMs), but current LLM-based systems commonly use KGs as auxiliary structures for text retrieval, leaving their intrinsic quality underexplored. In this work, we propose Wikontic, a multi-stage pipeline that constructs KGs from open-domain text by extracting candidate triplets with qualifiers, enforcing Wikidata-based type and relation constraints, and normalizing entities to reduce duplication. The resulting KGs are compact, ontology-consistent, and well-connected; on MuSiQue, the correct answer entity appears in 96% of generated triplets. On HotpotQA, our triplets-only setup achieves 76.0 F1, and on MuSiQue 59.8 F1, matching or surpassing several retrieval-augmented generation baselines that still require textual context. In addition, Wikontic attains state-of-the-art information-retention performance on the MINE-1 benchmark (86%), outperforming prior KG construction methods. Wikontic is also efficient at build time: KG construction uses less than 1,000 output tokens, about 3$\times$ fewer than AriGraph and $<$1/20 of GraphRAG. The proposed pipeline enhances the quality of the generated KG and offers a scalable solution for leveraging structured knowledge in LLMs.
Authors: Jennifer Rosina Andersson, Zheng Zhao
Abstract: This paper is concerned with differentiable resampling in the context of sequential Monte Carlo (e.g., particle filtering). We propose a new informative resampling method that is instantly differentiable, based on an ensemble score diffusion model. We theoretically prove that our diffusion resampling method provides a consistent resampling distribution, and we show empirically that it outperforms the state-of-the-art differentiable resampling methods on multiple filtering and parameter estimation benchmarks. Finally, we show that it achieves competitive end-to-end performance when used in learning a complex dynamics-decoder model with high-dimensional image observations.
Authors: Keisuke Toyama, Zhi Zhong, Akira Takahashi, Shusuke Takahashi, Yuki Mitsufuji
Abstract: In music information retrieval (MIR) research, the use of pretrained foundational audio encoders (FAEs) has recently become a trend. FAEs pretrained on large amounts of music and audio data have been shown to improve performance on MIR tasks such as music tagging and automatic music transcription. However, their use for music structure analysis (MSA) remains underexplored: only a small subset of FAEs has been examined for MSA, and the impact of factors such as learning methods, training data, and model context length on MSA performance remains unclear. In this study, we conduct comprehensive experiments on 11 types of FAEs to investigate how these factors affect MSA performance. Our results demonstrate that FAEs using self-supervised learning with masked language modeling on music data are particularly effective for MSA. These findings pave the way for future research in FAE and MSA.
Authors: Jonas Arruda, Niels Bracher, Ullrich K\"othe, Jan Hasenauer, Stefan T. Radev
Abstract: Diffusion models have recently emerged as powerful learners for simulation-based inference (SBI), enabling fast and accurate estimation of latent parameters from simulated and real data. Their score-based formulation offers a flexible way to learn conditional or joint distributions over parameters and observations, thereby providing a versatile solution to various modeling problems. In this tutorial review, we synthesize recent developments on diffusion models for SBI, covering design choices for training, inference, and evaluation. We highlight opportunities created by various concepts such as guidance, score composition, flow matching, consistency models, and joint modeling. Furthermore, we discuss how efficiency and statistical accuracy are affected by noise schedules, parameterizations, and samplers. Finally, we illustrate these concepts with case studies across parameter dimensionalities, simulation budgets, and model types, and outline open questions for future research.
Authors: Enoch Hyunwook Kang
Abstract: Field experiments (A/B tests) are often the most credible benchmark for methods (algorithms) in societal systems, but their cost and latency bottleneck rapid methodological progress. LLM-based persona simulation offers a cheap synthetic alternative, yet it is unclear whether replacing humans with personas preserves the benchmark interface that adaptive methods optimize against. We prove an if-and-only-if characterization: when (i) methods observe only the aggregate outcome (aggregate-only observation) and (ii) evaluation depends only on the submitted artifact and not on the method's identity or provenance (method-blind evaluation), swapping humans for personas is just panel change from the method's point of view, indistinguishable from changing the evaluation population (e.g., New York to Jakarta). Furthermore, we move from validity to usefulness: we define an information-theoretic discriminability of the induced aggregate channel and show that making persona benchmarking as decision-relevant as a field experiment is fundamentally a sample-size question, yielding explicit bounds on the number of independent persona evaluations required to reliably distinguish meaningfully different methods at a chosen resolution.
