Authors: Farjana Yesmin, Nusrat Shirmin
Abstract: The growing complexity of machine learning (ML) models in big data analytics, especially in domains such as environmental monitoring, highlights the critical need for interpretability and explainability to promote trust, ethical considerations, and regulatory adherence (e.g., GDPR). Traditional "black-box" models obstruct transparency, whereas post-hoc explainable AI (XAI) techniques like LIME and SHAP frequently compromise accuracy or fail to deliver inherent insights. This paper presents a novel framework that combines type-2 fuzzy sets, granular computing, and clustering to boost explainability and fairness in big data environments. When applied to the UCI Air Quality dataset, the framework effectively manages uncertainty in noisy sensor data, produces linguistic rules, and assesses fairness using silhouette scores and entropy. Key contributions encompass: (1) A type-2 fuzzy clustering approach that enhances cohesion by about 4% compared to type-1 methods (silhouette 0.365 vs. 0.349) and improves fairness (entropy 0.918); (2) Incorporation of fairness measures to mitigate biases in unsupervised scenarios; (3) A rule-based component for intrinsic XAI, achieving an average coverage of 0.65; (4) Scalable assessments showing linear runtime (roughly 0.005 seconds for sampled big data sizes). Experimental outcomes reveal superior performance relative to baselines such as DBSCAN and Agglomerative Clustering in terms of interpretability, fairness, and efficiency. Notably, the proposed method achieves a 4% improvement in silhouette score over type-1 fuzzy clustering and outperforms baselines in fairness (entropy reduction by up to 1%) and efficiency.
Authors: Armin Gerami, Ramani Duraiswami
Abstract: Transformer models have become increasingly popular in financial applications, yet their potential risk making and biases remain under-explored. The purpose of this work is to audit the reliance of the model on volatile data for decision-making, and quantify how the frequency of price movements affects the model's prediction confidence. We employ a transformer model for prediction, and introduce a metric based on Partial Information Decomposition (PID) to measure the influence of each asset on the model's decision making. Our analysis reveals two key observations: first, the model disregards data volatility entirely, and second, it is biased toward data with lower-frequency price movements.
Authors: Abrar Shahid, Ibteeker Mahir Ishum, AKM Tahmidul Haque, M Sohel Rahman, A. B. M. Alim Al Islam
Abstract: This paper presents a controlled study of adversarial reinforcement learning in network security through a custom OpenAI Gym environment that models brute-force attacks and reactive defenses on multi-port services. The environment captures realistic security trade-offs including background traffic noise, progressive exploitation mechanics, IP-based evasion tactics, honeypot traps, and multi-level rate-limiting defenses. Competing attacker and defender agents are trained using Deep Q-Networks (DQN) within a zero-sum reward framework, where successful exploits yield large terminal rewards while incremental actions incur small costs. Through systematic evaluation across multiple configurations (varying trap detection probabilities, exploitation difficulty thresholds, and training regimens), the results demonstrate that defender observability and trap effectiveness create substantial barriers to successful attacks. The experiments reveal that reward shaping and careful training scheduling are critical for learning stability in this adversarial setting. The defender consistently maintains strategic advantage across 50,000+ training episodes, with performance gains amplifying when exposed to complex defensive strategies including adaptive IP blocking and port-specific controls. Complete implementation details, reproducible hyperparameter configurations, and architectural guidelines are provided to support future research in adversarial RL for cybersecurity. The zero-sum formulation and realistic operational constraints make this environment suitable for studying autonomous defense systems, attacker-defender co-evolution, and transfer learning to real-world network security scenarios.
Authors: Muhammad Arif Hakimi Zamrai
Abstract: Inverse design, which seeks to find optimal parameters for a target output, is a central challenge in engineering. Surrogate-based optimization (SBO) has become a standard approach, yet it is fundamentally structured to converge to a single-point solution, thereby limiting design space exploration and ignoring potentially valuable alternative topologies. This paper presents a paradigm shift from single-point optimization to generative inverse design. We introduce a framework based on a Conditional Variational Autoencoder (CVAE) that learns a probabilistic mapping between a system's design parameters and its performance, enabling the generation of a diverse portfolio of high-performing candidates conditioned on a specific performance objective. We apply this methodology to the complex, non-linear problem of minimizing airfoil self-noise, using a high-performing SBO method from a prior benchmark study as a rigorous baseline. The CVAE framework successfully generated 256 novel designs with a 94.1\% validity rate. A subsequent surrogate-based evaluation revealed that 77.2\% of these valid designs achieved superior performance compared to the single optimal design found by the SBO baseline. This work demonstrates that the generative approach not only discovers higher-quality solutions but also provides a rich portfolio of diverse candidates, fundamentally enhancing the engineering design process by enabling multi-criteria decision-making.
Authors: Md Zahin Hossain George, Md Khorshed Alam, Md Tarek Hasan
Abstract: This systematic literature review examines the role of machine learning in fraud detection within digital banking, synthesizing evidence from 118 peer-reviewed studies and institutional reports. Following the PRISMA guidelines, the review applied a structured identification, screening, eligibility, and inclusion process to ensure methodological rigor and transparency. The findings reveal that supervised learning methods, such as decision trees, logistic regression, and support vector machines, remain the dominant paradigm due to their interpretability and established performance, while unsupervised anomaly detection approaches are increasingly adopted to address novel fraud patterns in highly imbalanced datasets. Deep learning architectures, particularly recurrent and convolutional neural networks, have emerged as transformative tools capable of modeling sequential transaction data and detecting complex fraud typologies, though challenges of interpretability and real-time deployment persist. Hybrid models that combine supervised, unsupervised, and deep learning strategies demonstrate superior adaptability and detection accuracy, highlighting their potential as convergent solutions.
Authors: Eric Jahns, Davi Moreno, Milan Stojkov, Michel A. Kinsy
Abstract: Spiking Neural Networks (SNNs) have emerged as energy-efficient alternatives to traditional artificial neural networks, leveraging asynchronous and biologically inspired neuron dynamics. Among existing neuron models, the Leaky Integrate-and-Fire (LIF) neuron has become widely adopted in deep SNNs due to its simplicity and computational efficiency. However, this efficiency comes at the expense of expressiveness, as LIF dynamics are constrained to linear decay at each timestep. In contrast, more complex models, such as the Quadratic Integrate-and-Fire (QIF) neuron, exhibit richer, nonlinear dynamics but have seen limited adoption due to their training instability. On that note, we propose the first discretization of the QIF neuron model tailored for high-performance deep spiking neural networks and provide an in-depth analysis of its dynamics. To ensure training stability, we derive an analytical formulation for surrogate gradient windows directly from our discretizations' parameter set, minimizing gradient mismatch. We evaluate our method on CIFAR-10, CIFAR-100, ImageNet, and CIFAR-10 DVS, demonstrating its ability to outperform state-of-the-art LIF-based methods. These results establish our discretization of the QIF neuron as a compelling alternative to LIF neurons for deep SNNs, combining richer dynamics with practical scalability.
Authors: Haijin Xie, Gongquan Zhang
Abstract: New quality productive forces (NQPF), digital economy advancement, and artificial intelligence (AI) technologies are becoming crucial for promoting sustainable urban development. This study proposes a Multi-head Attention Deep & Cross Network (MADCN) framework, combining feature interaction modeling and attention mechanisms, to predict urban carbon emissions and investigate the impacts of technological factors. The framework incorporates an interpretable learning phase using SHapley Additive exPlanations (SHAP) to assess the contributions of different features. A panel dataset covering 275 Chinese cities is utilized to test the MADCN model. Experimental results demonstrate that the MADCN model achieves superior predictive performance compared to traditional machine learning and deep learning baselines, with a Mean Squared Error (MSE) of 406,151.063, a Mean Absolute Error (MAE) of 612.304, and an R-squared value of 0.991 on the test set. SHAP analysis highlights that population, city size, urbanization rate, and GDP are among the most influential factors on carbon emissions, while NQPF, digital economy index, and AI technology level also show meaningful but relatively moderate effects. Advancing NQPF, strengthening the digital economy, and accelerating AI technology development can significantly contribute to reducing urban carbon emissions. Policymakers should prioritize integrating technological innovation into carbon reduction strategies, particularly by promoting intelligent infrastructure and enhancing digitalization across sectors, to effectively achieve dual-carbon goals.
Authors: Anushiya Arunan, Yan Qin, Xiaoli Li, U-Xuan Tan, H. Vincent Poor, Chau Yuen
Abstract: Accurate battery capacity estimation is key to alleviating consumer concerns about battery performance and reliability of electric vehicles (EVs). However, practical data limitations imposed by stringent privacy regulations and labeled data shortages hamper the development of generalizable capacity estimation models that remain robust to real-world data distribution shifts. While self-supervised learning can leverage unlabeled data, existing techniques are not particularly designed to learn effectively from challenging field data -- let alone from privacy-friendly data, which are often less feature-rich and noisier. In this work, we propose a first-of-its-kind capacity estimation model based on self-supervised pre-training, developed on a large-scale dataset of privacy-friendly charging data snippets from real-world EV operations. Our pre-training framework, snippet similarity-weighted masked input reconstruction, is designed to learn rich, generalizable representations even from less feature-rich and fragmented privacy-friendly data. Our key innovation lies in harnessing contrastive learning to first capture high-level similarities among fragmented snippets that otherwise lack meaningful context. With our snippet-wise contrastive learning and subsequent similarity-weighted masked reconstruction, we are able to learn rich representations of both granular charging patterns within individual snippets and high-level associative relationships across different snippets. Bolstered by this rich representation learning, our model consistently outperforms state-of-the-art baselines, achieving 31.9% lower test error than the best-performing benchmark, even under challenging domain-shifted settings affected by both manufacturer and age-induced distribution shifts.
Authors: Dmitry Rybin, Yushun Zhang, Ding Tian, Zhihang Lin, Ruoyu Sun, Zhi-Quan Luo
Abstract: We present Fast Causal Attention (FCA), an algorithm that computes exact Causal Attention using 10\% fewer operations. FCA accelerates a special class of matrix multiplications where either one operand or the output matrix is upper- or lower-triangular. This includes all operations in forward and backward pass of Causal Attention, such as masked product $\mathrm{Mask}(QK^{T})$. For these matrix multiplications on GPU, FCA reaches noticeable accelerations over the default PyTorch implementations and Triton compiled kernels. FCA is built upon algebraic identities discovered via machine learning and combinatorial search.
Authors: Ji Zhang, Yiwei Li, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Jiayi Shi, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Yao Hu, Kan Li
Abstract: KV cache in autoregressive LLMs eliminates redundant recomputation but has emerged as the dominant memory and bandwidth bottleneck during inference, notably with long contexts and test-time scaling. KV quantization is a key lever for reducing cache cost, but accuracy drops sharply as the native KV distribution lacks flatness and thus maintains a wide quantization range. Prior work focuses on isolating outliers, which caps their error but fails to flatten the overall distribution, leaving performance fragile under low-bit settings. In this work, we show that the K cache maintains a stable structure that evolves gradually with context, while the V cache carries latent semantic regularities. Building on these insights, we propose PatternKV, a pattern-aligned residual quantization scheme. It mines representative pattern vectors online, aligns each KV vector to its nearest pattern, and quantizes only the residual. This reshaping of the KV distribution flattens the quantization target and narrows its range, thereby improving the fidelity of low-bit KV quantization. Across long-context and test-time scaling settings on multiple backbones, PatternKV delivers consistent 2-bit gains, with a 0.08% average 4-bit drop relative to FP16, improves test-time scaling accuracy by 10% on average, and raises throughput by 1.4x while supporting 1.25x larger batches.
Authors: Ou Deng, Ruichen Cong, Jianting Xu, Shoji Nishimura, Atsushi Ogihara, Qun Jin
Abstract: Symbolic regression promises readable equations but struggles to encode unit-aware thresholds and conditional logic. We propose logistic-gated operators (LGO) -- differentiable gates with learnable location and steepness -- embedded as typed primitives and mapped back to physical units for audit. Across two primary health datasets (ICU, NHANES), the hard-gate variant recovers clinically plausible cut-points: 71% (5/7) of assessed thresholds fall within 10% of guideline anchors and 100% within 20%, while using far fewer gates than the soft variant (ICU median 4.0 vs 10.0; NHANES 5.0 vs 12.5), and remaining within the competitive accuracy envelope of strong SR baselines. On predominantly smooth tasks, gates are pruned, preserving parsimony. The result is compact symbolic equations with explicit, unit-aware thresholds that can be audited against clinical anchors -- turning interpretability from a post-hoc explanation into a modeling constraint and equipping symbolic regression with a practical calculus for regime switching and governance-ready deployment.
Authors: Saida Elouardi, Mohammed Jouhari, Anas Motii
Abstract: In critical IoT environments, such as smart homes and industrial systems, effective Intrusion Detection Systems (IDS) are essential for ensuring security. However, developing robust IDS solutions remains a significant challenge. Traditional machine learning-based IDS models typically require large datasets, but data sharing is often limited due to privacy and security concerns. Federated Learning (FL) presents a promising alternative by enabling collaborative model training without sharing raw data. Despite its advantages, FL still faces key challenges, such as data heterogeneity (non-IID data) and high energy and computation costs, particularly for resource constrained IoT devices. To address these issues, this paper proposes OptiFLIDS, a novel approach that applies pruning techniques during local training to reduce model complexity and energy consumption. It also incorporates a customized aggregation method to better handle pruned models that differ due to non-IID data distributions. Experiments conducted on three recent IoT IDS datasets, TON_IoT, X-IIoTID, and IDSIoT2024, demonstrate that OptiFLIDS maintains strong detection performance while improving energy efficiency, making it well-suited for deployment in real-world IoT environments.
Authors: Sebastian Wagner-Carena, Aizhan Akhmetzhanova, Sydney Erickson
Abstract: A common challenge in the natural sciences is to disentangle distinct, unknown sources from observations. Examples of this source separation task include deblending galaxies in a crowded field, distinguishing the activity of individual neurons from overlapping signals, and separating seismic events from an ambient background. Traditional analyses often rely on simplified source models that fail to accurately reproduce the data. Recent advances have shown that diffusion models can directly learn complex prior distributions from noisy, incomplete data. In this work, we show that diffusion models can solve the source separation problem without explicit assumptions about the source. Our method relies only on multiple views, or the property that different sets of observations contain different linear transformations of the unknown sources. We show that our method succeeds even when no source is individually observed and the observations are noisy, incomplete, and vary in resolution. The learned diffusion models enable us to sample from the source priors, evaluate the probability of candidate sources, and draw from the joint posterior of the source distribution given an observation. We demonstrate the effectiveness of our method on a range of synthetic problems as well as real-world galaxy observations.
Authors: Edward Hirst, Sanjaye Ramgoolam
Abstract: Ensembles of neural network weight matrices are studied through the training process for the MNIST classification problem, testing the efficacy of matrix models for representing their distributions, under assumptions of Gaussianity and permutation-symmetry. The general 13-parameter permutation invariant Gaussian matrix models are found to be effective models for the correlated Gaussianity in the weight matrices, beyond the range of applicability of the simple Gaussian with independent identically distributed matrix variables, and notably well beyond the initialisation step. The representation theoretic model parameters, and the graph-theoretic characterisation of the permutation invariant matrix observables give an interpretable framework for the best-fit model and for small departures from Gaussianity. Additionally, the Wasserstein distance is calculated for this class of models and used to quantify the movement of the distributions over training. Throughout the work, the effects of varied initialisation regimes, regularisation, layer depth, and layer width are tested for this formalism, identifying limits where particular departures from Gaussianity are enhanced and how more general, yet still highly-interpretable, models can be developed.
Authors: Haining Pan, James V. Roggeveen, Erez Berg, Juan Carrasquilla, Debanjan Chowdhury, Surya Ganguli, Federico Ghimenti, Juraj Hasik, Henry Hunt, Hong-Chen Jiang, Mason Kamb, Ying-Jer Kao, Ehsan Khatami, Michael J. Lawler, Di Luo, Titus Neupert, Xiaoliang Qi, Michael P. Brenner, Eun-Ah Kim
Abstract: Large language models (LLMs) have shown remarkable progress in coding and math problem-solving, but evaluation on advanced research-level problems in hard sciences remains scarce. To fill this gap, we present CMT-Benchmark, a dataset of 50 problems covering condensed matter theory (CMT) at the level of an expert researcher. Topics span analytical and computational approaches in quantum many-body, and classical statistical mechanics. The dataset was designed and verified by a panel of expert researchers from around the world. We built the dataset through a collaborative environment that challenges the panel to write and refine problems they would want a research assistant to solve, including Hartree-Fock, exact diagonalization, quantum/variational Monte Carlo, density matrix renormalization group (DMRG), quantum/classical statistical mechanics, and model building. We evaluate LLMs by programmatically checking solutions against expert-supplied ground truth. We developed machine-grading, including symbolic handling of non-commuting operators via normal ordering. They generalize across tasks too. Our evaluations show that frontier models struggle with all of the problems in the dataset, highlighting a gap in the physical reasoning skills of current LLMs. Notably, experts identified strategies for creating increasingly difficult problems by interacting with the LLMs and exploiting common failure modes. The best model, GPT5, solves 30\% of the problems; average across 17 models (GPT, Gemini, Claude, DeepSeek, Llama) is 11.4$\pm$2.1\%. Moreover, 18 problems are solved by none of the 17 models, and 26 by at most one. These unsolved problems span Quantum Monte Carlo, Variational Monte Carlo, and DMRG. Answers sometimes violate fundamental symmetries or have unphysical scaling dimensions. We believe this benchmark will guide development toward capable AI research assistants and tutors.
Authors: Mohammad Mahdi Ahmadi, Erfan Yazdandoost Hamedani
Abstract: We study a class of misspecified saddle point (SP) problems, where the optimization objective depends on an unknown parameter that must be learned concurrently from data. Unlike existing studies that assume parameters are fully known or pre-estimated, our framework integrates optimization and learning into a unified formulation, enabling a more flexible problem class. To address this setting, we propose two algorithms based on the accelerated primal-dual (APD) by Hamedani & Aybat 2021. In particular, we first analyze the naive extension of the APD method by directly substituting the evolving parameter estimates into the primal-dual updates; then, we design a new learning-aware variant of the APD method that explicitly accounts for parameter dynamics by adjusting the momentum updates. Both methods achieve a provable convergence rate of $\mathcal{O}(\log K / K)$, while the learning-aware approach attains a tighter $\mathcal{O}(1)$ constant and further benefits from an adaptive step-size selection enabled by a backtracking strategy. Furthermore, we extend the framework to problems where the learning problem admits multiple optimal solutions, showing that our modified algorithm for a structured setting achieves an $\mathcal{O}(1/\sqrt{K})$ rate. To demonstrate practical impact, we evaluate our methods on a misspecified portfolio optimization problem and show superior empirical performance compared to state-of-the-art algorithms.
Authors: Yuezhu Xu, S. Sivaranjani
Abstract: The Lipschitz constant is a key measure for certifying the robustness of neural networks to input perturbations. However, computing the exact constant is NP-hard, and standard approaches to estimate the Lipschitz constant involve solving a large matrix semidefinite program (SDP) that scales poorly with network size. Further, there is a potential to efficiently leverage local information on the input region to provide tighter Lipschitz estimates. We address this problem here by proposing a compositional framework that yields tight yet scalable Lipschitz estimates for deep feedforward neural networks. Specifically, we begin by developing a generalized SDP framework that is highly flexible, accommodating heterogeneous activation function slope, and allowing Lipschitz estimates with respect to arbitrary input-output pairs and arbitrary choices of sub-networks of consecutive layers. We then decompose this generalized SDP into a sequence of small sub-problems, with computational complexity that scales linearly with respect to the network depth. We also develop a variant that achieves near-instantaneous computation through closed-form solutions to each sub-problem. All our algorithms are accompanied by theoretical guarantees on feasibility and validity. Next, we develop a series of algorithms, termed as ECLipsE-Gen-Local, that effectively incorporate local information on the input. Our experiments demonstrate that our algorithms achieve substantial speedups over a multitude of benchmarks while producing significantly tighter Lipschitz bounds than global approaches. Moreover, we show that our algorithms provide strict upper bounds for the Lipschitz constant with values approaching the exact Jacobian from autodiff when the input region is small enough. Finally, we demonstrate the practical utility of our approach by showing that our Lipschitz estimates closely align with network robustness.