Authors: Jun Wang
Abstract: We present a theoretical study of continual and experiential learning in large language model agents that combine episodic memory with reinforcement learning. We argue that the key mechanism for continual adaptation, without updating model parameters, is reflection: the agent's ability to use past experience to guide future actions. Empirical findings suggest that episodic, experience-driven reflection enables generalised adaptation across a wide range of open-ended, long-horizon tasks. This indicates that efficient learning can occur during deployment and weakens the traditional separation between training and testing. Motivated by this, we introduce the Stateful Reflective Decision Process, a formal model of reflective memory dynamics. In this abstraction, an agent maintains an episodic memory and performs two core operations. Writing stores interaction outcomes and plays the role of policy evaluation. Reading retrieves relevant past cases to inform decisions and plays the role of policy improvement. This perspective treats reflective memory as a control object that can be analysed using classical reinforcement learning tools. We then develop a read-write reflective learning framework by integrating retrieval into soft policy iteration and establish convergence guarantees. We show that as memory grows and provides denser coverage of the state space, the resulting composite policy converges to the optimal solution. Overall, this framework connects practical memory-based methods with principled reinforcement learning, providing a rigorous mathematical basis for building reflective, memory-embedded agents capable of continual general-purpose learning.
Authors: Svenja Jedhoff, Elizaveta Semenova, Aura Raulo, Anne Meyer, Paul-Christian B\"urkner
Abstract: Graphs arise across diverse domains, from biology and chemistry to social and information networks, as well as in transportation and logistics. Inference on graph-structured data requires methods that are permutation-invariant, scalable across varying sizes and sparsities, and capable of capturing complex long-range dependencies, making posterior estimation on graph parameters particularly challenging. Amortized Bayesian Inference (ABI) is a simulation-based framework that employs generative neural networks to enable fast, likelihood-free posterior inference. We adapt ABI to graph data to address these challenges to perform inference on node-, edge-, and graph-level parameters. Our approach couples permutation-invariant graph encoders with flexible neural posterior estimators in a two-module pipeline: a summary network maps attributed graphs to fixed-length representations, and an inference network approximates the posterior over parameters. In this setting, several neural architectures can serve as the summary network. In this work we evaluate multiple architectures and assess their performance on controlled synthetic settings and two real-world domains - biology and logistics - in terms of recovery and calibration.
Authors: Ahmad Mohammad Saber, Saeed Jafari, Zhengmao Ouyang, Paul Budnarain, Amr Youssef, Deepa Kundur
Abstract: This paper presents a large language model (LLM)-based framework that adapts and fine-tunes compact LLMs for detecting cyberattacks on transformer current differential relays (TCDRs), which can otherwise cause false tripping of critical power transformers. The core idea is to textualize multivariate time-series current measurements from TCDRs, across phases and input/output sides, into structured natural-language prompts that are then processed by compact, locally deployable LLMs. Using this representation, we fine-tune DistilBERT, GPT-2, and DistilBERT+LoRA to distinguish cyberattacks from genuine fault-induced disturbances while preserving relay dependability. The proposed framework is evaluated against a broad set of state-of-the-art machine learning and deep learning baselines under nominal conditions, complex cyberattack scenarios, and measurement noise. Our results show that LLM-based detectors achieve competitive or superior cyberattack detection performance, with DistilBERT detecting up to 97.62% of attacks while maintaining perfect fault detection accuracy. Additional evaluations demonstrate robustness to prompt formulation variations, resilience under combined time-synchronization and false-data injection attacks, and stable performance under realistic measurement noise levels. The attention mechanisms of LLMs further enable intrinsic interpretability by highlighting the most influential time-phase regions of relay measurements. These results demonstrate that compact LLMs provide a practical, interpretable, and robust solution for enhancing cyberattack detection in modern digital substations. We provide the full dataset used in this study for reproducibility.
Authors: Josh Li, Fow-sen Choa
Abstract: Feedback-rich neural architectures can regenerate earlier representations and inject temporal context, making them a natural setting for strictly local synaptic plasticity. Existing literature raises doubt about whether a minimal, backpropagation-free feedback-Hebbian system can already express interpretable continual-learning-relevant behaviors under controlled training schedules. In this work, we introduce a compact prediction-reconstruction architecture with a dedicated feedback pathway providing lightweight, locally trainable temporal context for continual adaptation. All synapses are updated by a unified local rule combining centered Hebbian covariance, Oja-style stabilization, and a local supervised drive where targets are available. With a simple two-pair association task, learning is characterized through layer-wise activity snapshots, connectivity trajectories (row and column means of learned weights), and a normalized retention index across phases. Under sequential A to B training, forward output connectivity exhibits a long-term depression (LTD)-like suppression of the earlier association, while feedback connectivity preserves an A-related trace during acquisition of B. Under an alternating sequence, both associations are concurrently maintained rather than sequentially suppressed. Architectural controls and rule-term ablations isolate the role of dedicated feedback in regeneration and co-maintenance, alongside the role of the local supervised term in output selectivity and unlearning. Together, the results show that a compact feedback pathway trained with local plasticity can support regeneration and continual-learning-relevant dynamics in a minimal, mechanistically transparent setting.
Authors: Teddy Koker, Abhijeet Gangan, Mit Kotak, Jaime Marian, Tess Smidt
Abstract: Many materials properties depend on higher-order derivatives of the potential energy surface, yet machine learned interatomic potentials (MLIPs) trained with a standard loss on energy, force, and stress errors can exhibit error in curvature, degrading the prediction of vibrational properties. We introduce phonon fine-tuning (PFT), which directly supervises second-order force constants of materials by matching MLIP energy Hessians to DFT-computed force constants from finite displacement phonon calculations. To scale to large supercells, PFT stochastically samples Hessian columns and computes the loss with a single Hessian-vector product. We also use a simple co-training scheme to incorporate upstream data to mitigate catastrophic forgetting. On the MDR Phonon benchmark, PFT improves Nequix MP by 55% on average across phonon thermodynamic properties and achieves state-of-the-art accuracy among models trained on Materials Project trajectories. PFT also generalizes to improve properties beyond second-derivatives, improving thermal conductivity predictions that rely on third-order derivatives of the potential energy.
Authors: Sharan Sahu, Cameron J. Hogan, Martin T. Wells
Abstract: In this paper, we provide a comprehensive theoretical analysis of Stochastic Gradient Descent (SGD) and its momentum variants (Polyak Heavy-Ball and Nesterov) for tracking time-varying optima under strong convexity and smoothness. Our finite-time bounds reveal a sharp decomposition of tracking error into transient, noise-induced, and drift-induced components. This decomposition exposes a fundamental trade-off: while momentum is often used as a gradient-smoothing heuristic, under distribution shift it incurs an explicit drift-amplification penalty that diverges as the momentum parameter $\beta$ approaches 1, yielding systematic tracking lag. We complement these upper bounds with minimax lower bounds under gradient-variation constraints, proving this momentum-induced tracking penalty is not an analytical artifact but an information-theoretic barrier: in drift-dominated regimes, momentum is unavoidably worse because stale-gradient averaging forces systematic lag. Our results provide theoretical grounding for the empirical instability of momentum in dynamic settings and precisely delineate regime boundaries where vanilla SGD provably outperforms its accelerated counterparts.