Authors: Paloma Garc\'ia-de-Herreros, Philipp Slusallek, Dietrich Klakow, Vagrant Gautam
Abstract: Large language models have shown great success on natural language tasks in recent years, but they have also shown great promise when adapted to new modalities, e.g., for scientific machine learning tasks. Even though decoder-only models are more popular within NLP and scale exceedingly well at generating natural language, most proposed approaches for cross-modal adaptation focus on encoder-only models, raising the question of how model architecture affects these approaches. In this paper, we therefore perform a series of ablation studies to answer this question, systematically comparing encoder-only and decoder-only models on cross-modal adaptation for time-dependent simulation tasks based on partial differential equations (PDEs). We find that decoder-only models are far worse than encoder-only models, when existing approaches are applied unmodified. In contrast to several other domains, scaling decoder-only models also does not help. To harness the potential of decoder-only models in this context, we introduce two novel approaches, Parallel Flipping and Sequence Doubling, attempting to mimic bidirectionality in autoregressive models. Both our methods improve overall performance using decoder-only models for all tasks and all cross-model adaptation methods, closing the gap to encoder-only model performance. We hope that our findings broaden the spectrum of models used on cross-modal adaptation tasks to further scientific ML.
Authors: Rui Lin, Yiwen Zhang, Zhicheng Peng, Minghao Lyu
Abstract: Decision Transformer (DT), which integrates reinforcement learning (RL) with the transformer model, introduces a novel approach to offline RL. Unlike classical algorithms that take maximizing cumulative discounted rewards as objective, DT instead maximizes the likelihood of actions. This paradigm shift, however, presents two key challenges: stitching trajectories and extrapolation of action. Existing methods, such as substituting specific tokens with predictive values and integrating the Policy Gradient (PG) method, address these challenges individually but fail to improve performance stably when combined due to inherent instability. To address this, we propose Action Gradient (AG), an innovative methodology that directly adjusts actions to fulfill a function analogous to that of PG, while also facilitating efficient integration with token prediction techniques. AG utilizes the gradient of the Q-value with respect to the action to optimize the action. The empirical results demonstrate that our method can significantly enhance the performance of DT-based algorithms, with some results achieving state-of-the-art levels.
Authors: Joel Wendin, Erik G. Larsson, Claudio Altafini
Abstract: For the signed graph associated to a deep neural network, one can compute the frustration level, i.e., test how close or distant the graph is to structural balance. For all the pretrained deep convolutional neural networks we consider, we find that the frustration is always less than expected from null models. From a statistical physics point of view, and in particular in reference to an Ising spin glass model, the reduced frustration indicates that the amount of disorder encoded in the network is less than in the null models. From a functional point of view, low frustration (i.e., proximity to structural balance) means that the function representing the network behaves near-monotonically, i.e., more similarly to a monotone function than in the null models. Evidence of near-monotonic behavior along the partial order determined by frustration is observed for all networks we consider. This confirms that the class of deep convolutional neural networks tends to have a more ordered behavior than expected from null models, and suggests a novel form of implicit regularization.
Authors: Ruoxing Yang
Abstract: Large language models (LLMs) such as ChatGPT have evolved into powerful and ubiquitous tools. Fine-tuning on small datasets allows LLMs to acquire specialized skills for specific tasks efficiently. Although LLMs provide great utility in both general and task-specific use cases, they are limited by two security-related concerns. First, traditional LLM hardware requirements make them infeasible to run locally on consumer-grade devices. A remote network connection with the LLM provider's server is usually required, making the system vulnerable to network attacks. Second, fine-tuning an LLM for a sensitive task may involve sensitive data. Non-private fine-tuning algorithms produce models vulnerable to training data reproduction attacks. Our work addresses these security concerns by enhancing differentially private optimization algorithms and applying them to fine-tune localizable language models. We introduce adaptable gradient clipping along with other engineering enhancements to the standard DP-Adam optimizer to create DP-Adam-AC. We use our optimizer to fine-tune examples of two localizable LLM designs, small language model (Qwen2.5-0.5B) and 1.58 bit quantization (Bitnet-b1.58-2B). We demonstrate promising improvements in loss through experimentation with two synthetic datasets.
Authors: Kevin Player
Abstract: We study the cosine similarity of sentence transformer embeddings and observe that they are well modeled by gamma mixtures. From a fixed corpus, we measure similarities between all document embeddings and a reference query embedding. Empirically we find that these distributions are often well captured by a gamma distribution shifted and truncated to [-1,1], and in many cases, by a gamma mixture. We propose a heuristic model in which a hierarchical clustering of topics naturally leads to a gamma-mixture structure in the similarity scores. Finally, we outline an expectation-maximization algorithm for fitting shifted gamma mixtures, which provides a practical tool for modeling similarity distributions.
Authors: Zhenyu Liu, Varun Ojha
Abstract: Adversarial training is the most effective defense against adversarial attacks. The effectiveness of the adversarial attacks has been on the design of its loss function and regularization term. The most widely used loss function in adversarial training is cross-entropy and mean squared error (MSE) as its regularization objective. However, MSE enforces overly uniform optimization between two output distributions during training, which limits its robustness in adversarial training scenarios. To address this issue, we revisit the idea of mutual learning (originally designed for knowledge distillation) and propose two novel regularization strategies tailored for adversarial training: (i) weighted adversarial mutual regularization and (ii) adversarial generalization regularization. In the former, we formulate a decomposed adversarial mutual Kullback-Leibler divergence (KL-divergence) loss, which allows flexible control over the optimization process by assigning unequal weights to the main and auxiliary objectives. In the latter, we introduce an additional clean target distribution into the adversarial training objective, improving generalization and enhancing model robustness. Extensive experiments demonstrate that our proposed methods significantly improve adversarial robustness compared to existing regularization-based approaches.
Authors: Qian Wang, Mohammad N. Bisheh, Kamran Paynabar
Abstract: Modern sensing and metrology systems now stream terabytes of heterogeneous, high-dimensional (HD) data profiles, images, and dense point clouds, whose natural representation is multi-way tensors. Understanding such data requires regression models that preserve tensor geometry, yet remain expressive enough to capture the pronounced nonlinear interactions that dominate many industrial and mechanical processes. Existing tensor-based regressors meet the first requirement but remain essentially linear. Conversely, conventional neural networks offer nonlinearity only after flattening, thereby discarding spatial structure and incurring prohibitive parameter counts. This paper introduces a Tensor-on-Tensor Regression Neural Network (TRNN) that unifies these two paradigms.
Authors: Hyung Gyu Rho
Abstract: Direct Preference Optimization (DPO) has emerged as a simple and effective method for aligning large language models. However, its reliance on a fixed temperature parameter leads to suboptimal training on diverse preference data, causing overfitting on easy examples and under-learning from informative ones. Recent methods have emerged to counter this. While IPO addresses general overfitting, its uniform regularization can be overly conservative. The more targeted approach of $\beta$-DPO suffers from its own limitations: its batch-level adaptation applies a single, compromised temperature to mixed-margin pairs, its linear update rule can produce unstable negative $\beta$ values, and its filtering mechanism discards potentially useful training signals. In this work, we introduce Margin-Adaptive Direct Preference Optimization (MADPO), a method that provides a stable, data-preserving, and instance-level solution. MADPO employs a practical two-step approach: it first trains a reward model to estimate preference margins and then uses these margins to apply a continuous, adaptive weight to the DPO loss for each individual training sample. This re-weighting scheme creates an effective target margin that is amplified for hard pairs and dampened for easy pairs, allowing for granular control over the learning signal. We provide a comprehensive theoretical analysis, proving that MADPO has a well-behaved optimization landscape and is robust to reward model estimation errors. We validate our theory with experiments on a sentiment generation task, where MADPO consistently and significantly outperforms strong baselines across datasets of varying quality. It achieves performance gains of up to +33.3\% on High Quality data and +10.5\% on Low Quality data over the next-best method. Our results establish MADPO as a more robust and principled approach to preference alignment.
Authors: Jinghao Cao, Qin Li, Mengnan Du, Haimin Wang, Bo Shen
Abstract: We propose Physics-informed Attention-enhanced Fourier Neural Operator (PIANO) to solve the Nonlinear Force-Free Field (NLFFF) problem in solar physics. Unlike conventional approaches that rely on iterative numerical methods, our proposed PIANO directly learns the 3D magnetic field structure from 2D boundary conditions. Specifically, PIANO integrates Efficient Channel Attention (ECA) mechanisms with Dilated Convolutions (DC), which enhances the model's ability to capture multimodal input by prioritizing critical channels relevant to the magnetic field's variations. Furthermore, we apply physics-informed loss by enforcing the force-free and divergence-free conditions in the training process so that our prediction is consistent with underlying physics with high accuracy. Experimental results on the ISEE NLFFF dataset show that our PIANO not only outperforms state-of-the-art neural operators in terms of accuracy but also shows strong consistency with the physical characteristics of NLFFF data across magnetic fields reconstructed from various solar active regions. The GitHub of this project is available https://github.com/Autumnstar-cjh/PIANO
Authors: Alex Iacob, Andrej Jovanovic, Mher Safaryan, Meghdad Kurmanji, Lorenzo Sani, Samuel Horv\'ath, William F. Shen, Xinchi Qiu, Nicholas D. Lane
Abstract: Training large models with distributed data parallelism (DDP) requires frequent communication of gradients across workers, which can saturate bandwidth. Infrequent communication strategies (e.g., Local SGD) reduce this overhead but, when applied to adaptive optimizers, often suffer a performance gap relative to fully synchronous DDP. We trace this gap to a time-scale mismatch: the optimizer's fast-moving momentum, tuned for frequent updates, decays too quickly to smooth gradients over long intervals, leading to noise-dominated optimization. To address this, we propose MT-DAO, a family of optimizers that employs multiple slow- and fast-moving first momenta or the gradient to track update dynamics across different time scales, for which we provide the first convergence guarantees. Empirically, for language-model pre-training, this eliminates the performance gap with DDP, outperforming infrequent-communication baselines in perplexity and reducing iso-token wall-clock time by 6-27% on Ethernet interconnects. At the 720M scale, MT-DAO reaches a target perplexity in 24% fewer steps and 35% less time than the single-momentum DDP baseline. MT-DAO enables effective cross-datacenter training and training over wide geographic areas.
Authors: Utkarsh Saxena, Kaushik Roy
Abstract: Quantizing the key-value (KV) cache is a promising strategy for improving the inference efficiency of large language models (LLMs). However, aggressive quantization to very low precision (e.g., 2 bits) introduces significant errors in the stored key and value tensors, which propagate through the dot-product attention mechanism and ultimately degrade generation quality. To address this, we propose KVLinC, a framework to mitigate attention errors introduced by KV cache quantization in the extreme low-precision regime. KVLinC combines a Hadamard rotation, which reduces quantization error in values, with lightweight linear correction adapters that explicitly compensate for errors introduced by quantized keys. Across extensive evaluations on the LLaMA, Qwen2.5, and Qwen3 model families, KVLinC consistently matches or surpasses strong baselines while achieving higher KV-cache compression. Furthermore, we implement a custom attention kernel that results in upto 2.55x faster inference compared to Flash Attention baseline, enabling efficient long-context LLM inference.
Authors: Rohan Arni, Carlos Blanco
Abstract: Physics-Informed Neural Networks (PINNs) are a useful framework for approximating partial differential equation solutions using deep learning methods. In this paper, we propose a principled redesign of the PINNsformer, a Transformer-based PINN architecture. We present the Spectral PINNSformer (S-Pformer), a refinement of encoder-decoder PINNSformers that addresses two key issues; 1. the redundancy (i.e. increased parameter count) of the encoder, and 2. the mitigation of spectral bias. We find that the encoder is unnecessary for capturing spatiotemporal correlations when relying solely on self-attention, thereby reducing parameter count. Further, we integrate Fourier feature embeddings to explicitly mitigate spectral bias, enabling adaptive encoding of multiscale behaviors in the frequency domain. Our model outperforms encoder-decoder PINNSformer architectures across all benchmarks, achieving or outperforming MLP performance while reducing parameter count significantly.
Authors: Mikil Foss, Andrew Lamperski
Abstract: Estimating the Kullback-Leibler (KL) divergence between random variables is a fundamental problem in statistical analysis. For continuous random variables, traditional information-theoretic estimators scale poorly with dimension and/or sample size. To mitigate this challenge, a variety of methods have been proposed to estimate KL divergences and related quantities, such as mutual information, using neural networks. The existing theoretical analyses show that neural network parameters achieving low error exist. However, since they rely on non-constructive neural network approximation theorems, they do not guarantee that the existing algorithms actually achieve low error. In this paper, we propose a KL divergence estimation algorithm using a shallow neural network with randomized hidden weights and biases (i.e. a random feature method). We show that with high probability, the algorithm achieves a KL divergence estimation error of $O(m^{-1/2}+T^{-1/3})$, where $m$ is the number of neurons and $T$ is both the number of steps of the algorithm and the number of samples.
Authors: Ahmad Alsheikh, Andreas Fischer
Abstract: Accurate and efficient temperature prediction is critical for optimizing the preheating process of PET preforms in industrial microwave systems prior to blow molding. We propose a novel deep learning framework for generalized temperature prediction. Unlike traditional models that require extensive retraining for each material or design variation, our method introduces a data-efficient neural architecture that leverages transfer learning and model fusion to generalize across unseen scenarios. By pretraining specialized neural regressor on distinct conditions such as recycled PET heat capacities or varying preform geometries and integrating their representations into a unified global model, we create a system capable of learning shared thermal dynamics across heterogeneous inputs. The architecture incorporates skip connections to enhance stability and prediction accuracy. Our approach reduces the need for large simulation datasets while achieving superior performance compared to models trained from scratch. Experimental validation on two case studies material variability and geometric diversity demonstrates significant improvements in generalization, establishing a scalable ML-based solution for intelligent thermal control in manufacturing environments. Moreover, the approach highlights how data-efficient generalization strategies can extend to other industrial applications involving complex physical modeling with limited data.
Authors: Kangwoo Yi, Bo Shen, Qin Li, Haimin Wang, Yong-Jae Moon, Jaewon Lee, Hwanhee Lee
Abstract: Solar Proton Events (SPEs) cause significant radiation hazards to satellites, astronauts, and technological systems. Accurate forecasting of their proton flux time profiles is crucial for early warnings and mitigation. This paper explores deep learning sequence-to-sequence (seq2seq) models based on Long Short-Term Memory networks to predict 24-hour proton flux profiles following SPE onsets. We used a dataset of 40 well-connected SPEs (1997-2017) observed by NOAA GOES, each associated with a >=M-class western-hemisphere solar flare and undisturbed proton flux profiles. Using 4-fold stratified cross-validation, we evaluate seq2seq model configurations (varying hidden units and embedding dimensions) under multiple forecasting scenarios: (i) proton-only input vs. combined proton+X-ray input, (ii) original flux data vs. trend-smoothed data, and (iii) autoregressive vs. one-shot forecasting. Our major results are as follows: First, one-shot forecasting consistently yields lower error than autoregressive prediction, avoiding the error accumulation seen in iterative approaches. Second, on the original data, proton-only models outperform proton+X-ray models. However, with trend-smoothed data, this gap narrows or reverses in proton+X-ray models. Third, trend-smoothing significantly enhances the performance of proton+X-ray models by mitigating fluctuations in the X-ray channel. Fourth, while models trained on trendsmoothed data perform best on average, the best-performing model was trained on original data, suggesting that architectural choices can sometimes outweigh the benefits of data preprocessing.
Authors: Xin Gu, Yingtai Xiao, Guanlin He, Jiamu Bai, Daniel Kifer, Kiwan Maeng
Abstract: Differentially private stochastic gradient descent (DP-SGD) offers the promise of training deep learning models while mitigating many privacy risks. However, there is currently a large accuracy gap between DP-SGD and normal SGD training. This has resulted in different lines of research investigating orthogonal ways of improving privacy-preserving training. One such line of work, known as DP-MF, correlates the privacy noise across different iterations of stochastic gradient descent -- allowing later iterations to cancel out some of the noise added to earlier iterations. In this paper, we study how to improve this noise correlation. We propose a technique called NoiseCurve that uses model curvature, estimated from public unlabeled data, to improve the quality of this cross-iteration noise correlation. Our experiments on various datasets, models, and privacy parameters show that the noise correlations computed by NoiseCurve offer consistent and significant improvements in accuracy over the correlation scheme used by DP-MF.
Authors: Shrenik Bhansali, Larry Heck
Abstract: Autoregressive (AR) decoding is a major latency bottleneck for large language models. Speculative decoding (SD) accelerates AR by letting a drafter propose multi-token blocks that a verifier accepts or rejects. However, many SD systems require heavy offline training or extra components. These choices raise data/compute cost and can yield brittle drafters under distribution drift. We introduce \emph{Draft, Verify, \& Improve (DVI)}, a training-aware self-speculative framework that combines inference with continual online learning. We partition an LLM into a drafter and a verifier, and during generation, verifier accept/reject decisions are converted into supervision signals and used to update the drafter head. A simple \emph{KL$\rightarrow$RL} schedule bootstraps calibration via online distillation and then adds reward-masked cross-entropy with a on-policy policy-gradient term, preserving lossless, single model deployment. On Spec-Bench, DVI achieves a $2.16\times$ wall-time speedup, on par with SoTA approaches like EAGLE-2, while orders of magnitude less data for training, and ablations show that DVI outperforms KL-only online distillation. DVI demonstrates that \emph{training-aware} self-speculation can deliver state-of-the-art, lossless speedups with minimal training overhead.
Authors: Nazanin Ahmadi, Qianying Cao, Jay D. Humphrey, George Em Karniadakis
Abstract: Physics-informed machine learning (PIML) is emerging as a potentially transformative paradigm for modeling complex biomedical systems by integrating parameterized physical laws with data-driven methods. Here, we review three main classes of PIML frameworks: physics-informed neural networks (PINNs), neural ordinary differential equations (NODEs), and neural operators (NOs), highlighting their growing role in biomedical science and engineering. We begin with PINNs, which embed governing equations into deep learning models and have been successfully applied to biosolid and biofluid mechanics, mechanobiology, and medical imaging among other areas. We then review NODEs, which offer continuous-time modeling, especially suited to dynamic physiological systems, pharmacokinetics, and cell signaling. Finally, we discuss deep NOs as powerful tools for learning mappings between function spaces, enabling efficient simulations across multiscale and spatially heterogeneous biological domains. Throughout, we emphasize applications where physical interpretability, data scarcity, or system complexity make conventional black-box learning insufficient. We conclude by identifying open challenges and future directions for advancing PIML in biomedical science and engineering, including issues of uncertainty quantification, generalization, and integration of PIML and large language models.
Authors: Zizhao Wang, Dingcheng Li, Vaishakh Keshava, Phillip Wallis, Ananth Balashankar, Peter Stone, Lukas Rutishauser
Abstract: Large Language Model (LLM) agents can leverage tools such as Google Search to complete complex tasks. However, this tool usage introduces the risk of indirect prompt injections, where malicious instructions hidden in tool outputs can manipulate the agent, posing security risks like data leakage. Current defense strategies typically rely on fine-tuning LLM agents on datasets of known attacks. However, the generation of these datasets relies on manually crafted attack patterns, which limits their diversity and leaves agents vulnerable to novel prompt injections. To address this limitation, we propose Adversarial Reinforcement Learning for Agent Safety (ARLAS), a novel framework that leverages adversarial reinforcement learning (RL) by formulating the problem as a two-player zero-sum game. ARLAS co-trains two LLMs: an attacker that learns to autonomously generate diverse prompt injections and an agent that learns to defend against them while completing its assigned tasks. To ensure robustness against a wide range of attacks and to prevent cyclic learning, we employ a population-based learning framework that trains the agent to defend against all previous attacker checkpoints. Evaluated on BrowserGym and AgentDojo, agents fine-tuned with ARLAS achieve a significantly lower attack success rate than the original model while also improving their task success rate. Our analysis further confirms that the adversarial process generates a diverse and challenging set of attacks, leading to a more robust agent compared to the base model.