Authors: Miao Li, Hanyang Jiang, Sikai Cheng, Hengyu Fu, Yuhang Cai, Baihe Huang, Tinghan Ye, Xuanzhou Chen, Pascal Van Hentenryck
Abstract: Diffusion Language Models (DLMs) present a promising non-sequential paradigm for text generation, distinct from standard autoregressive (AR) approaches. However, current decoding strategies often adopt a reactive stance, underutilizing the global bidirectional context to dictate global trajectories. To address this, we propose Plan-Verify-Fill (PVF), a training-free paradigm that grounds planning via quantitative validation. PVF actively constructs a hierarchical skeleton by prioritizing high-leverage semantic anchors and employs a verification protocol to operationalize pragmatic structural stopping where further deliberation yields diminishing returns. Extensive evaluations on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate that PVF reduces the Number of Function Evaluations (NFE) by up to 65% compared to confidence-based parallel decoding across benchmark datasets, unlocking superior efficiency without compromising accuracy.
Authors: Pietro Barbiero, Mateo Espinosa Zarlenga, Francesco Giannini, Alberto Termine, Filippo Bonchi, Mateja Jamnik, Giuseppe Marra
Abstract: This paper argues that interpretability research in Artificial Intelligence (AI) is fundamentally ill-posed as existing definitions of interpretability fail to describe how interpretability can be formally tested or designed for. We posit that actionable definitions of interpretability must be formulated in terms of *symmetries* that inform model design and lead to testable conditions. Under a probabilistic view, we hypothesise that four symmetries (inference equivariance, information invariance, concept-closure invariance, and structural invariance) suffice to (i) formalise interpretable models as a subclass of probabilistic models, (ii) yield a unified formulation of interpretable inference (e.g., alignment, interventions, and counterfactuals) as a form of Bayesian inversion, and (iii) provide a formal framework to verify compliance with safety standards and regulations.
Authors: Yimeng Min, Carla P. Gomes
Abstract: We demonstrate that a single training trajectory can transform a graph neural network into an unsupervised heuristic for combinatorial optimization. Focusing on the Travelling Salesman Problem, we show that encoding global structural constraints as an inductive bias enables a non-autoregressive model to generate solutions via direct forward passes, without search, supervision, or sequential decision-making. At inference time, dropout and snapshot ensembling allow a single model to act as an implicit ensemble, reducing optimality gaps through increased solution diversity. Our results establish that graph neural networks do not require supervised training nor explicit search to be effective. Instead, they can internalize global combinatorial structure and function as strong, learned heuristics. This reframes the role of learning in combinatorial optimization: from augmenting classical algorithms to directly instantiating new heuristics.
Authors: Fabian Bleile, Sarah Lumpp, Mathias Drton
Abstract: Learning a stationary diffusion amounts to estimating the parameters of a stochastic differential equation whose stationary distribution matches a target distribution. We build on the recently introduced kernel deviation from stationarity (KDS), which enforces stationarity by evaluating expectations of the diffusion's generator in a reproducing kernel Hilbert space. Leveraging the connection between KDS and Stein discrepancies, we introduce the Stein-type KDS (SKDS) as an alternative formulation. We prove that a vanishing SKDS guarantees alignment of the learned diffusion's stationary distribution with the target. Furthermore, under broad parametrizations, SKDS is convex with an empirical version that is $\epsilon$-quasiconvex with high probability. Empirically, learning with SKDS attains comparable accuracy to KDS while substantially reducing computational cost and yields improvements over the majority of competitive baselines.
Authors: Thanawat Sornwanee
Abstract: We show how the differentiability method employed in the paper ``Differentiable Integer Linear Programming'', Geng, et al., 2025 as shown in its theorem 5 is incorrect. Moreover, there already exists some downstream work that inherits the same error. The underlying reason comes from that, though being continuous in expectation, the surrogate loss is discontinuous in almost every realization of the randomness, for the stochastic gradient descent.
Authors: Saurabh Jha, Rohan Arora, Bhavya, Noah Zheutlin, Paulina Toro Isaza, Laura Shwartz, Yu Deng, Daby Sow, Ruchi Mahindru, Ruchir Puri
Abstract: LLM agents excel when environments are mostly static and the needed information fits in a model's context window, but they often fail in open-ended investigations where explanations must be constructed by iteratively mining evidence from massive, heterogeneous operational data. These investigations exhibit hidden dependency structure: entities interact, signals co-vary, and the importance of a fact may only become clear after other evidence is discovered. Because the context window is bounded, agents must summarize intermediate findings before their significance is known, increasing the risk of discarding key evidence. ReAct-style agents are especially brittle in this regime. Their retrieve-summarize-reason loop makes conclusions sensitive to exploration order and introduces run-to-run non-determinism, producing a reliability gap where Pass-at-k may be high but Majority-at-k remains low. Simply sampling more rollouts or generating longer reasoning traces does not reliably stabilize results, since hypotheses cannot be autonomously checked as new evidence arrives and there is no explicit mechanism for belief bookkeeping and revision. In addition, ReAct entangles semantic reasoning with controller duties such as tool orchestration and state tracking, so execution errors and plan drift degrade reasoning while consuming scarce context. We address these issues by formulating investigation as abductive reasoning over a dependency graph and proposing EoG (Explanations over Graphs), a disaggregated framework in which an LLM performs bounded local evidence mining and labeling (cause vs symptom) while a deterministic controller manages traversal, state, and belief propagation to compute a minimal explanatory frontier. On a representative ITBench diagnostics task, EoG improves both accuracy and run-to-run consistency over ReAct baselines, including a 7x average gain in Majority-at-k entity F1.
Authors: Bohan Hou, Hongyi Jin, Guanjie Wang, Jinqi Chen, Yaxing Cai, Lijie Yang, Zihao Ye, Yaoyao Ding, Ruihang Lai, Tianqi Chen
Abstract: Scaling modern deep learning workloads demands coordinated placement of data and compute across device meshes, memory hierarchies, and heterogeneous accelerators. We present Axe Layout, a hardware-aware abstraction that maps logical tensor coordinates to a multi-axis physical space via named axes. Axe unifies tiling, sharding, replication, and offsets across inter-device distribution and on-device layouts, enabling collective primitives to be expressed consistently from device meshes to threads. Building on Axe, we design a multi-granularity, distribution-aware DSL and compiler that composes thread-local control with collective operators in a single kernel. Experiments show that our unified approach can bring performance close to hand-tuned kernels on across latest GPU devices and multi-device environments and accelerator backends.
Authors: Guilin Zhang, Kai Zhao, Jeffrey Friedman, Xu Chu
Abstract: Large Language Models (LLMs) are increasingly used to generate natural-language explanations in recommender systems, acting as explanation agents that reason over user behavior histories. While prior work has focused on explanation fluency and relevance under fixed inputs, the robustness of LLM-generated explanations to realistic user behavior noise remains largely unexplored. In real-world web platforms, interaction histories are inherently noisy due to accidental clicks, temporal inconsistencies, missing values, and evolving preferences, raising concerns about explanation stability and user trust. We present RobustExplain, the first systematic evaluation framework for measuring the robustness of LLM-generated recommendation explanations. RobustExplain introduces five realistic user behavior perturbations evaluated across multiple severity levels and a multi-dimensional robustness metric capturing semantic, keyword, structural, and length consistency. Our goal is to establish a principled, task-level evaluation framework and initial robustness baselines, rather than to provide a comprehensive leaderboard across all available LLMs. Experiments on four representative LLMs (7B--70B) show that current models exhibit only moderate robustness, with larger models achieving up to 8% higher stability. Our results establish the first robustness benchmarks for explanation agents and highlight robustness as a critical dimension for trustworthy, agent-driven recommender systems at web scale.