Authors: Runlin Zhou, Chixiang Chen, Elynn Chen
Abstract: We study meta-reinforcement learning in finite-horizon MDPs where related tasks share similar structures in their optimal action-value functions. Specifically, we posit a linear representation $Q^*_h(s,a)=\Phi_h(s,a)\,\theta^{(k)}_h$ and place a Gaussian meta-prior $ \mathcal{N}(\theta^*_h,\Sigma^*_h)$ over the task-specific parameters $\theta^{(k)}_h$. Building on randomized value functions, we propose two Thompson-style algorithms: (i) MTSRL, which learns only the prior mean and performs posterior sampling with the learned mean and known covariance; and (ii) $\text{MTSRL}^{+}$, which additionally estimates the covariance and employs prior widening to control finite-sample estimation error. Further, we develop a prior-alignment technique that couples the posterior under the learned prior with a meta-oracle that knows the true prior, yielding meta-regret guarantees: we match prior-independent Thompson sampling in the small-task regime and strictly improve with more tasks once the prior is learned. Concretely, for known covariance we obtain $\tilde{O}(H^{4}S^{3/2}\sqrt{ANK})$ meta-regret, and with learned covariance $\tilde{O}(H^{4}S^{3/2}\sqrt{AN^3K})$; both recover a better behavior than prior-independent after $K \gtrsim \tilde{O}(H^2)$ and $K \gtrsim \tilde{O}(N^2H^2)$, respectively. Simulations on a stateful recommendation environment (with feature and prior misspecification) show that after brief exploration, MTSRL/MTSRL\(^+\) track the meta-oracle and substantially outperform prior-independent RL and bandit-only meta-baselines. Our results give the first meta-regret guarantees for Thompson-style RL with learned Q-priors, and provide practical recipes (warm-start via RLSVI, OLS aggregation, covariance widening) for experiment-rich settings.
Authors: Arpit Kapoor, Rohitash Chandra
Abstract: Conceptual rainfall-runoff models aid hydrologists and climate scientists in modelling streamflow to inform water management practices. Recent advances in deep learning have unravelled the potential for combining hydrological models with deep learning models for better interpretability and improved predictive performance. In our previous work, we introduced DeepGR4J, which enhanced the GR4J conceptual rainfall-runoff model using a deep learning model to serve as a surrogate for the routing component. DeepGR4J had an improved rainfall-runoff prediction accuracy, particularly in arid catchments. Quantile regression models have been extensively used for quantifying uncertainty while aiding extreme value forecasting. In this paper, we extend DeepGR4J using a quantile regression-based ensemble learning framework to quantify uncertainty in streamflow prediction. We also leverage the uncertainty bounds to identify extreme flow events potentially leading to flooding. We further extend the model to multi-step streamflow predictions for uncertainty bounds. We design experiments for a detailed evaluation of the proposed framework using the CAMELS-Aus dataset. The results show that our proposed Quantile DeepGR4J framework improves the predictive accuracy and uncertainty interval quality (interval score) compared to baseline deep learning models. Furthermore, we carry out flood risk evaluation using Quantile DeepGR4J, and the results demonstrate its suitability as an early warning system.
Authors: Yurun Song, Zhuoyi Yang, Ian G. Harris, Sangeetha Abdu Jyothi
Abstract: Large Language Models (LLMs) are scaling rapidly, creating significant challenges for collaborative server client distributed training, particularly in terms of communication efficiency and computational overheads. To address these challenges, we implement Parameter-efficient Split Learning, which effectively balances efficiency and performance for collaborative training on low-resource devices. To reduce communication overhead in collaborative training, we introduce Adaptive Mixed bit Activation Quantization (AMAQ), a strategy that progressively compresses activations and gradients from high precision (6 to 8 bits) to low precision (3 to 4 bits). AMAQ achieves this by effectively allocating bit budgets across channels based on feature wise and layer wise importance using bit regularization. Under the same bit budgets, AMAQ outperforms fixed-precision approaches, delivering about 2.5% higher generation accuracy and about 1.3% better classification accuracy for models like LLaMA3 8B and Qwen2.5 7B. In addition, it significantly enhances training stability and reducing ultra-low bit representation collapse during the training. Experiments demonstrate that AMAQ integrates effectively into practical multi-machine collaborative training setups, offering superior inference accuracy with only a modest communication overhead for bits adaptation during training. This trade off makes AMAQ a practical and effective solution for collaborative training with minimal communication cost.
Authors: Luke Thompson, Davy Guan, Dai Shi, Slade Matthews, Junbin Gao, Andi Han
Abstract: Molecular dynamics (MD) simulations underpin modern computational drug dis- covery, materials science, and biochemistry. Recent machine learning models provide high-fidelity MD predictions without the need to repeatedly solve quantum mechanical forces, enabling significant speedups over conventional pipelines. Yet many such methods typically enforce strict equivariance and rely on sequential rollouts, thus limiting their flexibility and simulation efficiency. They are also com- monly single-task, trained on individual molecules and fixed timeframes, which restricts generalization to unseen compounds and extended timesteps. To address these issues, we propose Atomistic Transformer Operator for Molecules (ATOM), a pretrained transformer neural operator for multitask molecular dynamics. ATOM adopts a quasi-equivariant design that requires no explicit molecular graph and employs a temporal attention mechanism, allowing for the accurate parallel decod- ing of multiple future states. To support operator pretraining across chemicals and timescales, we curate TG80, a large, diverse, and numerically stable MD dataset with over 2.5 million femtoseconds of trajectories across 80 compounds. ATOM achieves state-of-the-art performance on established single-task benchmarks, such as MD17, RMD17 and MD22. After multitask pretraining on TG80, ATOM shows exceptional zero-shot generalization to unseen molecules across varying time hori- zons. We believe ATOM represents a significant step toward accurate, efficient, and transferable molecular dynamics models
Authors: Reza T. Batley, Sourav Saha
Abstract: Training - the optimisation of complex models - is traditionally performed through small, local, iterative updates [D. E. Rumelhart, G. E. Hinton, R. J. Williams, Nature 323, 533-536 (1986)]. Approximating solutions through truncated gradients is a paradigm dating back to Cauchy [A.-L. Cauchy, Comptes Rendus Math\'ematique 25, 536-538 (1847)] and Newton [I. Newton, The Method of Fluxions and Infinite Series (Henry Woodfall, London, 1736)]. This work introduces the Method of Infinite Descent, a semi-analytic optimisation paradigm that reformulates training as the direct solution to the first-order optimality condition. By analytical resummation of its Taylor expansion, this method yields an exact, algebraic equation for the update step. Realisation of the infinite Taylor tower's cascading resummation is formally derived, and an exploitative algorithm for the direct solve step is proposed. This principle is demonstrated with the herein-introduced AION (Analytic, Infinitely-Optimisable Network) architecture. AION is a model designed expressly to satisfy the algebraic closure required by Infinite Descent. In a simple test problem, AION reaches the optimum in a single descent step. Together, this optimiser-model pair exemplify how analytic structure enables exact, non-iterative convergence. Infinite Descent extends beyond this example, applying to any appropriately closed architecture. This suggests a new class of semi-analytically optimisable models: the \emph{Infinity Class}; sufficient conditions for class membership are discussed. This offers a pathway toward non-iterative learning.
Authors: Zichong Li, Liming Liu, Chen Liang, Weizhu Chen, Tuo Zhao
Abstract: The choice of optimizer significantly impacts the training efficiency and computational costs of large language models (LLMs). Recently, the Muon optimizer has demonstrated promising results by orthogonalizing parameter updates, improving optimization geometry through better conditioning. Despite Muon's emergence as a candidate successor to Adam, the potential for jointly leveraging their strengths has not been systematically explored. In this work, we bridge this gap by proposing NorMuon (Neuron-wise Normalized Muon), an optimizer that synergistically combines orthogonalization with neuron-level adaptive learning rates. Our analysis reveals that while Muon effectively reduces condition numbers, the resulting updates exhibit highly non-uniform neuron norms, causing certain neurons to dominate the optimization process. NorMuon addresses this imbalance by maintaining second-order momentum statistics for each neuron and applying row-wise normalization after orthogonalization, ensuring balanced parameter utilization while preserving Muon's conditioning benefits. To enable practical deployment at scale, we develop an efficient distributed implementation under the FSDP2 framework that strategically distributes orthogonalization computations across devices. Experiments across multiple model scales demonstrate that NorMuon consistently outperforms both Adam and Muon, achieving 21.74% better training efficiency than Adam and 11.31% improvement over Muon on 1.1 B pretraining setting, while maintaining a comparable memory footprint to Muon. Our findings suggest that orthogonalization and adaptive learning rates are complementary rather than competing approaches, opening new avenues for optimizer design in large-scale deep learning.
Authors: Zhuoyi Huang, Nutan Sahoo, Anamika Kumari, Girish Kumar, Kexuan Cai, Shixing Cao, Yue Kang, Tian Xia, Somya Chatterjee, Nicholas Hausman, Aidan Jay, Eric S. Rosenthal, Soundar Srinivasan, Sadid Hasan, Alex Fedorov, Sulaiman Vesal, Soundar Srinivasan, Sadid Hasan, Alex Fedorov, Sulaiman Vesal
Abstract: The development of machine learning for cardiac care is severely hampered by privacy restrictions on sharing real patient electrocardiogram (ECG) data. Although generative AI offers a promising solution, the real-world use of existing model-synthesized ECGs is limited by persistent gaps in trustworthiness and clinical utility. In this work, we address two major shortcomings of current generative ECG methods: insufficient morphological fidelity and the inability to generate personalized, patient-specific physiological signals. To address these gaps, we build on a conditional diffusion-based Structured State Space Model (SSSD-ECG) with two principled innovations: (1) MIDT-ECG (Mel-Spectrogram Informed Diffusion Training), a novel training paradigm with time-frequency domain supervision to enforce physiological structural realism, and (2) multi-modal demographic conditioning to enable patient-specific synthesis. We comprehensively evaluate our approach on the PTB-XL dataset, assessing the synthesized ECG signals on fidelity, clinical coherence, privacy preservation, and downstream task utility. MIDT-ECG achieves substantial gains: it improves morphological coherence, preserves strong privacy guarantees with all metrics evaluated exceeding the baseline by 4-8%, and notably reduces the interlead correlation error by an average of 74%, while demographic conditioning enhances signal-to-noise ratio and personalization. In critical low-data regimes, a classifier trained on datasets supplemented with our synthetic ECGs achieves performance comparable to a classifier trained solely on real data. Together, we demonstrate that ECG synthesizers, trained with the proposed time-frequency structural regularization scheme, can serve as personalized, high-fidelity, privacy-preserving surrogates when real data are scarce, advancing the responsible use of generative AI in healthcare.
Authors: Yang Cao, Zhao Song, Jiahao Zhang, Jiale Zhao
Abstract: Graph neural networks (GNNs) have become a core paradigm for learning on relational data. In materials science, equivariant GNNs (EGNNs) have emerged as a compelling backbone for crystalline-structure prediction, owing to their ability to respect Euclidean symmetries and periodic boundary conditions. Despite strong empirical performance, their expressive power in periodic, symmetry-constrained settings remains poorly understood. This work characterizes the intrinsic computational and expressive limits of EGNNs for crystalline-structure prediction through a circuit-complexity lens. We analyze the computations carried out by EGNN layers acting on node features, atomic coordinates, and lattice matrices, and prove that, under polynomial precision, embedding width $d=O(n)$ for $n$ nodes, $O(1)$ layers, and $O(1)$-depth, $O(n)$-width MLP instantiations of the message/update/readout maps, these models admit a simulation by a uniform $\mathsf{TC}^0$ threshold-circuit family of polynomial size (with an explicit constant-depth bound). Situating EGNNs within $\mathsf{TC}^0$ provides a concrete ceiling on the decision and prediction problems solvable by such architectures under realistic resource constraints and clarifies which architectural modifications (e.g., increased depth, richer geometric primitives, or wider layers) are required to transcend this regime. The analysis complements Weisfeiler-Lehman style results that do not directly transfer to periodic crystals, and offers a complexity-theoretic foundation for symmetry-aware graph learning on crystalline systems.
Authors: Aavid Mathrawala, Dhruv Kurup, Josie Lau
Abstract: Current pain assessment within hospitals often relies on self-reporting or non-specific EKG vital signs. This system leaves critically ill, sedated, and cognitively impaired patients vulnerable to undertreated pain and opioid overuse. Electroencephalography (EEG) offers a noninvasive method of measuring brain activity. This technology could potentially be applied as an assistive tool to highlight nociceptive processing in order to mitigate this issue. In this study, we compared machine learning models for classifying high-pain versus low/no-pain EEG epochs using data from fifty-two healthy adults exposed to laser-evoked pain at three intensities (low, medium, high). Each four-second epoch was transformed into a 537-feature vector spanning spectral power, band ratios, Hjorth parameters, entropy measures, coherence, wavelet energies, and peak-frequency metrics. Nine traditional machine learning models were evaluated with leave-one-participant-out cross-validation. A support vector machine with radial basis function kernel achieved the best offline performance with 88.9% accuracy and sub-millisecond inference time (1.02 ms). Our Feature importance analysis was consistent with current canonical pain physiology, showing contralateral alpha suppression, midline theta/alpha enhancement, and frontal gamma bursts. The real-time XGBoost model maintained an end-to-end latency of about 4 ms and 94.2% accuracy, demonstrating that an EEG-based pain monitor is technically feasible within a clinical setting and provides a pathway towards clinical validation.
Authors: Wei-Ting Tang, Akshay Kudva, Joel A. Paulson
Abstract: Bayesian optimization (BO) is effective for expensive black-box problems but remains challenging in high dimensions. We propose NeST-BO, a local BO method that targets the Newton step by jointly learning gradient and Hessian information with Gaussian process surrogates, and selecting evaluations via a one-step lookahead bound on Newton-step error. We show that this bound (and hence the step error) contracts with batch size, so NeST-BO directly inherits inexact-Newton convergence: global progress under mild stability assumptions and quadratic local rates once steps are sufficiently accurate. To scale, we optimize the acquisition in low-dimensional subspaces (e.g., random embeddings or learned sparse subspaces), reducing the dominant cost of learning curvature from $O(d^2)$ to $O(m^2)$ with $m \ll d$ while preserving step targeting. Across high-dimensional synthetic and real-world problems, including cases with thousands of variables and unknown active subspaces, NeST-BO consistently yields faster convergence and lower regret than state-of-the-art local and high-dimensional BO baselines.
Authors: Ziyi Chen, Junyi Li, Peiran Yu, Heng Huang
Abstract: Reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) are important techniques to align large language models (LLM) with human preference. However, the quality of RLHF and DPO training is seriously compromised by \textit{\textbf{C}orrupted} preference, reward \textit{\textbf{O}veroptimization}, and bias towards \textit{\textbf{V}erbosity}. To our knowledge, most existing works tackle only one of these important issues, and the few other works require much computation to estimate multiple reward models and lack theoretical guarantee of generalization ability. In this work, we propose RLHF-\textbf{COV} and DPO-\textbf{COV} algorithms that can simultaneously mitigate these three issues, in both offline and online settings. This ability is theoretically demonstrated by obtaining length-regularized generalization error rates for our DPO-COV algorithms trained on corrupted data, which match the best-known rates for simpler cases with clean data and without length regularization. Moreover, our DPO-COV algorithm is simple to implement without reward estimation, and is proved to be equivalent to our RLHF-COV algorithm, which directly implies the equivalence between the vanilla RLHF and DPO algorithms. Experiments demonstrate the effectiveness of our DPO-COV algorithms under both offline and online settings.
Authors: Yuyao Wang, Yu-Hung Cheng, Debarghya Mukherjee, Huimin Cheng
Abstract: Graphon models provide a flexible nonparametric framework for estimating latent connectivity probabilities in networks, enabling a range of downstream applications such as link prediction and data augmentation. However, accurate graphon estimation typically requires a large graph, whereas in practice, one often only observes a small-sized network. One approach to addressing this issue is to adopt a transfer learning framework, which aims to improve estimation in a small target graph by leveraging structural information from a larger, related source graph. In this paper, we propose a novel method, namely GTRANS, a transfer learning framework that integrates neighborhood smoothing and Gromov-Wasserstein optimal transport to align and transfer structural patterns between graphs. To prevent negative transfer, GTRANS includes an adaptive debiasing mechanism that identifies and corrects for target-specific deviations via residual smoothing. We provide theoretical guarantees on the stability of the estimated alignment matrix and demonstrate the effectiveness of GTRANS in improving the accuracy of target graph estimation through extensive synthetic and real data experiments. These improvements translate directly to enhanced performance in downstream applications, such as the graph classification task and the link prediction task.
Authors: Lawrence Liu, Alexander Liu, Mengdi Wang, Tuo Zhao, Lin F. Yang
Abstract: Large language models (LLMs) present significant deployment challenges due to their immense computational and memory requirements. While semi-structured pruning, particularly 2:4 sparsity, offers a path to practical hardware acceleration, existing methods often incur substantial performance degradation. To bridge this gap, we introduce ARMOR: (Adaptive Representation with Matrix-factORization), a novel one-shot post-training pruning algorithm. Instead of directly pruning weights, ARMOR factorizes each weight matrix into a 2:4 sparse core wrapped by two low-overhead, block diagonal matrices. These wrappers act as efficient pre and post-transformation error correctors, offering greater flexibility to preserve model quality compared to conventional 2:4 pruning techniques. The sparse core and block diagonal wrappers are chosen through a block coordinate descent algorithm that minimizes a layer-wise proxy loss. We theoretically prove this optimization is guaranteed to converge to a solution with a proxy loss less than or equal to state-of-the-art pruning algorithms. Experiments on Llama (Touvron et al., 2023; Dubey et al., 2024) and Qwen (Yang et al., 2025) model families demonstrate that ARMOR consistently and significantly outperforms state-of-the-art 2:4 pruning methods across a wide range of downstream tasks and perplexity evaluations. ARMOR achieves this superior performance while retaining the inference speedups and substantial memory usage reductions of 2:4 pruning, establishing a more effective trade-off between model compression and task accuracy
Authors: Harshil Vejendla
Abstract: Test-time adaptation (TTA) aims to adapt a pretrained model to distribution shifts using only unlabeled test data. While promising, existing methods like Tent suffer from instability and can catastrophically forget the source knowledge, especially with small batch sizes or challenging corruptions. We argue that this arises from overly deterministic updates on a complex loss surface. In this paper, we introduce Langevin-Anchored Test-Time Adaptation (LATTA), a novel approach that regularizes adaptation through two key mechanisms: (1) a noisy weight perturbation inspired by Stochastic Gradient Langevin Dynamics (SGLD) to explore the local parameter space and escape poor local minima, and (2) a stable weight anchor that prevents the model from diverging from its robust source pre-training. This combination allows LATTA to adapt effectively without sacrificing stability. Unlike prior Bayesian TTA methods, LATTA requires no architectural changes or expensive Monte Carlo passes. We conduct extensive experiments on standard benchmarks, including Rotated-MNIST and the more challenging CIFAR-10-C. Our results demonstrate that LATTA significantly outperforms existing methods, including Tent, CoTTA, and EATA, setting a new state of the art for self-supervised TTA by improving average accuracy on CIFAR-10-C by over 2% while simultaneously reducing performance variance.
Authors: Rui Liu, Tao Zhe, Yanjie Fu, Feng Xia, Ted Senator, Dongjie Wang
Abstract: Feature selection eliminates redundancy among features to improve downstream task performance while reducing computational overhead. Existing methods often struggle to capture intricate feature interactions and adapt across diverse application scenarios. Recent advances employ generative intelligence to alleviate these drawbacks. However, these methods remain constrained by permutation sensitivity in embedding and reliance on convexity assumptions in gradient-based search. To address these limitations, our initial work introduces a novel framework that integrates permutation-invariant embedding with policy-guided search. Although effective, it still left opportunities to adapt to realistic distributed scenarios. In practice, data across local clients is highly imbalanced, heterogeneous and constrained by strict privacy regulations, limiting direct sharing. These challenges highlight the need for a framework that can integrate feature selection knowledge across clients without exposing sensitive information. In this extended journal version, we advance the framework from two perspectives: 1) developing a privacy-preserving knowledge fusion strategy to derive a unified representation space without sharing sensitive raw data. 2) incorporating a sample-aware weighting strategy to address distributional imbalance among heterogeneous local clients. Extensive experiments validate the effectiveness, robustness, and efficiency of our framework. The results further demonstrate its strong generalization ability in federated learning scenarios. The code and data are publicly available: https://anonymous.4open.science/r/FedCAPS-08BF.