Authors: Guilin Zhang, Kai Zhao, Jeffrey Friedman, Xu Chu
Abstract: Recommendation systems must optimize multiple objectives while satisfying hard business constraints such as fairness and coverage. For example, an e-commerce platform may require every recommendation list to include items from multiple sellers and at least one newly listed product; violating such constraints--even once--is unacceptable in production. Prior work on multi-objective recommendation and recent LLM-based recommender agents largely treat constraints as soft penalties or focus on item scoring and interaction, leading to frequent violations in real-world deployments. How to leverage LLMs for coordinating constrained optimization in recommendation systems remains underexplored. We propose DualAgent-Rec, an LLM-coordinated dual-agent framework for constrained multi-objective e-commerce recommendation. The framework separates optimization into an Exploitation Agent that prioritizes accuracy under hard constraints and an Exploration Agent that promotes diversity through unconstrained Pareto search. An LLM-based coordinator adaptively allocates resources between agents based on optimization progress and constraint satisfaction, while an adaptive epsilon-relaxation mechanism guarantees feasibility of final solutions. Experiments on the Amazon Reviews 2023 dataset demonstrate that DualAgent-Rec achieves 100% constraint satisfaction and improves Pareto hypervolume by 4-6% over strong baselines, while maintaining competitive accuracy-diversity trade-offs. These results indicate that LLMs can act as effective orchestration agents for deployable and constraint-compliant recommendation systems.
Authors: Ankang Zhang, Ming Chi, Xiaoling Wang, Lintao Ye
Abstract: Stabilizing a dynamical system is a fundamental problem that serves as a cornerstone for many complex tasks in the field of control systems. The problem becomes challenging when the system model is unknown. Among the Reinforcement Learning (RL) algorithms that have been successfully applied to solve problems pertaining to unknown linear dynamical systems, the policy gradient (PG) method stands out due to its ease of implementation and can solve the problem in a model-free manner. However, most of the existing works on PG methods for unknown linear dynamical systems assume full-state feedback. In this paper, we take a step towards model-free learning for partially observable linear dynamical systems with output feedback and focus on the fundamental stabilization problem of the system. We propose an algorithmic framework that stretches the boundary of PG methods to the problem without global convergence guarantees. We show that by leveraging zeroth-order PG update based on system trajectories and its convergence to stationary points, the proposed algorithms return a stabilizing output feedback policy for discrete-time linear dynamical systems. We also explicitly characterize the sample complexity of our algorithm and verify the effectiveness of the algorithm using numerical examples.
Authors: Shuntaro Nagashima, Hideaki Iiduka
Abstract: The Muon optimizer has recently attracted attention due to its orthogonalized first-order updates, and a deeper theoretical understanding of its convergence behavior is essential for guiding practical applications; however, existing convergence guarantees are either coarse or obtained under restrictive analytical settings. In this work, we establish sharper convergence guarantees for the Muon optimizer through a direct and simplified analysis that does not rely on restrictive assumptions on the update rule. Our results improve upon existing bounds by achieving faster convergence rates while covering a broader class of problem settings. These findings provide a more accurate theoretical characterization of Muon and offer insights applicable to a broader class of orthogonalized first-order methods.
Authors: Shir Rozenfeld, Rahul Pankajakshan, Itay Zloczower, Eyal Lenga, Gilad Gressel, Yisroel Mirsky
Abstract: Large language models (LLMs) are increasingly paired with activation-based monitoring to detect and prevent harmful behaviors that may not be apparent at the surface-text level. However, existing activation safety approaches, trained on broad misuse datasets, struggle with poor precision, limited flexibility, and lack of interpretability. This paper introduces a new paradigm: rule-based activation safety, inspired by rule-sharing practices in cybersecurity. We propose modeling activations as cognitive elements (CEs), fine-grained, interpretable factors such as ''making a threat'' and ''payment processing'', that can be composed to capture nuanced, domain-specific behaviors with higher precision. Building on this representation, we present a practical framework that defines predicate rules over CEs and detects violations in real time. This enables practitioners to configure and update safeguards without retraining models or detectors, while supporting transparency and auditability. Our results show that compositional rule-based activation safety improves precision, supports domain customization, and lays the groundwork for scalable, interpretable, and auditable AI governance. We will release GAVEL as an open-source framework and provide an accompanying automated rule creation tool.