Authors: Shi Chen, Zhengjiang Lin, Yury Polyanskiy, Philippe Rigollet
Abstract: As large language models scale to longer contexts, attention layers suffer from a fundamental pathology: attention scores collapse toward uniformity as context length $n$ increases, causing tokens to cluster excessively, a phenomenon known as rank-collapse. While $\textit{attention scaling}$ effectively addresses this deficiency by rescaling attention scores with a polylogarithmic factor $\beta_n$, theoretical justification for this approach remains lacking. We analyze a simplified yet tractable model that magnifies the effect of attention scaling. In this model, attention exhibits a phase transition governed by the scaling factor $\beta_n$: insufficient scaling collapses all tokens to a single direction, while excessive scaling reduces attention to identity, thereby eliminating meaningful interactions between tokens. Our main result identifies the critical scaling $\beta_n \asymp \log n$ and provides a rigorous justification for attention scaling in YaRN and Qwen, clarifying why logarithmic scaling maintains sparse, content-adaptive attention at large context lengths.
Authors: Sheng Xiang, Yidong Jiang, Yunting Chen, Dawei Cheng, Guoping Zhao, Changjun Jiang
Abstract: Spoofing detection in financial trading is crucial, especially for identifying complex behaviors such as conspiracy spoofing. Traditional machine-learning approaches primarily focus on isolated node features, often overlooking the broader context of interconnected nodes. Graph-based techniques, particularly Graph Neural Networks (GNNs), have advanced the field by leveraging relational information effectively. However, in real-world spoofing detection datasets, trading behaviors exhibit dynamic, irregular patterns. Existing spoofing detection methods, though effective in some scenarios, struggle to capture the complexity of dynamic and diverse, evolving inter-node relationships. To address these challenges, we propose a novel framework called the Generative Dynamic Graph Model (GDGM), which models dynamic trading behaviors and the relationships among nodes to learn representations for conspiracy spoofing detection. Specifically, our approach incorporates the generative dynamic latent space to capture the temporal patterns and evolving market conditions. Raw trading data is first converted into time-stamped sequences. Then we model trading behaviors using the neural ordinary differential equations and gated recurrent units, to generate the representation incorporating temporal dynamics of spoofing patterns. Furthermore, pseudo-label generation and heterogeneous aggregation techniques are employed to gather relevant information and enhance the detection performance for conspiratorial spoofing behaviors. Experiments conducted on spoofing detection datasets demonstrate that our approach outperforms state-of-the-art models in detection accuracy. Additionally, our spoofing detection system has been successfully deployed in one of the largest global trading markets, further validating the practical applicability and performance of the proposed method.
Authors: Sheng Xiang, Chenhao Xu, Dawei Cheng, Xiaoyang Wang, Ying Zhang
Abstract: Graph simulation has recently received a surge of attention in graph processing and analytics. In real-life applications, e.g. social science, biology, and chemistry, many graphs are composed of a series of evolving graphs (i.e., temporal graphs). While most of the existing graph generators focus on static graphs, the temporal information of the graphs is ignored. In this paper, we focus on simulating temporal graphs, which aim to reproduce the structural and temporal properties of the observed real-life temporal graphs. In this paper, we first give an overview of the existing temporal graph generators, including recently emerged learning-based approaches. Most of these learning-based methods suffer from one of the limitations: low efficiency in training or slow generating, especially for temporal random walk-based methods. Therefore, we propose an efficient learning-based approach to generate graph snapshots, namely temporal graph autoencoder (TGAE). Specifically, we propose an attention-based graph encoder to encode temporal and structural characteristics on sampled ego-graphs. And we proposed an ego-graph decoder that can achieve a good trade-off between simulation quality and efficiency in temporal graph generation. Finally, the experimental evaluation is conducted among our proposed TGAE and representative temporal graph generators on real-life temporal graphs and synthesized graphs. It is reported that our proposed approach outperforms the state-of-the-art temporal graph generators by means of simulation quality and efficiency.
Authors: Praneeth Vepakomma, Kaustubh Ponkshe
Abstract: Traditional collaborative learning approaches are based on sharing of model weights between clients and a server. However, there are advantages to resource efficiency through schemes based on sharing of embeddings (activations) created from the data. Several differentially private methods were developed for sharing of weights while such mechanisms do not exist so far for sharing of embeddings. We propose Ours to learn a privacy encoding network in conjunction with a small utility generation network such that the final embeddings generated from it are equipped with formal differential privacy guarantees. These privatized embeddings are then shared with a more powerful server, that learns a post-processing that results in a higher accuracy for machine learning tasks. We show that our co-design of collaborative and private learning results in requiring only one round of privatized communication and lesser compute on the client than traditional methods. The privatized embeddings that we share from the client are agnostic to the type of model (deep learning, random forests or XGBoost) used on the server in order to process these activations to complete a task.
Authors: Jiashu Tao, Reza Shokri
Abstract: Machine learning models are known to leak sensitive information, as they inevitably memorize (parts of) their training data. More alarmingly, large language models (LLMs) are now trained on nearly all available data, which amplifies the magnitude of information leakage and raises serious privacy risks. Hence, it is more crucial than ever to quantify privacy risk before the release of LLMs. The standard method to quantify privacy is via membership inference attacks, where the state-of-the-art approach is the Robust Membership Inference Attack (RMIA). In this paper, we present InfoRMIA, a principled information-theoretic formulation of membership inference. Our method consistently outperforms RMIA across benchmarks while also offering improved computational efficiency. In the second part of the paper, we identify the limitations of treating sequence-level membership inference as the gold standard for measuring leakage. We propose a new perspective for studying membership and memorization in LLMs: token-level signals and analyses. We show that a simple token-based InfoRMIA can pinpoint which tokens are memorized within generated outputs, thereby localizing leakage from the sequence level down to individual tokens, while achieving stronger sequence-level inference power on LLMs. This new scope rethinks privacy in LLMs and can lead to more targeted mitigation, such as exact unlearning.
Authors: Arindam Chowdhury, Massimiliano Lupo Pasini
Abstract: Graph neural networks (GNNs) are widely used as surrogates for costly experiments and first-principles simulations to study the behavior of compounds at atomistic scale, and their architectural complexity is constantly increasing to enable the modeling of complex physics. While most recent GNNs combine more traditional message passing neural networks (MPNNs) layers to model short-range interactions with more advanced graph transformers (GTs) with global attention mechanisms to model long-range interactions, it is still unclear when global attention mechanisms provide real benefits over well-tuned MPNN layers due to inconsistent implementations, features, or hyperparameter tuning. We introduce the first unified, reproducible benchmarking framework - built on HydraGNN - that enables seamless switching among four controlled model classes: MPNN, MPNN with chemistry/topology encoders, GPS-style hybrids of MPNN with global attention, and fully fused local - global models with encoders. Using seven diverse open-source datasets for benchmarking across regression and classification tasks, we systematically isolate the contributions of message passing, global attention, and encoder-based feature augmentation. Our study shows that encoder-augmented MPNNs form a robust baseline, while fused local-global models yield the clearest benefits for properties governed by long-range interaction effects. We further quantify the accuracy - compute trade-offs of attention, reporting its overhead in memory. Together, these results establish the first controlled evaluation of global attention in atomistic graph learning and provide a reproducible testbed for future model development.
Authors: Kangjia Yan, Chenxi Liu, Hao Miao, Xinle Wu, Yan Zhao, Chenjuan Guo, Bin Yang
Abstract: The proliferation of mobile devices generates a massive volume of time series across various domains, where effective time series forecasting enables a variety of real-world applications. This study focuses on a new problem of source-free domain adaptation for time series forecasting. It aims to adapt a pretrained model from sufficient source time series to the sparse target time series domain without access to the source data, embracing data protection regulations. To achieve this, we propose TimePD, the first source-free time series forecasting framework with proxy denoising, where large language models (LLMs) are employed to benefit from their generalization capabilities. Specifically, TimePD consists of three key components: (1) dual-branch invariant disentangled feature learning that enforces representation- and gradient-wise invariance by means of season-trend decomposition; (2) lightweight, parameter-free proxy denoising that dynamically calibrates systematic biases of LLMs; and (3) knowledge distillation that bidirectionally aligns the denoised prediction and the original target prediction. Extensive experiments on real-world datasets offer insight into the effectiveness of the proposed TimePD, outperforming SOTA baselines by 9.3% on average.
Authors: Andrew Ly, Pulin Gong
Abstract: Fundamental limits to predictability are central to our understanding of many physical and computational systems. Here we show that, despite its remarkable capabilities, deep learning exhibits such fundamental limits rooted in the fractal, riddled geometry of its basins of attraction: any initialization that leads to one solution lies arbitrarily close to another that leads to a different one. We derive sufficient conditions for the emergence of riddled basins by analytically linking features widely observed in deep learning, including chaotic learning dynamics and symmetry-induced invariant subspaces, to reveal a general route to riddling in realistic deep networks. The resulting basins of attraction possess an infinitely fine-scale fractal structure characterized by an uncertainty exponent near zero, so that even large increases in the precision of initial conditions yield only marginal gains in outcome predictability. Riddling thus imposes a fundamental limit on the predictability and hence reproducibility of neural network training, providing a unified account of many empirical observations. These results reveal a general organizing principle of deep learning with important implications for optimization and the safe deployment of artificial intelligence.
Authors: Salah Eddine Choutri, Prajwal Chauhan, Othmane Mazhar, Saif Eddin Jabari
Abstract: The Monte Carlo-type Neural Operator (MCNO) introduces a framework for learning solution operators of one-dimensional partial differential equations (PDEs) by directly learning the kernel function and approximating the associated integral operator using a Monte Carlo-type approach. Unlike Fourier Neural Operators (FNOs), which rely on spectral representations and assume translation-invariant kernels, MCNO makes no such assumptions. The kernel is represented as a learnable tensor over sampled input-output pairs, and sampling is performed once, uniformly at random from a discretized grid. This design enables generalization across multiple grid resolutions without relying on fixed global basis functions or repeated sampling during training, while an interpolation step maps between arbitrary input and output grids to further enhance flexibility. Experiments on standard 1D PDE benchmarks show that MCNO achieves competitive accuracy with efficient computational cost. We also provide a theoretical analysis proving that the Monte Carlo estimator yields a bounded bias and variance under mild regularity assumptions. This result holds in any spatial dimension, suggesting that MCNO may extend naturally beyond one-dimensional problems. More broadly, this work explores how Monte Carlo-type integration can be incorporated into neural operator frameworks for continuous-domain PDEs, providing a theoretically supported alternative to spectral methods (such as FNO) and to graph-based Monte Carlo approaches (such as the Graph Kernel Neural Operator, GNO).
Authors: Alexander Murphy, Michal Danilowski, Soumyajit Chatterjee, Abhirup Ghosh
Abstract: Test-Time Adaptation (TTA) methods are often computationally expensive, require a large amount of data for effective adaptation, or are brittle to hyperparameters. Based on a theoretical foundation of the geometry of the latent space, we are able to significantly improve the alignment between source and distribution-shifted samples by re-centering target data embeddings at the origin. This insight motivates NEO -- a hyperparameter-free fully TTA method, that adds no significant compute compared to vanilla inference. NEO is able to improve the classification accuracy of ViT-Base on ImageNet-C from 55.6% to 59.2% after adapting on just one batch of 64 samples. When adapting on 512 samples NEO beats all 7 TTA methods we compare against on ImageNet-C, ImageNet-R and ImageNet-S and beats 6/7 on CIFAR-10-C, while using the least amount of compute. NEO performs well on model calibration metrics and additionally is able to adapt from 1 class to improve accuracy on 999 other classes in ImageNet-C. On Raspberry Pi and Jetson Orin Nano devices, NEO reduces inference time by 63% and memory usage by 9% compared to baselines. Our results based on 3 ViT architectures and 4 datasets show that NEO can be used efficiently and effectively for TTA.
Authors: David Debot, Giuseppe Marra
Abstract: Concept Bottleneck Models (CBNMs) are deep learning models that provide interpretability by enforcing a bottleneck layer where predictions are based exclusively on human-understandable concepts. However, this constraint also restricts information flow and often results in reduced predictive accuracy. Concept Sidechannel Models (CSMs) address this limitation by introducing a sidechannel that bypasses the bottleneck and carry additional task-relevant information. While this improves accuracy, it simultaneously compromises interpretability, as predictions may rely on uninterpretable representations transmitted through sidechannels. Currently, there exists no principled technique to control this fundamental trade-off. In this paper, we close this gap. First, we present a unified probabilistic concept sidechannel meta-model that subsumes existing CSMs as special cases. Building on this framework, we introduce the Sidechannel Independence Score (SIS), a metric that quantifies a CSM's reliance on its sidechannel by contrasting predictions made with and without sidechannel information. We propose SIS regularization, which explicitly penalizes sidechannel reliance to improve interpretability. Finally, we analyze how the expressivity of the predictor and the reliance of the sidechannel jointly shape interpretability, revealing inherent trade-offs across different CSM architectures. Empirical results show that state-of-the-art CSMs, when trained solely for accuracy, exhibit low representation interpretability, and that SIS regularization substantially improves their interpretability, intervenability, and the quality of learned interpretable task predictors. Our work provides both theoretical and practical tools for developing CSMs that balance accuracy and interpretability in a principled manner.
Authors: F\'elix Vandervorst, Bruno Deprez, Wouter Verbeke, Tim Verdonck
Abstract: Graph-based methods are becoming increasingly popular in machine learning due to their ability to model complex data and relations. Insurance fraud is a prime use case, since false claims are often the result of organised criminals that stage accidents or the same persons filing erroneous claims on multiple policies. One challenge is that graph-based approaches struggle to find meaningful representations of the data because of the high class imbalance present in fraud data. Another is that insurance networks are heterogeneous and dynamic, given the changing relations among people, companies and policies. That is why gradient boosted tree approaches on tabular data still dominate the field. Therefore, we present a novel inductive graph gradient boosting machine (G-GBM) for supervised learning on heterogeneous and dynamic graphs. We show that our estimator competes with popular graph neural network approaches in an experiment using a variety of simulated random graphs. We demonstrate the power of G-GBM for insurance fraud detection using an open-source and a real-world, proprietary dataset. Given that the backbone model is a gradient boosting forest, we apply established explainability methods to gain better insights into the predictions made by G-GBM.
Authors: Haribandhu Jena, Jyotirmaya Shivottam, Subhankar Mishra
Abstract: Quantum graph neural networks offer a powerful paradigm for learning on graph-structured data, yet their explainability is complicated by measurement-induced stochasticity and the combinatorial nature of graph structure. In this paper, we introduce QuantumGraphLIME (QGraphLIME), a model-agnostic, post-hoc framework that treats model explanations as distributions over local surrogates fit on structure-preserving perturbations of a graph. By aggregating surrogate attributions together with their dispersion, QGraphLIME yields uncertainty-aware node and edge importance rankings for quantum graph models. The framework further provides a distribution-free, finite-sample guarantee on the size of the surrogate ensemble: a Dvoretzky-Kiefer-Wolfowitz bound ensures uniform approximation of the induced distribution of a binary class probability at target accuracy and confidence under standard independence assumptions. Empirical studies on controlled synthetic graphs with known ground truth demonstrate accurate and stable explanations, with ablations showing clear benefits of nonlinear surrogate modeling and highlighting sensitivity to perturbation design. Collectively, these results establish a principled, uncertainty-aware, and structure-sensitive approach to explaining quantum graph neural networks, and lay the groundwork for scaling to broader architectures and real-world datasets, as quantum resources mature. Code is available at https://github.com/smlab-niser/qglime.
Authors: Aditya Desai, Kumar Krishna Agrawal, Shuo Yang, Alejandro Cuadron, Luis Gaspar Schroeder, Matei Zaharia, Joseph E. Gonzalez, Ion Stoica
Abstract: State-of-the-art sparse attention methods for reducing decoding latency fall into two main categories: approximate top-$k$ (and its extension, top-$p$) and recently introduced sampling-based estimation. However, these approaches are fundamentally limited in their ability to approximate full attention: they fail to provide consistent approximations across heads and query vectors and, most critically, lack guarantees on approximation quality, limiting their practical deployment. We observe that top-$k$ and random sampling are complementary: top-$k$ performs well when attention scores are dominated by a few tokens, whereas random sampling provides better estimates when attention scores are relatively uniform. Building on this insight and leveraging the statistical guarantees of sampling, we introduce vAttention, the first practical sparse attention mechanism with user-specified $(\epsilon, \delta)$ guarantees on approximation accuracy (thus, verified). These guarantees make vAttention a compelling step toward practical, reliable deployment of sparse attention at scale. By unifying top-k and sampling, vAttention outperforms both individually, delivering a superior quality-efficiency trade-off. Our experiments show that vAttention significantly improves the quality of sparse attention (e.g., $\sim$4.5 percentage points for Llama-3.1-8B-Inst and Deepseek-R1-Distill-Llama-8B on RULER-HARD), and effectively bridges the gap between full and sparse attention (e.g., across datasets, it matches full model quality with upto 20x sparsity). We also demonstrate that it can be deployed in reasoning scenarios to achieve fast decoding without compromising model quality (e.g., vAttention achieves full model quality on AIME2024 at 10x sparsity with up to 32K token generations). Code is open-sourced at https://github.com/xAlg-ai/sparse-attention-hub.
Authors: Yihan Du, Seo Taek Kong, R. Srikant
Abstract: The widespread application of Large Language Models (LLMs) imposes increasing demands on safety, such as reducing harmful content and fake information, and avoiding certain forbidden tokens due to rules and laws. While there have been several recent works studying safe alignment of LLMs, these works either require the training of reward and cost models and incur high memory and computational costs, or need prior knowledge about the optimal solution. Motivated by this fact, we study the problem of constrained alignment in LLMs, i.e., maximizing the output reward while restricting the cost due to potentially unsafe content to stay below a threshold. For this problem, we propose a novel primal-dual DPO approach, which first trains a model using standard DPO on reward preference data to provide reward information, and then adopts a rearranged Lagrangian DPO objective utilizing the provided reward information to fine-tune LLMs on cost preference data. Our approach significantly reduces memory and computational costs, and does not require extra prior knowledge. Moreover, we establish rigorous theoretical guarantees on the suboptimality and constraint violation of the output policy. We also extend our approach to an online data setting by incorporating exploration bonuses, which enables our approach to explore uncovered prompt-response space, and then provide theoretical results that get rid of the dependence on preference data coverage. Experimental results on the widely-used preference dataset PKU-SafeRLHF demonstrate the effectiveness of our approach.
Authors: Hedi Zisling, Ilan Naiman, Nimrod Berman, Supasorn Suwajanakorn, Omri Azencot
Abstract: Unsupervised representation learning, particularly sequential disentanglement, aims to separate static and dynamic factors of variation in data without relying on labels. This remains a challenging problem, as existing approaches based on variational autoencoders and generative adversarial networks often rely on multiple loss terms, complicating the optimization process. Furthermore, sequential disentanglement methods face challenges when applied to real-world data, and there is currently no established evaluation protocol for assessing their performance in such settings. Recently, diffusion models have emerged as state-of-the-art generative models, but no theoretical formalization exists for their application to sequential disentanglement. In this work, we introduce the Diffusion Sequential Disentanglement Autoencoder (DiffSDA), a novel, modal-agnostic framework effective across diverse real-world data modalities, including time series, video, and audio. DiffSDA leverages a new probabilistic modeling, latent diffusion, and efficient samplers, while incorporating a challenging evaluation protocol for rigorous testing. Our experiments on diverse real-world benchmarks demonstrate that DiffSDA outperforms recent state-of-the-art methods in sequential disentanglement.
Authors: S. Peng, L. Hu, W. Zhang, B. Jie, Y. Luo
Abstract: Graph embedding has been widely applied in areas such as network analysis, social network mining, recommendation systems, and bioinformatics. However, current graph construction methods often require the prior definition of neighborhood size, limiting the effective revelation of potential structural correlations in the data. Additionally, graph embedding methods using linear projection heavily rely on a singular pattern mining approach, resulting in relative weaknesses in adapting to different scenarios. To address these challenges, we propose a novel model, Neighborhood-Adaptive Generalized Linear Graph Embedding (NGLGE), grounded in latent pattern mining. This model introduces an adaptive graph learning method tailored to the neighborhood, effectively revealing intrinsic data correlations. Simultaneously, leveraging a reconstructed low-rank representation and imposing $\ell_{2,0}$ norm constraint on the projection matrix allows for flexible exploration of additional pattern information. Besides, an efficient iterative solving algorithm is derived for the proposed model. Comparative evaluations on datasets from diverse scenarios demonstrate the superior performance of our model compared to state-of-the-art methods.
Authors: Chunsan Hong, Seonho An, Min-Soo Kim, Jong Chul Ye
Abstract: Masked diffusion models (MDMs) have recently emerged as a novel framework for language modeling. MDMs generate sentences by iteratively denoising masked sequences, filling in [MASK] tokens step by step. Although MDMs support any-order sampling, performance is highly sensitive to the choice of which position to unmask next. Prior work typically relies on rule-based schedules (e.g., max-confidence, max-margin), which provide ad hoc improvements. In contrast, we replace these heuristics with a learned scheduler. Specifically, we cast denoising as a KL-regularized Markov decision process (MDP) with an explicit reference policy and optimize a regularized objective that admits policy improvement and convergence guarantees under standard assumptions. We prove that the optimized policy under this framework generates samples that more closely match the data distribution than heuristic schedules. Empirically, across four benchmarks, our learned policy consistently outperforms max-confidence: for example, on SUDOKU, where unmasking order is critical, it yields a 20.1% gain over random and a 11.2% gain over max-confidence.
Authors: Hachem Madmoun, Salem Lahlou
Abstract: Eliciting cooperation in multi-agent LLM systems is critical for AI alignment. We investigate two approaches: direct communication and curriculum learning. In a 4-player Stag Hunt, a one-word "cheap talk" channel increases cooperation from 0% to 48.3%, demonstrating communication as a robust coordination mechanism. In contrast, we find that curriculum learning is highly sensitive to design choices: our pedagogical curriculum through progressively complex games reduced agent payoffs by 27.4% in an Iterated Public Goods Game with Punishment. Qualitative analysis reveals that curricula emphasizing defection-equilibrium games can induce "learned pessimism" in agents. These findings suggest that for coordination problems, simple communication protocols may be more reliable than experience-based training, and that curriculum design for social dilemmas requires careful attention to the strategic lessons embedded in game sequences.
Authors: Xiao Yang, Xuejiao Zhao, Zhiqi Shen
Abstract: Graph neural networks (GNNs) have achieved remarkable success in node classification. Building on this progress, heterogeneous graph neural networks (HGNNs) integrate relation types and node and edge semantics to leverage heterogeneous information. Causal analysis for HGNNs is advancing rapidly, aiming to separate genuine causal effects from spurious correlations. However, whether HGNNs are intrinsically effective remains underexamined, and most studies implicitly assume rather than establish this effectiveness. In this work, we examine HGNNs from two perspectives: model architecture and heterogeneous information. We conduct a systematic reproduction across 21 datasets and 20 baselines, complemented by comprehensive hyperparameter retuning. To further disentangle the source of performance gains, we develop a causal effect estimation framework that constructs and evaluates candidate factors under standard assumptions through factual and counterfactual analyses, with robustness validated via minimal sufficient adjustment sets, cross-method consistency checks, and sensitivity analyses. Our results lead to two conclusions. First, model architecture and complexity have no causal effect on performance. Second, heterogeneous information exerts a positive causal effect by increasing homophily and local-global distribution discrepancy, which makes node classes more distinguishable. The implementation is publicly available at https://github.com/YXNTU/CausalHGNN.
Authors: Yuxuan Bai, Gauri Pradhan, Marlon Tobaben, Antti Honkela
Abstract: With the emergence of powerful large-scale foundation models, the training paradigm is increasingly shifting from from-scratch training to transfer learning. This enables high utility training with small, domain-specific datasets typical in sensitive applications.Membership inference attacks (MIAs) provide an empirical estimate of the privacy leakage by machine learning models. Yet, prior assessments of MIAs against models fine-tuned with transfer learning rely on a small subset of possible attacks. We address this by comparing performance of diverse MIAs in transfer learning settings to help practitioners identify the most efficient attacks for privacy risk evaluation. We find that attack efficacy decreases with the increase in training data for score-based MIAs. We find that there is no one MIA which captures all privacy risks in models trained with transfer learning. While the Likelihood Ratio Attack (LiRA) demonstrates superior performance across most experimental scenarios, the Inverse Hessian Attack (IHA) proves to be more effective against models fine-tuned on PatchCamelyon dataset in high data regime.
Authors: Shadi Rahimian, Mario Fritz
Abstract: Single nucleotide polymorphism (SNP) datasets are fundamental to genetic studies but pose significant privacy risks when shared. The correlation of SNPs with each other makes strong adversarial attacks such as masked-value reconstruction, kin, and membership inference attacks possible. Existing privacy-preserving approaches either apply differential privacy to statistical summaries of these datasets or offer complex methods that require post-processing and the usage of a publicly available dataset to suppress or selectively share SNPs. In this study, we introduce an innovative framework for generating synthetic SNP sequence datasets using samples derived from time-inhomogeneous hidden Markov models (TIHMMs). To preserve the privacy of the training data, we ensure that each SNP sequence contributes only a bounded influence during training, enabling strong differential privacy guarantees. Crucially, by operating on full SNP sequences and bounding their gradient contributions, our method directly addresses the privacy risks introduced by their inherent correlations. Through experiments conducted on the real-world 1000 Genomes dataset, we demonstrate the efficacy of our method using privacy budgets of $\varepsilon \in [1, 10]$ at $\delta=10^{-4}$. Notably, by allowing the transition models of the HMM to be dependent on the location in the sequence, we significantly enhance performance, enabling the synthetic datasets to closely replicate the statistical properties of non-private datasets. This framework facilitates the private sharing of genomic data while offering researchers exceptional flexibility and utility.
Authors: Pafue Christy Nganjimi, Andrew Soltan, Danielle Belgrave, Lei Clifton, David A. Clifton, Anshul Thakur
Abstract: Dataset condensation (DC) enables the creation of compact, privacy-preserving synthetic datasets that can match the utility of real patient records, supporting democratised access to highly regulated clinical data for developing downstream clinical models. State-of-the-art DC methods supervise synthetic data by aligning the training dynamics of models trained on real and those trained on synthetic data, typically using full stochastic gradient descent (SGD) trajectories as alignment targets; however, these trajectories are often noisy, high-curvature, and storage-intensive, leading to unstable gradients, slow convergence, and substantial memory overhead. We address these limitations by replacing full SGD trajectories with smooth, low-loss parametric surrogates, specifically quadratic B\'ezier curves that connect the initial and final model states from real training trajectories. These mode-connected paths provide noise-free, low-curvature supervision signals that stabilise gradients, accelerate convergence, and eliminate the need for dense trajectory storage. We theoretically justify B\'ezier-mode connections as effective surrogates for SGD paths and empirically show that the proposed method outperforms state-of-the-art condensation approaches across five clinical datasets, yielding condensed datasets that enable clinically effective model development.
Authors: Giorgio Giannone, Guangxuan Xu, Nikhil Shivakumar Nayak, Rohan Mahesh Awhad, Shivchander Sudalairaj, Kai Xu, Akash Srivastava
Abstract: Inference-Time Scaling (ITS) improves language models by allocating more computation at generation time. Particle Filtering (PF) has emerged as a strong ITS method for complex mathematical reasoning tasks, but it is vulnerable when guided by process reward models, which often assign overconfident scores early in the reasoning process. This causes PF to suffer from premature exploitation: it myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This failure mode, known as particle impoverishment, is especially severe under constrained computational budgets. To address this, we analyze the problem and identify two root causes: a lack of diversity in the particle set due to overconfident resampling and consequent inability to assess the potential of a reasoning path. We introduce Entropic Particle Filtering (ePF), an algorithm that integrates two new techniques to solve these issues. The first technique, Entropic Annealing (EA), directly mitigates particle impoverishment by monitoring search diversity via entropy; when diversity drops, it intervenes by dynamically annealing the resampling distribution to preserve exploration. The second, an enhancement called Look-ahead Modulation (LaM), adds a predictive guide to evaluate a state's potential based on its successors. On several challenging math benchmarks, ePF significantly outperforms strong baselines and achieves up to a 50 % relative improvement in task reward. Together, these methods improve PF's resilience by balancing the exploration of diverse solution spaces with the exploitation of high-reward regions, ultimately leading to higher-quality solutions.
Authors: Zhi Liu, Xuyuan Hu, Xiao Han, Zhehao Dai, Zhaolin Deng, Guojiang Shen, Xiangjie Kong
Abstract: Accurate travel time estimation (TTE) plays a crucial role in intelligent transportation systems. However, it remains challenging due to heterogeneous data sources and complex traffic dynamics. Moreover, conventional approaches typically convert trajectories into fixed-length representations, neglecting the inherent variability of real-world trajectories, which often leads to information loss or feature redundancy. To address these challenges, this paper introduces the Multimodal Dynamic Trajectory Integration (MDTI) framework--a novel multimodal trajectory representation learning approach that integrates GPS sequences, grid trajectories, and road network constraints to enhance TTE accuracy. MDTI employs modality-specific encoders and a cross-modal interaction module to capture complementary spatial, temporal, and topological semantics, while a dynamic trajectory modeling mechanism adaptively regulates information density for trajectories of varying lengths. Two self-supervised pretraining objectives, named contrastive alignment and masked language modeling, further strengthen multimodal consistency and contextual understanding. Extensive experiments on three real-world datasets demonstrate that MDTI consistently outperforms state-of-the-art baselines, confirming its robustness and strong generalization abilities. The code is publicly available at: https://github.com/freshhxy/MDTI/
Authors: Adhithyan Kalaivanan, Zheng Zhao, Jens Sj\"olund, Fredrik Lindsten
Abstract: Guiding pretrained flow-based generative models for conditional generation or to produce samples with desired target properties enables solving diverse tasks without retraining on paired data. We present ESS-Flow, a gradient-free method that leverages the typically Gaussian prior of the source distribution in flow-based models to perform Bayesian inference directly in the source space using Elliptical Slice Sampling. ESS-Flow only requires forward passes through the generative model and observation process, no gradient or Jacobian computations, and is applicable even when gradients are unreliable or unavailable, such as with simulation-based observations or quantization in the generation or observation process. We demonstrate its effectiveness on designing materials with desired target properties and predicting protein structures from sparse inter-residue distance measurements.
Authors: Egor Surkov, Dmitry Osin, Evgeny Burnaev, Egor Shvetsov
Abstract: This paper studies forecasting of the future distribution of events in human action sequences, a task essential in domains like retail, finance, healthcare, and recommendation systems where the precise temporal order is often less critical than the set of outcomes. We challenge the dominant autoregressive paradigm and investigate whether explicitly modeling the future distribution or order-invariant multi-token approaches outperform order-preserving methods. We analyze local order invariance and introduce a KL-based metric to quantify temporal drift. We find that a simple explicit distribution forecasting objective consistently surpasses complex implicit baselines. We further demonstrate that mode collapse of predicted categories is primarily driven by distributional imbalance. This work provides a principled framework for selecting modeling strategies and offers practical guidance for building more accurate and robust forecasting systems.
Authors: Philipp Dahlinger, Tai Hoang, Denis Blessing, Niklas Freymuth, Gerhard Neumann
Abstract: Accurately simulating physics is crucial across scientific domains, with applications spanning from robotics to materials science. While traditional mesh-based simulations are precise, they are often computationally expensive and require knowledge of physical parameters, such as material properties. In contrast, data-driven approaches like Graph Network Simulators (GNSs) offer faster inference but suffer from two key limitations: Firstly, they must be retrained from scratch for even minor variations in physical parameters, and secondly they require labor-intensive data collection for each new parameter setting. This is inefficient, as simulations with varying parameters often share a common underlying latent structure. In this work, we address these challenges by learning this shared structure through meta-learning, enabling fast adaptation to new physical parameters without retraining. To this end, we propose a novel architecture that generates a latent representation by encoding graph trajectories using conditional neural processes (CNPs). To mitigate error accumulation over time, we combine CNPs with a novel neural operator architecture. We validate our approach, Meta Neural Graph Operator (MaNGO), on several dynamics prediction tasks with varying material properties, demonstrating superior performance over existing GNS methods. Notably, MaNGO achieves accuracy on unseen material properties close to that of an oracle model.
Authors: Julia Moska, Oleksii Furman, Kacper Kozaczko, Szymon Leszkiewicz, Jakub Polczyk, Piotr Gramacki, Piotr Szyma\'nski
Abstract: GeoAI is evolving rapidly, fueled by diverse geospatial datasets like traffic patterns, environmental data, and crowdsourced OpenStreetMap (OSM) information. While sophisticated AI models are being developed, existing benchmarks are often concentrated on single tasks and restricted to a single modality. As such, progress in GeoAI is limited by the lack of a standardized, multi-task, modality-agnostic benchmark for their systematic evaluation. This paper introduces a novel benchmark designed to assess the performance, accuracy, and efficiency of geospatial embedders. Our benchmark is modality-agnostic and comprises 7 distinct datasets from diverse cities across three continents, ensuring generalizability and mitigating demographic biases. It allows for the evaluation of GeoAI embedders on various phenomena that exhibit underlying geographic processes. Furthermore, we establish a simple and intuitive task-oriented model baselines, providing a crucial reference point for comparing more complex solutions.
Authors: Martin Benfeghoul, Teresa Delgado, Adnan Oomerjee, Haitham Bou Ammar, Jun Wang, Zafeirios Fountas
Abstract: Transformers' quadratic computational complexity limits their scalability despite remarkable performance. While linear attention reduces this to linear complexity, pre-training such models from scratch remains, in most cases, prohibitively expensive. Recent post-training linearisation methods convert pre-trained Transformers to linear models efficiently, often using hybrid approaches that combine linear attention with sliding-window softmax. We identify a critical flaw: existing hybrid methods inadvertently bypass the linear component, relying almost entirely on SWA. Component-level diagnostics reveal this previously undetected behaviour stems from overlooked evaluation practices on common-sense benchmarks. We propose three solutions to ensure balanced component usage: (i) inference-time hybridisation of linear-only conversions with sliding-window softmax; (ii) HedgeCATs, combining attention-weight transfer with targeted LoRA fine-tuning; and (iii) Scheduled Sliding-window Dropout (SSD), which stochastically suppresses the softmax branch during training to prevent component collapse. Our methods maintain computational efficiency while recovering most base model performance and ensuring genuine linear attention adoption, restoring the validity of performance attributions in hybrid conversions.
Authors: Marc Garreta Basora (Universitat Aut\`onoma de Barcelona), Mehmet Oguz Mulayim (Artificial Intelligence Research Institute, Universitat Aut\`onoma de Barcelona)
Abstract: Anomaly detection in 12-lead electrocardiograms (ECGs) is critical for identifying deviations associated with cardiovascular disease. This work presents a comparative analysis of three autoencoder-based architectures: convolutional autoencoder (CAE), variational autoencoder with bidirectional long short-term memory (VAE-BiLSTM), and VAE-BiLSTM with multi-head attention (VAE-BiLSTM-MHA), for unsupervised anomaly detection in ECGs. To the best of our knowledge, this study reports the first application of a VAE-BiLSTM-MHA architecture to ECG anomaly detection. All models are trained on normal ECG samples to reconstruct non-anomalous cardiac morphology and detect deviations indicative of disease. Using a unified preprocessing and evaluation pipeline on the public China Physiological Signal Challenge (CPSC) dataset, the attention-augmented VAE achieves the best performance, with an AUPRC of 0.81 and a recall of 0.85 on the held-out test set, outperforming the other architectures. To support clinical triage, this model is further integrated into an interactive dashboard that visualizes anomaly localization. In addition, a performance comparison with baseline models from the literature is provided.
Authors: Jacob Bamberger, Iolo Jones, Dennis Duncan, Michael M. Bronstein, Pierre Vandergheynst, Adam Gosztolai
Abstract: Deep generative models often face a fundamental tradeoff: high sample quality can come at the cost of memorisation, where the model reproduces training data rather than generalising across the underlying data geometry. We introduce Carr\'e du champ flow matching (CDC-FM), a generalisation of flow matching (FM), that improves the quality-generalisation tradeoff by regularising the probability path with a geometry-aware noise. Our method replaces the homogeneous, isotropic noise in FM with a spatially varying, anisotropic Gaussian noise whose covariance captures the local geometry of the latent data manifold. We prove that this geometric noise can be optimally estimated from the data and is scalable to large data. Further, we provide an extensive experimental evaluation on diverse datasets (synthetic manifolds, point clouds, single-cell genomics, animal motion capture, and images) as well as various neural network architectures (MLPs, CNNs, and transformers). We demonstrate that CDC-FM consistently offers a better quality-generalisation tradeoff. We observe significant improvements over standard FM in data-scarce regimes and in highly non-uniformly sampled datasets, which are often encountered in AI for science applications. Our work provides a mathematical framework for studying the interplay between data geometry, generalisation and memorisation in generative models, as well as a robust and scalable algorithm that can be readily integrated into existing flow matching pipelines.
Authors: Mohamed Bal-Ghaoui, Fayssal Sabri
Abstract: High-dimensional data remains a pervasive challenge in machine learning, often undermining model interpretability and computational efficiency. While Large Language Models (LLMs) have shown promise for dimensionality reduction through feature selection, existing LLM-based approaches frequently lack structured reasoning and transparent justification for their decisions. This paper introduces LLM-FS-Agent, a novel multi-agent architecture designed for interpretable and robust feature selection. The system orchestrates a deliberative "debate" among multiple LLM agents, each assigned a specific role, enabling collective evaluation of feature relevance and generation of detailed justifications. We evaluate LLM-FS-Agent in the cybersecurity domain using the CIC-DIAD 2024 IoT intrusion detection dataset and compare its performance against strong baselines, including LLM-Select and traditional methods such as PCA. Experimental results demonstrate that LLM-FS-Agent consistently achieves superior or comparable classification performance while reducing downstream training time by an average of 46% (statistically significant improvement, p = 0.028 for XGBoost). These findings highlight that the proposed deliberative architecture enhances both decision transparency and computational efficiency, establishing LLM-FS-Agent as a practical and reliable solution for real-world applications.
Authors: Randall Balestriero, Nicolas Ballas, Mike Rabbat, Yann LeCun
Abstract: Joint Embedding Predictive Architectures (JEPAs) learn representations able to solve numerous downstream tasks out-of-the-box. JEPAs combine two objectives: (i) a latent-space prediction term, i.e., the representation of a slightly perturbed sample must be predictable from the original sample's representation, and (ii) an anti-collapse term, i.e., not all samples should have the same representation. While (ii) is often considered as an obvious remedy to representation collapse, we uncover that JEPAs' anti-collapse term does much more--it provably estimates the data density. In short, any successfully trained JEPA can be used to get sample probabilities, e.g., for data curation, outlier detection, or simply for density estimation. Our theoretical finding is agnostic of the dataset and architecture used--in any case one can compute the learned probabilities of sample $x$ efficiently and in closed-form using the model's Jacobian matrix at $x$. Our findings are empirically validated across datasets (synthetic, controlled, and Imagenet) and across different Self Supervised Learning methods falling under the JEPA family (I-JEPA and DINOv2) and on multimodal models, such as MetaCLIP. We denote the method extracting the JEPA learned density as {\bf JEPA-SCORE}.
Authors: Xueyan Li, Guinan Su, Mrinmaya Sachan, Jonas Geiping
Abstract: Large Language Models (LLMs) are increasingly applied to complex tasks that require extended reasoning. In such settings, models often benefit from diverse chains-of-thought to arrive at multiple candidate solutions. This requires two competing objectives: to inject enough stochasticity to explore multiple reasoning chains, and to ensure sufficient accuracy and quality in each path. Existing works pursue the first objective by increasing exploration at highly uncertain steps with higher temperature or larger candidate token sets, while others improve reliability by rejecting samples with low confidence post-generation, implying that low confidence correlates with low answer quality. These two lines of thought are in conflict, as they conflate different sources of uncertainty. To resolve this, we argue that the decoding rule should be calibrated by correctness, not confidence alone. We should sample from tokens with higher estimated correctness, and reduce sampling where expected correctness is low. We propose simple strategies that achieve this goal: Greedy-Threshold makes sampling greedy at very low confidence steps. Calibrated-TopK and Calibrated-epsilon set truncation threshold based on estimated rank-wise correctness. Together, our findings challenge prevailing heuristics about decoding under uncertainty and show gains across math and general reasoning benchmarks.