Authors: Nikhil Raghav, Avisek Gupta, Swagatam Das, Md Sahidullah
Abstract: Speaker diarization aims to segment audio recordings into regions corresponding to individual speakers. Although unsupervised speaker diarization is inherently challenging, the prospect of identifying speaker regions without pretraining or weak supervision motivates research on clustering techniques. In this work, we share the notable observation that measuring multiple kernel similarities of speaker embeddings to thereafter craft a sparse graph for spectral clustering in a principled manner is sufficient to achieve state-of-the-art performances in a fully unsupervised setting. Specifically, we consider four polynomial kernels and a degree one arccosine kernel to measure similarities in speaker embeddings, using which sparse graphs are constructed in a principled manner to emphasize local similarities. Experiments show the proposed approach excels in unsupervised speaker diarization over a variety of challenging environments in the DIHARD-III, AMI, and VoxConverse corpora. To encourage further research, our implementations are available at https://github.com/nikhilraghav29/MK-SGC-SC.
Authors: Skyler Wu, Yash Nair, Emmanuel J. Cand\`es
Abstract: Exhaustively evaluating many large language models (LLMs) on a large suite of benchmarks is expensive. We cast benchmarking as finite-population inference and, under a fixed query budget, seek tight confidence intervals (CIs) for model accuracy with valid frequentist coverage. We propose Factorized Active Querying (FAQ), which (a) leverages historical information through a Bayesian factor model; (b) adaptively selects questions using a hybrid variance-reduction/active-learning sampling policy; and (c) maintains validity through Proactive Active Inference -- a finite-population extension of active inference (Zrnic & Cand\`es, 2024) that enables direct question selection while preserving coverage. With negligible overhead cost, FAQ delivers up to $5\times$ effective sample size gains over strong baselines on two benchmark suites, across varying historical-data missingness levels: this means that it matches the CI width of uniform sampling while using up to $5\times$ fewer queries. We release our source code and our curated datasets to support reproducible evaluation and future research.
Authors: Gaku Omiya, Pierre-Louis Poirion, Akiko Takeda
Abstract: Randomized subspace methods reduce per-iteration cost; however, in nonconvex optimization, most analyses are expectation-based, and high-probability bounds remain scarce even under sub-Gaussian noise. We first prove that randomized subspace SGD (RS-SGD) admits a high-probability convergence bound under sub-Gaussian noise, achieving the same order of oracle complexity as prior in-expectation results. Motivated by the prevalence of heavy-tailed gradients in modern machine learning, we then propose randomized subspace normalized SGD (RS-NSGD), which integrates direction normalization into subspace updates. Assuming the noise has bounded $p$-th moments, we establish both in-expectation and high-probability convergence guarantees, and show that RS-NSGD can achieve better oracle complexity than full-dimensional normalized SGD.
Authors: Yichi Zhang, Fengqing Zhu
Abstract: Training learned image compression (LIC) models entails navigating a challenging optimization landscape defined by the fundamental trade-off between rate and distortion. Standard first-order optimizers, such as SGD and Adam, struggle with \emph{gradient conflicts} arising from competing objectives, leading to slow convergence and suboptimal rate-distortion performance. In this work, we demonstrate that a simple utilization of a second-order quasi-Newton optimizer, \textbf{SOAP}, dramatically improves both training efficiency and final performance across diverse LICs. Our theoretical and empirical analyses reveal that Newton preconditioning inherently resolves the intra-step and inter-step update conflicts intrinsic to the R-D objective, facilitating faster, more stable convergence. Beyond acceleration, we uncover a critical deployability benefit: second-order trained models exhibit significantly fewer activation and latent outliers. This substantially enhances robustness to post-training quantization. Together, these results establish second-order optimization, achievable as a seamless drop-in replacement of the imported optimizer, as a powerful, practical tool for advancing the efficiency and real-world readiness of LICs.