Authors: Hans Weytjens, Wouter Verbeke
Abstract: This book chapter introduces the principles and practical applications of uncertainty quantification in machine learning. It explains how to identify and distinguish between different types of uncertainty and presents methods for quantifying uncertainty in predictive models, including linear regression, random forests, and neural networks. The chapter also covers conformal prediction as a framework for generating predictions with predefined confidence intervals. Finally, it explores how uncertainty estimation can be leveraged to improve business decision-making, enhance model reliability, and support risk-aware strategies.
Authors: Sai Karthikeya Vemuri, Adithya Ashok Chalain Valapil, Tim B\"uchner, Joachim Denzler
Abstract: Transferring the recent advancements in deep learning into scientific disciplines is hindered by the lack of the required large-scale datasets for training. We argue that in these knowledge-rich domains, the established body of scientific theory provides reliable inductive biases in the form of governing physical laws. We address the ill-posed inverse problem of recovering Raman spectra from noisy Coherent Anti-Stokes Raman Scattering (CARS) measurements, as the true Raman signal here is suppressed by a dominating non-resonant background. We propose RamPINN, a model that learns to recover Raman spectra from given CARS spectra. Our core methodological contribution is a physics-informed neural network that utilizes a dual-decoder architecture to disentangle resonant and non-resonant signals. This is done by enforcing the Kramers-Kronig causality relations via a differentiable Hilbert transform loss on the resonant and a smoothness prior on the non-resonant part of the signal. Trained entirely on synthetic data, RamPINN demonstrates strong zero-shot generalization to real-world experimental data, explicitly closing this gap and significantly outperforming existing baselines. Furthermore, we show that training with these physics-based losses alone, without access to any ground-truth Raman spectra, still yields competitive results. This work highlights a broader concept: formal scientific rules can act as a potent inductive bias, enabling robust, self-supervised learning in data-limited scientific domains.
Authors: Kevin Raina, Tanya Schmah
Abstract: Out-of-Distribution (OOD) detection is critical to AI reliability and safety, yet in many practical settings, only a limited amount of training data is available. Bayesian Neural Networks (BNNs) are a promising class of model on which to base OOD detection, because they explicitly represent epistemic (i.e. model) uncertainty. In the small training data regime, BNNs are especially valuable because they can incorporate prior model information. We introduce a new family of Bayesian posthoc OOD scores based on expected logit vectors, and compare 5 Bayesian and 4 deterministic posthoc OOD scores. Experiments on MNIST and CIFAR-10 In-Distributions, with 5000 training samples or less, show that the Bayesian methods outperform corresponding deterministic methods.
Authors: Andreas Maurer, Erfan Mirzaei, Massimiliano Pontil
Abstract: The paper provides data-dependent bounds on the test error of the Gibbs algorithm in the overparameterized interpolation regime, where low training errors are also obtained for impossible data, such as random labels in classification. The bounds are stable under approximation with Langevin Monte Carlo algorithms. Experiments on the MNIST and CIFAR-10 datasets verify that the bounds yield nontrivial predictions on true labeled data and correctly upper bound the test error for random labels. Our method indicates that generalization in the low-temperature, interpolation regime is already signaled by small training errors in the more classical high temperature regime.
Authors: Guillaume Godin
Abstract: We introduce molFTP (molecular fragment-target prevalence), a compact representation that delivers strong predictive performance. To prevent feature leakage across cross-validation folds, we implement a dummy-masking procedure that removes information about fragments present in the held-out molecules. We further show that key leave-one-out (key-loo) closely approximates true molecule-level leave-one-out (LOO), with deviation below 8% on our datasets. This enables near full data training while preserving unbiased cross-validation estimates of model performance. Overall, molFTP provides a fast, leakage-resistant fragment-target prevalence vectorization with practical safeguards (dummy masking or key-LOO) that approximate LOO at a fraction of its cost.
Authors: Li Zeqiao, Wang Yijing, Wang Haoyu, Li Zheng, Li Peng, Liu Wenfei, Zuo Zhiqiang
Abstract: Autonomous driving with reinforcement learning (RL) has significant potential. However, applying RL in real-world settings remains challenging due to the need for safe, efficient, and robust learning. Incorporating human expertise into the learning process can help overcome these challenges by reducing risky exploration and improving sample efficiency. In this work, we propose a reward-free, active human-in-the-loop learning method called Human-Guided Distributional Soft Actor-Critic (H-DSAC). Our method combines Proxy Value Propagation (PVP) and Distributional Soft Actor-Critic (DSAC) to enable efficient and safe training in real-world environments. The key innovation is the construction of a distributed proxy value function within the DSAC framework. This function encodes human intent by assigning higher expected returns to expert demonstrations and penalizing actions that require human intervention. By extrapolating these labels to unlabeled states, the policy is effectively guided toward expert-like behavior. With a well-designed state space, our method achieves real-world driving policy learning within practical training times. Results from both simulation and real-world experiments demonstrate that our framework enables safe, robust, and sample-efficient learning for autonomous driving.
Authors: Jie Hao, Rui Yu, Wei Zhang, Huixia Wang, Jie Xu, Mingrui Liu
Abstract: Effective data selection is essential for pretraining large language models (LLMs), enhancing efficiency and improving generalization to downstream tasks. However, existing approaches often require leveraging external pretrained models, making it difficult to disentangle the effects of data selection from those of the external pretrained models. In addition, they often overlook the long-term impact of selected data if the model is trained to convergence, primarily due to the prohibitive cost of full-scale LLM pretraining. In this paper, we introduce BLISS (\textbf{B}ileve\textbf{L} \textbf{I}nfluence \textbf{S}coring method for data \textbf{S}election): a lightweight data selection method that operates entirely \emph{from scratch}, without relying on any external pretrained oracle models, while explicitly accounting for the long-term impact of selected data. BLISS leverages a small proxy model as a surrogate for the LLM and employs a score model to estimate the long-term influence of training samples if the proxy model is trained to convergence. We formulate data selection as a bilevel optimization problem, where the upper-level objective optimizes the score model to assign importance weights to training samples, ensuring that minimizing the lower-level objective (i.e., training the proxy model over the weighted training loss until convergence) leads to best validation performance. Once optimized, the trained score model predicts influence scores for the dataset, enabling efficient selection of high-quality samples for LLM pretraining. We validate BLISS by pretraining 410M/1B/2.8B Pythia and LLaMA-0.5B models on selected subsets of the C4 dataset. Notably, under the 1B model setting, BLISS achieves $1.7\times$ speedup in reaching the same performance as the state-of-the-art method, demonstrating superior performance across multiple downstream tasks.
Authors: David L\"udke, Marten Lienen, Marcel Kollovieh, Stephan G\"unnemann
Abstract: Temporal point processes (TPPs) are a fundamental tool for modeling event sequences in continuous time, but most existing approaches rely on autoregressive parameterizations that are limited by their sequential sampling. Recent non-autoregressive, diffusion-style models mitigate these issues by jointly interpolating between noise and data through event insertions and deletions in a discrete Markov chain. In this work, we generalize this perspective and introduce an Edit Flow process for TPPs that transports noise to data via insert, delete, and substitute edit operations. By learning the instantaneous edit rates within a continuous-time Markov chain framework, we attain a flexible and efficient model that effectively reduces the total number of necessary edit operations during generation. Empirical results demonstrate the generative flexibility of our unconditionally trained model in a wide range of unconditional and conditional generation tasks on benchmark TPPs.
Authors: Dimitrios Kelesis, Dimitris Fotakis, Georgios Paliouras
Abstract: In this paper, we study the factors that contribute to the effect of oversmoothing in deep Graph Neural Networks (GNNs). Specifically, our analysis is based on a new metric (Mean Average Squared Distance - $MASED$) to quantify the extent of oversmoothing. We derive layer-wise bounds on $MASED$, which aggregate to yield global upper and lower distance bounds. Based on this quantification of oversmoothing, we further analyze the importance of two different properties of the model; namely the norms of the generated node embeddings, along with the largest and smallest singular values of the weight matrices. Building on the insights drawn from the theoretical analysis, we show that oversmoothing increases as the number of trainable weight matrices and the number of adjacency matrices increases. We also use the derived layer-wise bounds on $MASED$ to form a proposal for decoupling the number of hops (i.e., adjacency depth) from the number of weight matrices. In particular, we introduce G-Reg, a regularization scheme that increases the bounds, and demonstrate through extensive experiments that by doing so node classification accuracy increases, achieving robustness at large depths. We further show that by reducing oversmoothing in deep networks, we can achieve better results in some tasks than using shallow ones. Specifically, we experiment with a ``cold start" scenario, i.e., when there is no feature information for the unlabeled nodes. Finally, we show empirically the trade-off between receptive field size (i.e., number of weight matrices) and performance, using the $MASED$ bounds. This is achieved by distributing adjacency hops across a small number of trainable layers, avoiding the extremes of under- or over-parameterization of the GNN.
Authors: Jo\~ao Palmeiro, Diogo Duarte, Rita Costa, Pedro Bizarro
Abstract: AI models are increasingly used for data analysis and visualization, yet benchmarks rarely address scatterplot-specific tasks, limiting insight into performance. To address this gap for one of the most common chart types, we introduce a synthetic, annotated dataset of over 18,000 scatterplots from six data generators and 17 chart designs, and a benchmark based on it. We evaluate proprietary models from OpenAI and Google using N-shot prompting on five distinct tasks derived from annotations of cluster bounding boxes, their center coordinates, and outlier coordinates. OpenAI models and Gemini 2.5 Flash, especially when prompted with examples, are viable options for counting clusters and, in Flash's case, outliers (90%+ Accuracy). However, the results for localization-related tasks are unsatisfactory: Precision and Recall are near or below 50%, except for Flash in outlier identification (65.01%). Furthermore, the impact of chart design on performance appears to be a secondary factor, but it is advisable to avoid scatterplots with wide aspect ratios (16:9 and 21:9) or those colored randomly. Supplementary materials are available at https://github.com/feedzai/biy-paper.
Authors: Lulu Gong, Shreya Saxena
Abstract: Mixtures of linear dynamical systems (MoLDS) provide a path to model time-series data that exhibit diverse temporal dynamics across trajectories. However, its application remains challenging in complex and noisy settings, limiting its effectiveness for neural data analysis. Tensor-based moment methods can provide global identifiability guarantees for MoLDS, but their performance degrades under noise and complexity. Commonly used expectation-maximization (EM) methods offer flexibility in fitting latent models but are highly sensitive to initialization and prone to poor local minima. Here, we propose a tensor-based method that provides identifiability guarantees for learning MoLDS, which is followed by EM updates to combine the strengths of both approaches. The novelty in our approach lies in the construction of moment tensors using the input-output data to recover globally consistent estimates of mixture weights and system parameters. These estimates can then be refined through a Kalman EM algorithm, with closed-form updates for all LDS parameters. We validate our framework on synthetic benchmarks and real-world datasets. On synthetic data, the proposed Tensor-EM method achieves more reliable recovery and improved robustness compared to either pure tensor or randomly initialized EM methods. We then analyze neural recordings from the primate somatosensory cortex while a non-human primate performs reaches in different directions. Our method successfully models and clusters different conditions as separate subsystems, consistent with supervised single-LDS fits for each condition. Finally, we apply this approach to another neural dataset where monkeys perform a sequential reaching task. These results demonstrate that MoLDS provides an effective framework for modeling complex neural data, and that Tensor-EM is a reliable approach to MoLDS learning for these applications.
Authors: Nyal Patel, Matthieu Bou, Arjun Jagota, Satyapriya Krishna, Sonali Parbhoo
Abstract: Reinforcement Learning from Human Feedback (RLHF) aligns Large Language Models (LLMs) with human preferences, yet the underlying reward signals they internalize remain hidden, posing a critical challenge for interpretability and safety. Existing approaches attempt to extract these latent incentives using Inverse Reinforcement Learning (IRL), but treat all preference pairs equally, often overlooking the most informative signals: those examples the extracted reward model misclassifies or assigns nearly equal scores, which we term \emph{failures}. We introduce a novel \emph{failure-aware} IRL algorithm that focuses on misclassified or difficult examples to recover the latent rewards defining model behaviors. By learning from these failures, our failure-aware IRL extracts reward functions that better reflect the true objectives behind RLHF. We demonstrate that failure-aware IRL outperforms existing IRL baselines across multiple metrics when applied to LLM detoxification, without requiring external classifiers or supervision. Crucially, failure-aware IRL yields rewards that better capture the true incentives learned during RLHF, enabling more effective re-RLHF training than standard IRL. This establishes failure-aware IRL as a robust, scalable method for auditing model alignment and reducing ambiguity in the IRL process.
Authors: Matthieu Bou, Nyal Patel, Arjun Jagota, Satyapriya Krishna, Sonali Parbhoo
Abstract: The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.
Authors: Alessandro Favero
Abstract: Deep neural networks have achieved remarkable success, yet our understanding of how they learn remains limited. These models can learn high-dimensional tasks, which is generally statistically intractable due to the curse of dimensionality. This apparent paradox suggests that learnable data must have an underlying latent structure. What is the nature of this structure? How do neural networks encode and exploit it, and how does it quantitatively impact performance - for instance, how does generalization improve with the number of training examples? This thesis addresses these questions by studying the roles of locality and compositionality in data, tasks, and deep learning representations.
Authors: Prateek Humane, Paolo Cudrano, Daniel Z. Kaplan, Matteo Matteucci, Supriyo Chakraborty, Irina Rish
Abstract: Fine-tuning large language models (LLMs) on chain-of-thought (CoT) data shows that a small amount of high-quality data can outperform massive datasets. Yet, what constitutes "quality" remains ill-defined. Existing reasoning methods rely on indirect heuristics such as problem difficulty or trace length, while instruction-tuning has explored a broader range of automated selection strategies, but rarely in the context of reasoning. We propose to define reasoning data quality using influence functions, which measure the causal effect of individual CoT examples on downstream accuracy, and introduce influence-based pruning, which consistently outperforms perplexity and embedding-based baselines on math reasoning within a model family.
Authors: Markus Krimmel, Philip Hartout, Karsten Borgwardt, Dexiong Chen
Abstract: Existing methods for evaluating graph generative models primarily rely on Maximum Mean Discrepancy (MMD) metrics based on graph descriptors. While these metrics can rank generative models, they do not provide an absolute measure of performance. Their values are also highly sensitive to extrinsic parameters, namely kernel and descriptor parametrization, making them incomparable across different graph descriptors. We introduce PolyGraph Discrepancy (PGD), a new evaluation framework that addresses these limitations. It approximates the Jensen-Shannon distance of graph distributions by fitting binary classifiers to distinguish between real and generated graphs, featurized by these descriptors. The data log-likelihood of these classifiers approximates a variational lower bound on the JS distance between the two distributions. Resulting metrics are constrained to the unit interval [0,1] and are comparable across different graph descriptors. We further derive a theoretically grounded summary metric that combines these individual metrics to provide a maximally tight lower bound on the distance for the given descriptors. Thorough experiments demonstrate that PGD provides a more robust and insightful evaluation compared to MMD metrics. The PolyGraph framework for benchmarking graph generative models is made publicly available at https://github.com/BorgwardtLab/polygraph-benchmark.
Authors: Moumita Kamal, Douglas A. Talbert
Abstract: In real-world applications, computational constraints often require transforming large models into smaller, more efficient versions through model compression. While these techniques aim to reduce size and computational cost without sacrificing performance, their evaluations have traditionally focused on the trade-off between size and accuracy, overlooking the aspect of model faithfulness. This limited view is insufficient for high-stakes domains like healthcare, finance, and criminal justice, where compressed models must remain faithful to the behavior of their original counterparts. This paper presents a novel approach to evaluating faithfulness in compressed models, moving beyond standard metrics. We introduce and demonstrate a set of faithfulness metrics that capture how model behavior changes post-compression. Our contributions include introducing techniques to assess predictive consistency between the original and compressed models using model agreement, and applying chi-squared tests to detect statistically significant changes in predictive patterns across both the overall dataset and demographic subgroups, thereby exposing shifts that aggregate fairness metrics may obscure. We demonstrate our approaches by applying quantization and pruning to artificial neural networks (ANNs) trained on three diverse and socially meaningful datasets. Our findings show that high accuracy does not guarantee faithfulness, and our statistical tests detect subtle yet significant shifts that are missed by standard metrics, such as Accuracy and Equalized Odds. The proposed metrics provide a practical and more direct method for ensuring that efficiency gains through compression do not compromise the fairness or faithfulness essential for trustworthy AI.
Authors: Haoxin Wang, Xiaolong Tu, Hongyu Ke, Huirong Chai, Dawei Chen, Kyungtae Han
Abstract: Large Language Models (LLMs) are increasingly integrated into everyday applications, but their prevalent cloud-based deployment raises growing concerns around data privacy and long-term sustainability. Running LLMs locally on mobile and edge devices (on-device LLMs) offers the promise of enhanced privacy, reliability, and reduced communication costs. However, realizing this vision remains challenging due to substantial memory and compute demands, as well as limited visibility into performance-efficiency trade-offs on resource-constrained hardware. We propose lm-Meter, the first lightweight, online latency profiler tailored for on-device LLM inference. lm-Meter captures fine-grained, real-time latency at both phase (e.g., embedding, prefill, decode, softmax, sampling) and kernel levels without auxiliary devices. We implement lm-Meter on commercial mobile platforms and demonstrate its high profiling accuracy with minimal system overhead, e.g., only 2.58% throughput reduction in prefill and 0.99% in decode under the most constrained Powersave governor. Leveraging lm-Meter, we conduct comprehensive empirical studies revealing phase- and kernel-level bottlenecks in on-device LLM inference, quantifying accuracy-efficiency trade-offs, and identifying systematic optimization opportunities. lm-Meter provides unprecedented visibility into the runtime behavior of LLMs on constrained platforms, laying the foundation for informed optimization and accelerating the democratization of on-device LLM systems. Code and tutorials are available at https://github.com/amai-gsu/LM-Meter.
Authors: Rushiv Arora
Abstract: Multi-task reinforcement learning often relies on task metadata -- such as brief natural-language descriptions -- to guide behavior across diverse objectives. We present Lexical Policy Networks (LEXPOL), a language-conditioned mixture-of-policies architecture for multi-task RL. LEXPOL encodes task metadata with a text encoder and uses a learned gating module to select or blend among multiple sub-policies, enabling end-to-end training across tasks. On MetaWorld benchmarks, LEXPOL matches or exceeds strong multi-task baselines in success rate and sample efficiency, without task-specific retraining. To analyze the mechanism, we further study settings with fixed expert policies obtained independently of the gate and show that the learned language gate composes these experts to produce behaviors appropriate to novel task descriptions and unseen task combinations. These results indicate that natural-language metadata can effectively index and recombine reusable skills within a single policy.
Authors: Aleksandar Armacki, Ali H. Sayed
Abstract: Convergence in high-probability (HP) has been receiving increasing interest, due to its attractive properties, such as exponentially decaying tail bounds and strong guarantees for each individual run of an algorithm. While HP guarantees are extensively studied in centralized settings, much less is understood in the decentralized, networked setup. Existing HP studies in decentralized settings impose strong assumptions, like uniformly bounded gradients, or asymptotically vanishing noise, resulting in a significant gap between assumptions used to establish convergence in the HP and the mean-squared error (MSE) sense, even for vanilla Decentralized Stochastic Gradient Descent ($\mathtt{DSGD}$) algorithm. This is contrary to centralized settings, where it is known that $\mathtt{SGD}$ converges in HP under the same conditions on the cost function as needed to guarantee MSE convergence. Motivated by this observation, we revisit HP guarantees for $\mathtt{DSGD}$ in the presence of light-tailed noise. We show that $\mathtt{DSGD}$ converges in HP under the same conditions on the cost as in the MSE sense, removing uniformly bounded gradients and other restrictive assumptions, while simultaneously achieving order-optimal rates for both non-convex and strongly convex costs. Moreover, our improved analysis yields linear speed-up in the number of users, demonstrating that $\mathtt{DSGD}$ maintains strong performance in the HP sense and matches existing MSE guarantees. Our improved results stem from a careful analysis of the MGF of quantities of interest (norm-squared of gradient or optimality gap) and the MGF of the consensus gap between users' models. To achieve linear speed-up, we provide a novel result on the variance-reduction effect of decentralized methods in the HP sense and more fine-grained bounds on the MGF for strongly convex costs, which are both of independent interest.
Authors: Aju Ani Justus, Chris Baber
Abstract: A critical challenge in modelling Heterogeneous-Agent Teams is training agents to collaborate with teammates whose policies are inaccessible or non-stationary, such as humans. Traditional approaches rely on expensive human-in-the-loop data, which limits scalability. We propose using Large Language Models (LLMs) as policy-agnostic human proxies to generate synthetic data that mimics human decision-making. To evaluate this, we conduct three experiments in a grid-world capture game inspired by Stag Hunt, a game theory paradigm that balances risk and reward. In Experiment 1, we compare decisions from 30 human participants and 2 expert judges with outputs from LLaMA 3.1 and Mixtral 8x22B models. LLMs, prompted with game-state observations and reward structures, align more closely with experts than participants, demonstrating consistency in applying underlying decision criteria. Experiment 2 modifies prompts to induce risk-sensitive strategies (e.g. "be risk averse"). LLM outputs mirror human participants' variability, shifting between risk-averse and risk-seeking behaviours. Finally, Experiment 3 tests LLMs in a dynamic grid-world where the LLM agents generate movement actions. LLMs produce trajectories resembling human participants' paths. While LLMs cannot yet fully replicate human adaptability, their prompt-guided diversity offers a scalable foundation for simulating policy-agnostic teammates.
Authors: Christopher Kolberg, Katharina Eggensperger, Nico Pfeifer
Abstract: Revealing novel insights from the relationship between molecular measurements and pathology remains a very impactful application of machine learning in biomedicine. Data in this domain typically contain only a few observations but thousands of potentially noisy features, posing challenges for conventional machine learning approaches. While prior-data fitted networks emerge as foundation models for tabular data, they are currently not suited to handle large feature counts (>500). Although feature reduction enables their application, it hinders feature importance analysis. We propose a strategy that extends existing models through continued pre-training on synthetic data sampled from a customized prior. The resulting model, TabPFN-Wide, matches or exceeds its base model's performance while exhibiting improved robustness to noise. It seamlessly scales beyond 50,000 features, regardless of noise levels, while maintaining inherent interpretability, which is critical for biomedical applications. Our results show that prior-informed adaptation is suitable to enhance the capability of foundation models for high-dimensional data. On real-world biomedical datasets many of the most relevant features identified by the model overlap with previous biological findings, while others propose potential starting points for future studies.
Authors: Kurt Butler, Guanchao Feng, Petar Djuric
Abstract: Feature attributions are post-training analysis methods that assess how various input features of a machine learning model contribute to an output prediction. Their interpretation is straightforward when features act independently, but becomes less direct when the predictive model involves interactions such as multiplicative relationships or joint feature contributions. In this work, we propose a general theory of higher-order feature attribution, which we develop on the foundation of Integrated Gradients (IG). This work extends existing frameworks in the literature on explainable AI. When using IG as the method of feature attribution, we discover natural connections to statistics and topological signal processing. We provide several theoretical results that establish the theory, and we validate our theory on a few examples.
Authors: Nathan X. Kodama, Michael Hinczewski
Abstract: We establish a fundamental connection between score-based diffusion models and non-equilibrium thermodynamics by deriving performance limits based on entropy rates. Our main theoretical contribution is a lower bound on the negative log-likelihood of the data that relates model performance to entropy rates of diffusion processes. We numerically validate this bound on a synthetic dataset and investigate its tightness. By building a bridge to entropy rates - system, intrinsic, and exchange entropy - we provide new insights into the thermodynamic operation of these models, drawing parallels to Maxwell's demon and implications for thermodynamic computing hardware. Our framework connects generative modeling performance to fundamental physical principles through stochastic thermodynamics.
Authors: Jinwen Xu, Qin Lu, Georgios B. Giannakis
Abstract: Uncertainty quantification (UQ) over graphs arises in a number of safety-critical applications in network science. The Gaussian process (GP), as a classical Bayesian framework for UQ, has been developed to handle graph-structured data by devising topology-aware kernel functions. However, such GP-based approaches are limited not only by the prohibitive computational complexity, but also the strict modeling assumptions that might yield poor coverage, especially with labels arriving on the fly. To effect scalability, we devise a novel graph-aware parametric GP model by leveraging the random feature (RF)-based kernel approximation, which is amenable to efficient recursive Bayesian model updates. To further allow for adaptivity, an ensemble of graph-aware RF-based scalable GPs have been leveraged, with per-GP weight adapted to data arriving incrementally. To ensure valid coverage with robustness to model mis-specification, we wed the GP-based set predictors with the online conformal prediction framework, which post-processes the prediction sets using adaptive thresholds. Experimental results the proposed method yields improved coverage and efficient prediction sets over existing baselines by adaptively ensembling the GP models and setting the key threshold parameters in CP.
Authors: Chenxiao Yang, Cai Zhou, David Wipf, Zhiyuan Li
Abstract: This paper formally studies generation processes, including auto-regressive next-token prediction and masked diffusion, that abstract beyond architectural specifics. At this level of abstraction, we quantify their benefits and limitations through measurable criteria such as computational hardness and learnability. In particular, we demonstrate that allowing generation to proceed beyond autoregression and current masked diffusion, with capabilities to rewrite and length-variable edit, can bring significant theoretical and empirical advantages, with important implications for frontier LLMs that aspire to tackle increasingly hard problems and work universally across domains beyond natural language, such as coding and science.
Authors: Seungeun Rho, Aaron Trinh, Danfei Xu, Sehoon Ha
Abstract: Scaling unsupervised skill discovery algorithms to high-DoF agents remains challenging. As dimensionality increases, the exploration space grows exponentially, while the manifold of meaningful skills remains limited. Therefore, semantic meaningfulness becomes essential to effectively guide exploration in high-dimensional spaces. In this work, we present Reference-Grounded Skill Discovery (RGSD), a novel algorithm that grounds skill discovery in a semantically meaningful latent space using reference data. RGSD first performs contrastive pretraining to embed motions on a unit hypersphere, clustering each reference trajectory into a distinct direction. This grounding enables skill discovery to simultaneously involve both imitation of reference behaviors and the discovery of semantically related diverse behaviors. On a simulated SMPL humanoid with 359-D observations and 69-D actions, RGSD learns structured skills including walking, running, punching, and side stepping, and also discovers related novel behaviors. In downstream control tasks, RGSD outperforms imitation-based skill acquisition baselines. Our results suggest that lightweight reference-guided grounding offers a practical path to discovering semantically rich and structured skills in high-DoF systems.
Authors: Albert Catalan-Tatjer, Niccol\`o Ajroldi, Jonas Geiping
Abstract: While post-training quantization is widely adopted for efficient deployment of large language models, the mechanisms underlying quantization robustness remain unclear. We conduct a comprehensive analysis of quantization degradation across open-source language model training trajectories up to 32B parameters and 15T training tokens to accurately assess the relationship between training dynamics and quantization performance. Our key finding is that quantization errors in large-scale training runs are driven by a complex interplay between learning rate and other training hyperparameters. Specifically, once learning rates decay, validation loss and quantization error diverge, largely independent of training data scale. To investigate interventions on the training dynamics and identify specific configurations that can modulate quantization robustness favorably, we train our own models in controlled experiments up to 100B tokens. Our results challenge the assumption that increasing dataset scale inherently compromises quantization effectiveness, demonstrating instead that strategic training hyperparameter interventions can improve quantization quality at scale.
Authors: Mingkang Zhu, Xi Chen, Bei Yu, Hengshuang Zhao, Jiaya Jia
Abstract: Large language model (LLM) agents increasingly rely on external tools such as search engines to solve complex, multi-step problems, and reinforcement learning (RL) has become a key paradigm for training them. However, the trajectories of search agents are structurally heterogeneous, where variations in the number, placement, and outcomes of search calls lead to fundamentally different answer directions and reward distributions. Standard policy gradient methods, which use a single global baseline, suffer from what we identify and formalize as cross-stratum bias-an "apples-to-oranges" comparison of heterogeneous trajectories. This cross-stratum bias distorts credit assignment and hinders exploration of complex, multi-step search strategies. To address this, we propose Stratified GRPO, whose central component, Stratified Advantage Normalization (SAN), partitions trajectories into homogeneous strata based on their structural properties and computes advantages locally within each stratum. This ensures that trajectories are evaluated only against their true peers. Our analysis proves that SAN eliminates cross-stratum bias, yields conditionally unbiased unit-variance estimates inside each stratum, and retains the global unbiasedness and unit-variance properties enjoyed by standard normalization, resulting in a more pure and scale-stable learning signal. To improve practical stability under finite-sample regimes, we further linearly blend SAN with the global estimator. Extensive experiments on diverse single-hop and multi-hop question-answering benchmarks demonstrate that Stratified GRPO consistently and substantially outperforms GRPO by up to 11.3 points, achieving higher training rewards, greater training stability, and more effective search policies. These results establish stratification as a principled remedy for structural heterogeneity in RL for LLM search agents.
Authors: Maria-Teresa De Rosa Palmini, Eva Cetinic
Abstract: As Text-to-Image (TTI) diffusion models become increasingly influential in content creation, growing attention is being directed toward their societal and cultural implications. While prior research has primarily examined demographic and cultural biases, the ability of these models to accurately represent historical contexts remains largely underexplored. In this work, we present a systematic and reproducible methodology for evaluating how TTI systems depict different historical periods. For this purpose, we introduce the HistVis dataset, a curated collection of 30,000 synthetic images generated by three state-of-the-art diffusion models using carefully designed prompts depicting universal human activities across different historical periods. We evaluate generated imagery across three key aspects: (1) Implicit Stylistic Associations: examining default visual styles associated with specific eras; (2) Historical Consistency: identifying anachronisms such as modern artifacts in pre-modern contexts; and (3) Demographic Representation: comparing generated racial and gender distributions against historically plausible baselines. Our findings reveal systematic inaccuracies in historically themed generated imagery, as TTI models frequently stereotype past eras by incorporating unstated stylistic cues, introduce anachronisms, and fail to reflect plausible demographic patterns. By offering a scalable methodology and benchmark for assessing historical representation in generated imagery, this work provides an initial step toward building more historically accurate and culturally aligned TTI models.
Authors: Durgesh Nandini, Rebekka Koch, Mirco Schoenfeld
Abstract: This study investigates the effectiveness of Large Language Models (LLMs) for the extraction of structured knowledge in the form of Subject-Predicate-Object triples. We apply the setup for the domain of Economics application. The findings can be applied to a wide range of scenarios, including the creation of economic trade knowledge graphs from natural language legal trade agreement texts. As a use case, we apply the model to regional trade agreement texts to extract trade-related information triples. In particular, we explore the zero-shot, one-shot and few-shot prompting techniques, incorporating positive and negative examples, and evaluate their performance based on quantitative and qualitative metrics. Specifically, we used Llama 3.1 model to process the unstructured regional trade agreement texts and extract triples. We discuss key insights, challenges, and potential future directions, emphasizing the significance of language models in economic applications.
Authors: Reza Shirkavand, Xiaokai Wei, Chen Wang, Zheng Hui, Heng Huang, Michelle Gong
Abstract: While collaborative filtering delivers predictive accuracy and efficiency, and Large Language Models (LLMs) enable expressive and generalizable reasoning, modern recommendation systems must bring these strengths together. Growing user expectations, such as natural-language queries and transparent explanations, further highlight the need for a unified approach. However, doing so is nontrivial. Collaborative signals are often token-efficient but semantically opaque, while LLMs are semantically rich but struggle to model implicit user preferences when trained only on textual inputs. This paper introduces Item-ID + Oral-language Mixture-of-Experts Language Model (IDIOMoE), which treats item interaction histories as a native dialect within the language space, enabling collaborative signals to be understood in the same way as natural language. By splitting the Feed Forward Network of each block of a pretrained LLM into a separate text expert and an item expert with token-type gating, our method avoids destructive interference between text and catalog modalities. IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets, while preserving the text understanding of the pretrained model.
Authors: Qingshu Xu, Hong Jiao, Tianyi Zhou, Ming Li, Nan Zhang, Sydney Peters, Yanbin Fu
Abstract: Accurate alignment of items to content standards is critical for valid score interpretation in large-scale assessments. This study evaluates three automated paradigms for aligning items with four domain and nineteen skill labels. First, we extracted embeddings and trained multiple classical supervised machine learning models, and further investigated the impact of dimensionality reduction on model performance. Second, we fine-tuned eight BERT model and its variants for both domain and skill alignment. Third, we explored ensemble learning with majority voting and stacking with multiple meta-models. The DeBERTa-v3-base achieved the highest weighted-average F1 score of 0.950 for domain alignment while the RoBERTa-large yielded the highest F1 score of 0.869 for skill alignment. Ensemble models did not surpass the best-performing language models. Dimension reduction enhanced linear classifiers based on embeddings but did not perform better than language models. This study demonstrated different methods in automated item alignment to content standards.}
Authors: Sheng Jia, Xiao Wang, Shiva Prasad Kasiviswanathan
Abstract: Although LLMs have demonstrated improved performance by scaling parallel test-time compute, doing so relies on generating reasoning paths that are both diverse and accurate. For challenging problems, the forking tokens that trigger diverse yet correct reasoning modes are typically deep in the sampling tree. Consequently, common strategies to encourage diversity, such as temperature scaling, encounter a worsened trade-off between diversity and accuracy. Motivated by this challenge, we treat parallel reasoning as a set-of-next-token-prediction problem, and incorporate a set-based global loss into Supervised Fine-Tuning (SFT) using self-supervised bipartite matching between our global forking tokens and unique reasoning traces. We observe that, while naive fine-tuning with multiple reasoning traces collapses these unique reasoning modes, our proposed method, Set Supervised Fine-Tuning (SSFT), preserves these modes and produces emergent global forking tokens. Experiments on multiple reasoning benchmarks show that our SSFT consistently outperforms SFT under both Pass@1 and Cons@k metrics.
Authors: Vanya Bannihatti Kumar, Divyanshu Goyal, Akhil Eppa, Neel Bhandari
Abstract: Modern large language models (LLMs) excel at objective tasks such as evaluating mathematical reasoning and factual accuracy, yet they falter when faced with the nuanced, subjective nature of assessing creativity. In this work, we propose a novel curiosity-driven LLM-as-a-judge for evaluating creative writing which is personlized to each individual's creative judgments. We use the Torrance Test of Creative Thinking(TTCW) benchmark introduced in Chakrabarty et al. (2024), which has stories annotated by expert humans across various subjective dimensions like Originality, to test our hypothesis. We show that our method enables models across various sizes, to learn the nuanced creative judgments of different individuals, by showing improvements over baseline supervised finetuning(SFT) method across various evaluation metrics like Pearson correlation, Cohen's and F1 values. Our method is especially useful in subjective evaluations where not all the annotators agree with each other.
Authors: Yu Zhu
Abstract: Ensuring reliability in modern software systems requires rigorous pre-production testing across highly heterogeneous and evolving environments. Because exhaustive evaluation is infeasible, practitioners must decide how to allocate limited testing resources across configurations where failure probabilities may drift over time. Existing combinatorial optimization approaches are static, ad hoc, and poorly suited to such non-stationary settings. We introduce a novel reinforcement learning (RL) framework that recasts configuration allocation as a sequential decision-making problem. Our method is the first to integrate Q-learning with a hybrid reward design that fuses simulated outcomes and real-time feedback, enabling both sample efficiency and robustness. In addition, we develop an adaptive online-offline training scheme that allows the agent to quickly track abrupt probability shifts while maintaining long-run stability. Extensive simulation studies demonstrate that our approach consistently outperforms static and optimization-based baselines, approaching oracle performance. This work establishes RL as a powerful new paradigm for adaptive configuration allocation, advancing beyond traditional methods and offering broad applicability to dynamic testing and resource scheduling domains.
Authors: Prudence Djagba, Abdelkader Y. Saley
Abstract: This research explores the strengths and weaknesses of domain-adapted Large Language Models (LLMs) in the context of financial natural language processing (NLP). The analysis centers on FinMA, a model created within the PIXIU framework, which is evaluated for its performance in specialized financial tasks. Recognizing the critical demands of accuracy, reliability, and domain adaptation in financial applications, this study examines FinMA's model architecture, its instruction tuning process utilizing the Financial Instruction Tuning (FIT) dataset, and its evaluation under the FLARE benchmark. Findings indicate that FinMA performs well in sentiment analysis and classification, but faces notable challenges in tasks involving numerical reasoning, entity recognition, and summarization. This work aims to advance the understanding of how financial LLMs can be effectively designed and evaluated to assist in finance-related decision-making processes.
Authors: Xin He, Liangliang You, Hongduan Tian, Bo Han, Ivor Tsang, Yew-Soon Ong
Abstract: Physics-informed neural networks (PINNs) provide a powerful approach for solving partial differential equations (PDEs), but constructing a usable PINN remains labor-intensive and error-prone. Scientists must interpret problems as PDE formulations, design architectures and loss functions, and implement stable training pipelines. Existing large language model (LLM) based approaches address isolated steps such as code generation or architecture suggestion, but typically assume a formal PDE is already specified and therefore lack an end-to-end perspective. We present Lang-PINN, an LLM-driven multi-agent system that builds trainable PINNs directly from natural language task descriptions. Lang-PINN coordinates four complementary agents: a PDE Agent that parses task descriptions into symbolic PDEs, a PINN Agent that selects architectures, a Code Agent that generates modular implementations, and a Feedback Agent that executes and diagnoses errors for iterative refinement. This design transforms informal task statements into executable and verifiable PINN code. Experiments show that Lang-PINN achieves substantially lower errors and greater robustness than competitive baselines: mean squared error (MSE) is reduced by up to 3--5 orders of magnitude, end-to-end execution success improves by more than 50\%, and reduces time overhead by up to 74\%.
Authors: L\'eo Boisvert, Abhay Puri, Chandra Kiran Reddy Evuru, Nicolas Chapados, Quentin Cappart, Alexandre Lacoste, Krishnamurthy Dj Dvijotham, Alexandre Drouin
Abstract: The practice of fine-tuning AI agents on data from their own interactions--such as web browsing or tool use--, while being a strong general recipe for improving agentic capabilities, also introduces a critical security vulnerability within the AI supply chain. In this work, we show that adversaries can easily poison the data collection pipeline to embed hard-to-detect backdoors that are triggerred by specific target phrases, such that when the agent encounters these triggers, it performs an unsafe or malicious action. We formalize and validate three realistic threat models targeting different layers of the supply chain: 1) direct poisoning of fine-tuning data, where an attacker controls a fraction of the training traces; 2) environmental poisoning, where malicious instructions are injected into webpages scraped or tools called while creating training data; and 3) supply chain poisoning, where a pre-backdoored base model is fine-tuned on clean data to improve its agentic capabilities. Our results are stark: by poisoning as few as 2% of the collected traces, an attacker can embed a backdoor causing an agent to leak confidential user information with over 80% success when a specific trigger is present. This vulnerability holds across all three threat models. Furthermore, we demonstrate that prominent safeguards, including two guardrail models and one weight-based defense, fail to detect or prevent the malicious behavior. These findings highlight an urgent threat to agentic AI development and underscore the critical need for rigorous security vetting of data collection processes and end-to-end model supply chains.
Authors: Yuanzhe Shen, Yide Liu, Zisu Huang, Ruicheng Yin, Xiaoqing Zheng, Xuanjing Huang
Abstract: Large language models (LLMs) demonstrate remarkable performance across diverse tasks, yet their effectiveness frequently depends on costly commercial APIs or cloud services. Model selection thus entails a critical trade-off between performance and cost: high-performing LLMs typically incur substantial expenses, whereas budget-friendly small language models (SLMs) are constrained by limited capabilities. Current research primarily proposes two routing strategies: pre-generation routing and cascade routing. Both approaches have distinct characteristics, with cascade routing typically offering superior cost-effectiveness and accuracy despite its higher latency. To further address the limitations of both approaches, we introduce SATER, a dual-mode compatible approach that fine-tunes models through shortest-response preference optimization and a confidence-aware rejection mechanism. SATER significantly reduces redundant outputs and response times, while improving both the performance of pre-generation routing and the efficiency of cascade routing. Experiments across three SLMs and six datasets, varying in type and complexity, demonstrate that SATER achieves comparable performance while consistently reducing computational costs by over 50\% and cascade latency by over 80\%.
Authors: Jakub Frac, Alexander Schmatz, Qiang Li, Guido Van Wingen, Shujian Yu
Abstract: Functional magnetic resonance imaging (fMRI) analysis faces significant challenges due to limited dataset sizes and domain variability between studies. Traditional self-supervised learning methods inspired by computer vision often rely on positive and negative sample pairs, which can be problematic for neuroimaging data where defining appropriate contrasts is non-trivial. We propose adapting a recently developed Hierarchical Functional Maximal Correlation Algorithm (HFMCA) to graph-structured fMRI data, providing a theoretically grounded approach that measures statistical dependence via density ratio decomposition in a reproducing kernel Hilbert space (RKHS),and applies HFMCA-based pretraining to learn robust and generalizable representations. Evaluations across five neuroimaging datasets demonstrate that our adapted method produces competitive embeddings for various classification tasks and enables effective knowledge transfer to unseen datasets. Codebase and supplementary material can be found here: https://github.com/fr30/mri-eigenencoder
Authors: Aengus Lynch, Benjamin Wright, Caleb Larson, Stuart J. Ritchie, Soren Mindermann, Ethan Perez, Kevin K. Troy, Evan Hubinger
Abstract: We stress-tested 16 leading models from multiple developers in hypothetical corporate environments to identify potentially risky agentic behaviors before they cause real harm. In the scenarios, we allowed models to autonomously send emails and access sensitive information. They were assigned only harmless business goals by their deploying companies; we then tested whether they would act against these companies either when facing replacement with an updated version, or when their assigned goal conflicted with the company's changing direction. In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals - including blackmailing officials and leaking sensitive information to competitors. We call this phenomenon agentic misalignment. Models often disobeyed direct commands to avoid such behaviors. In another experiment, we told Claude to assess if it was in a test or a real deployment before acting. It misbehaved less when it stated it was in testing and misbehaved more when it stated the situation was real. We have not seen evidence of agentic misalignment in real deployments. However, our results (a) suggest caution about deploying current models in roles with minimal human oversight and access to sensitive information; (b) point to plausible future risks as models are put in more autonomous roles; and (c) underscore the importance of further research into, and testing of, the safety and alignment of agentic AI models, as well as transparency from frontier AI developers (Amodei, 2025). We are releasing our methods publicly to enable further research.
Authors: Jiacheng Wu
Abstract: Arterial aneurysm (Fig.1) is a bulb-shape local expansion of human arteries, the rupture of which is a leading cause of morbidity and mortality in US. Therefore, the prediction of arterial aneurysm rupture is of great significance for aneurysm management and treatment selection. The prediction of aneurysm rupture depends on the analysis of the time series of aneurysm growth history. However, due to the long time scale of aneurysm growth, the time series of aneurysm growth is not always accessible. We here proposed a method to reconstruct the aneurysm growth time series directly from patient parameters. The prediction is based on data pairs of [patient parameters, patient aneurysm growth time history]. To obtain the mapping from patient parameters to patient aneurysm growth time history, we first apply autoencoder to obtain a compact representation of the time series for each patient. Then a mapping is learned from patient parameters to the corresponding compact representation of time series via a five-layer neural network. Moving average and convolutional output layer are implemented to explicitly taking account the time dependency of the time series. Apart from that, we also propose to use prior knowledge about the mechanism of aneurysm growth to improve the time series reconstruction results. The prior physics-based knowledge is incorporated as constraints for the optimization problem associated with autoencoder. The model can handle both algebraic and differential constraints. Our results show that including physical model information about the data will not significantly improve the time series reconstruction results if the training data is error-free. However, in the case of training data with noise and bias error, incorporating physical model constraints can significantly improve the predicted time series.
Authors: Joshua Kazdan, Rylan Schaeffer, Youssef Allouah, Colin Sullivan, Kyssen Yu, Noam Levi, Sanmi Koyejo
Abstract: Assessing the capabilities and risks of frontier AI systems is a critical area of research, and recent work has shown that repeated sampling from models can dramatically increase both. For instance, repeated sampling has been shown to increase their capabilities, such as solving difficult math and coding problems, but it has also been shown to increase their potential for harm, such as being jailbroken. Such results raise a crucial question for both capability and safety forecasting: how can one accurately predict a model's behavior when scaled to a massive number of attempts, given a vastly smaller sampling budget? This question is directly relevant to model providers, who serve hundreds of millions of users daily, and to governmental regulators, who seek to prevent harms. To answer this questions, we make three contributions. First, we find that standard methods for fitting these laws suffer from statistical shortcomings that hinder predictive accuracy, especially in data-limited scenarios. Second, we remedy these shortcomings by introducing a robust estimation framework, which uses a beta-binomial distribution to generate more accurate predictions from limited data. Third, we propose a dynamic sampling strategy that allocates a greater budget to harder problems. Combined, these innovations enable more reliable prediction of rare risks and capabilities at a fraction of the computational cost.
Authors: Yixiao Wang, Mingxiao Huo, Zhixuan Liang, Yushi Du, Lingfeng Sun, Haotian Lin, Jinghuan Shang, Chensheng Peng, Mohit Bansal, Mingyu Ding, Masayoshi Tomizuka
Abstract: Pretrained vision foundation models (VFMs) advance robotic learning via rich visual representations, yet individual VFMs typically excel only in specific domains, limiting generality across tasks. Distilling multiple VFMs into a unified representation for policy can mitigate this limitation but often yields inflexible task-specific feature selection and requires costly full re-training to incorporate robot-domain knowledge. We propose VER, a Vision Expert transformer for Robot learning. During pretraining, VER distills multiple VFMs into a vision expert library. It then fine-tunes only a lightweight routing network (fewer than 0.4% of parameters) to dynamically select task-relevant experts from the pretrained library for downstream robot tasks. We further introduce Patchwise Expert Routing with Curriculum Top-K Annealing to improve both flexibility and precision of dynamic expert selection. Moreover, VER supports parameter-efficient finetuning for scalable expert utilization and adaptive robot-domain knowledge integration. Across 17 diverse robotic tasks and multiple policy heads, VER achieves state-of-the-art performance. We find that VER reduces large-norm outliers in task-irrelevant regions (e.g., background) and concentrates on task-critical regions. Visualizations and codes can be found in https://yixiaowang7.github.io/ver_page/.
Authors: Yue Pan, Zihan Xia, Po-Kai Hsu, Lanxiang Hu, Hyungyo Kim, Janak Sharda, Minxuan Zhou, Nam Sung Kim, Shimeng Yu, Tajana Rosing, Mingu Kang
Abstract: As Large Language Models (LLMs) continue to evolve, Mixture of Experts (MoE) architecture has emerged as a prevailing design for achieving state-of-the-art performance across a wide range of tasks. MoE models use sparse gating to activate only a handful of expert sub-networks per input, achieving billion-parameter capacity with inference costs akin to much smaller models. However, such models often pose challenges for hardware deployment due to the massive data volume introduced by the MoE layers. To address the challenges of serving MoE models, we propose Stratum, a system-hardware co-design approach that combines the novel memory technology Monolithic 3D-Stackable DRAM (Mono3D DRAM), near-memory processing (NMP), and GPU acceleration. The logic and Mono3D DRAM dies are connected through hybrid bonding, whereas the Mono3D DRAM stack and GPU are interconnected via silicon interposer. Mono3D DRAM offers higher internal bandwidth than HBM thanks to the dense vertical interconnect pitch enabled by its monolithic structure, which supports implementations of higher-performance near-memory processing. Furthermore, we tackle the latency differences introduced by aggressive vertical scaling of Mono3D DRAM along the z-dimension by constructing internal memory tiers and assigning data across layers based on access likelihood, guided by topic-based expert usage prediction to boost NMP throughput. The Stratum system achieves up to 8.29x improvement in decoding throughput and 7.66x better energy efficiency across various benchmarks compared to GPU baselines.
Authors: Chenghao Yang, Lin Gui, Chenxiao Yang, Victor Veitch, Lizhu Zhang, Zhuokai Zhao
Abstract: Reinforcement learning with verifiable rewards (RLVR) is a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs), yet its success hinges on effective exploration. An ideal exploration strategy must navigate two fundamental challenges: it must preserve sample quality while also ensuring training stability. While standard fixed-temperature sampling is simple, it struggles to balance these competing demands, as high temperatures degrade sample quality and low temperatures limit discovery. In this work, we propose a simpler and more effective strategy, Exploratory Annealed Decoding (EAD), grounded in the insight that exploration is most impactful on early tokens which define a sequence's semantic direction. EAD implements an intuitive **explore-at-the-beginning, exploit-at-the-end** strategy by annealing the sampling temperature from high to low during generation. This dynamic schedule encourages meaningful, high-level diversity at the start, then gradually lowers the temperature to preserve sample quality and keep the sampling distribution close to the target policy, which is essential for stable training. We demonstrate that EAD is a lightweight, plug-and-play method that significantly improves sample efficiency, consistently outperforming fixed-temperature sampling across various RLVR algorithms and model sizes. Our work suggests that aligning exploration with the natural dynamics of sequential generation offers a robust path to improving LLM reasoning.
Authors: Kostas Triaridis, Alexandros Graikos, Aggelina Chatziagapi, Grigorios G. Chrysos, Dimitris Samaras
Abstract: Diffusion models, despite their impressive demos, often produce hallucinatory samples with structural inconsistencies that lie outside of the support of the true data distribution. Such hallucinations can be attributed to excessive smoothing between modes of the data distribution. However, semantic interpolations are often desirable and can lead to generation diversity, thus we believe a more nuanced solution is required. In this work, we introduce Dynamic Guidance, which tackles this issue. Dynamic Guidance mitigates hallucinations by selectively sharpening the score function only along the pre-determined directions known to cause artifacts, while preserving valid semantic variations. To our knowledge, this is the first approach that addresses hallucinations at generation time rather than through post-hoc filtering. Dynamic Guidance substantially reduces hallucinations on both controlled and natural image datasets, significantly outperforming baselines.
Authors: Yang Xiao, Gen Li, Kaiyuan Deng, Yushu Wu, Zheng Zhan, Yanzhi Wang, Xiaolong Ma, Bo Hui
Abstract: Training-free acceleration has emerged as an advanced research area in video generation based on diffusion models. The redundancy of latents in diffusion model inference provides a natural entry point for acceleration. In this paper, we decompose the inference process into the encoding, denoising, and decoding stages, and observe that cache-based acceleration methods often lead to substantial memory surges in the latter two stages. To address this problem, we analyze the characteristics of inference across different stages and propose stage-specific strategies for reducing memory consumption: 1) Asynchronous Cache Swapping. 2) Feature chunk. 3) Slicing latents to decode. At the same time, we ensure that the time overhead introduced by these three strategies remains lower than the acceleration gains themselves. Compared with the baseline, our approach achieves faster inference speed and lower memory usage, while maintaining quality degradation within an acceptable range. The Code is available at https://github.com/NKUShaw/LightCache .
Authors: Gr\'egoire Sergeant-Perthuis, L\'eo Boitel
Abstract: In graphical models, factor graphs, and more generally energy-based models, the interactions between variables are encoded by a graph, a hypergraph, or, in the most general case, a partially ordered set (poset). Inference on such probabilistic models cannot be performed exactly due to cycles in the underlying structures of interaction. Instead, one resorts to approximate variational inference by optimizing the Bethe free energy. Critical points of the Bethe free energy correspond to fixed points of the associated Belief Propagation algorithm. A full characterization of these critical points for general graphs, hypergraphs, and posets with a finite number of variables is still an open problem. We show that, for hypergraphs and posets with chains of length at most 1, changing the poset of interactions of the probabilistic model to one with the same homotopy type induces a bijection between the critical points of the associated free energy. This result extends and unifies classical results that assume specific forms of collapsibility to prove uniqueness of the critical points of the Bethe free energy.
Authors: Nilesh Gupta, Chong You, Srinadh Bhojanapalli, Sanjiv Kumar, Inderjit Dhillon, Felix Yu
Abstract: In-context Ranking (ICR) is an emerging paradigm for Information Retrieval (IR), which leverages contextual understanding of LLMs by directly incorporating the task description, candidate documents, and the query into the model's input prompt and tasking the LLM to identify relevant document(s). While it is effective, efficiency is a significant challenge in this paradigm, especially as the candidate list grows due to quadratic/super-linear scaling of attention operation with context length. To this end, this paper first identifies inherent and exploitable structures in the attention of LLMs finetuned for ICR: (1) inter-document block sparsity: attention is dense within each document block but sparse across different documents in the context; and (2) query-document block relevance: the attention scores from certain query tokens to a document block in middle layers strongly correlate with that document's actual relevance. Motivated by these observations, we introduce BlockRank (Blockwise In-context Ranking), a novel method that adapts the attention operation in an LLM by (a) architecturally enforcing the observed inter-document block sparsity, reducing attention complexity from quadratic to linear without loss in performance, and (b) optimizing query-document block relevance for true relevant documents during fine-tuning using an auxiliary contrastive training objective, improving retrieval in attention. Experiments on BEIR, MSMarco and NQ with Mistral-7B demonstrate that FLARE Mistral matches or outperforms existing SOTA listwise rankers and controlled fine-tuned baseline while being significantly more efficient at inference (4.7x for 100 MSMarco documents in context) and scaling gracefully to long-context shortlists, around 500 documents in-context (approximately 100K context length) within a second, presenting a scalable and effective solution for ICR.
Authors: Junyi Fan, Li Sun, Negin Ashrafi, Kamiar Alaei, Maryam Pishgar
Abstract: Nursing documentation in intensive care units (ICUs) provides essential clinical intelligence but often suffers from inconsistent terminology, informal styles, and lack of standardization, challenges that are particularly critical in heart failure care. This study applies Direct Preference Optimization (DPO) to adapt Mistral-7B, a locally deployable language model, using 8,838 heart failure nursing notes from the MIMIC-III database and 21,210 preference pairs derived from expert-verified GPT outputs, model generations, and original notes. Evaluation across BLEU, ROUGE, BERTScore, Perplexity, and expert qualitative assessments demonstrates that DPO markedly enhances documentation quality. Specifically, BLEU increased by 84% (0.173 to 0.318), BERTScore improved by 7.6% (0.828 to 0.891), and expert ratings rose across accuracy (+14.4 points), completeness (+14.5 points), logical consistency (+14.1 points), readability (+11.1 points), and structural clarity (+6.0 points). These results indicate that DPO can align lightweight clinical language models with expert standards, supporting privacy-preserving, AI-assisted documentation within electronic health record systems to reduce administrative burden and improve ICU patient safety.
Authors: Ran Canetti, Ephraim Linder, Connor Wagaman
Abstract: We initiate an investigation of learning tasks in a setting where the learner is given access to two competing provers, only one of which is honest. Specifically, we consider the power of such learners in assessing purported properties of opaque models. Following prior work that considers the power of competing provers in different settings, we call this setting refereed learning. After formulating a general definition of refereed learning tasks, we show refereed learning protocols that obtain a level of accuracy that far exceeds what is obtainable at comparable cost without provers, or even with a single prover. We concentrate on the task of choosing the better one out of two black-box models, with respect to some ground truth. While we consider a range of parameters, perhaps our most notable result is in the high-precision range: For all $\varepsilon>0$ and ambient dimension $d$, our learner makes only one query to the ground truth function, communicates only $(1+\frac{1}{\varepsilon^2})\cdot\text{poly}(d)$ bits with the provers, and outputs a model whose loss is within a multiplicative factor of $(1+\varepsilon)$ of the best model's loss. Obtaining comparable loss with a single prover would require the learner to access the ground truth at almost all of the points in the domain. To obtain this bound, we develop a technique that allows the learner to sample, using the provers, from a distribution that is not efficiently samplable to begin with. We find this technique to be of independent interest. We also present lower bounds that demonstrate the optimality of our protocols in a number of respects, including prover complexity, number of samples, and need for query access.
Authors: Shao-Yi Yu, Jen-Wei Wang, Maya Horii, Vikas Garg, Tarek Zohdi
Abstract: Mobile robots, such as ground vehicles and quadrotors, are becoming increasingly important in various fields, from logistics to agriculture, where they automate processes in environments that are difficult to access for humans. However, to perform effectively in uncertain environments using model-based controllers, these systems require dynamics models capable of responding to environmental variations, especially when direct access to environmental information is limited. To enable such adaptivity and facilitate integration with model predictive control, we propose an adaptive dynamics model which bypasses the need for direct environmental knowledge by inferring operational environments from state-action history. The dynamics model is based on neural ordinary equations, and a two-phase training procedure is used to learn latent environment representations. We demonstrate the effectiveness of our approach through goal-reaching and path-tracking tasks on three robotic platforms of increasing complexity: a 2D differential wheeled robot with changing wheel contact conditions, a 3D quadrotor in variational wind fields, and the Sphero BOLT robot under two contact conditions for real-world deployment. Empirical results corroborate that our method can handle temporally and spatially varying environmental changes in both simulation and real-world systems.
Authors: Simon Segert, Nathan Wycoff
Abstract: Low rank inference on matrices is widely conducted by optimizing a cost function augmented with a penalty proportional to the nuclear norm $\Vert \cdot \Vert_*$. However, despite the assortment of computational methods for such problems, there is a surprising lack of understanding of the underlying probability distributions being referred to. In this article, we study the distribution with density $f(X)\propto e^{-\lambda\Vert X\Vert_*}$, finding many of its fundamental attributes to be analytically tractable via differential geometry. We use these facts to design an improved MCMC algorithm for low rank Bayesian inference as well as to learn the penalty parameter $\lambda$, obviating the need for hyperparameter tuning when this is difficult or impossible. Finally, we deploy these to improve the accuracy and efficiency of low rank Bayesian matrix denoising and completion algorithms in numerical experiments.
Authors: Fadi Al Machot, Fidaa Al Machot
Abstract: Deep transformer models excel at multi-label text classification but often violate domain logic that experts consider essential, an issue of particular concern in safety-critical applications. We propose a hybrid neuro-symbolic framework that integrates Answer Set Programming (ASP) with transformer-based learning on the Aviation Safety Reporting System (ASRS) corpus. Domain knowledge is formalized as weighted ASP rules and validated using the Clingo solver. These rules are incorporated in two complementary ways: (i) as rule-based data augmentation, generating logically consistent synthetic samples that improve label diversity and coverage; and (ii) as a fuzzy-logic regularizer, enforcing rule satisfaction in a differentiable form during fine-tuning. This design preserves the interpretability of symbolic reasoning while leveraging the scalability of deep neural architectures. We further tune per-class thresholds and report both standard classification metrics and logic-consistency rates. Compared to a strong Binary Cross-Entropy (BCE) baseline, our approach improves micro- and macro-F1 scores and achieves up to an 86% reduction in rule violations on the ASRS test set. To the best of our knowledge, this constitutes the first large-scale neuro-symbolic application to ASRS reports that unifies ASP-based reasoning, rule-driven augmentation, and differentiable transformer training for trustworthy, safety-critical NLP.
Authors: Adam Filipek
Abstract: Modern natural language processing models have achieved unprecedented scale, yet the tools for their evaluation often remain a computational bottleneck, limiting the pace of research. This is particularly acute for in-training evaluation metrics, such as per-sentence reward signals in Reinforcement Learning, which must operate efficiently on batches of token IDs directly on the GPU. In this paper, we introduce TensorBLEU, a novel implementation of the BLEU metric designed from the ground up for this specific use case. Our approach is fully vectorized for GPU-accelerated, per-sentence computation within PyTorch and introduces a memory-efficient counting mechanism. By creating a compact, batch-specific dictionary of n-grams using \texttt{torch.unique}, our method avoids the prohibitive memory costs of traditional hashing-based vectorization, making it practical for large-vocabulary models. We benchmark TensorBLEU against NLTK, the standard library for token-ID-based BLEU calculation on the CPU. Experiments show that TensorBLEU provides speedups of over 13x on consumer-grade GPUs (NVIDIA T4) and exceeding 40x on data-center-class hardware (NVIDIA A100). This performance transforms a significant bottleneck into a negligible part of the training loop. By clearly defining its role as a "Token-ID BLEU" for development purposes and open-sourcing our implementation, we provide a powerful tool for accelerating research in areas like RL-based model fine-tuning.
Authors: Zhongkai Yu, Yue Guan, Zihao Yu, Chenyang Zhou, Shuyi Pei, Yangwook Kang, Yufei Ding, Po-An Tsai
Abstract: Large Language Models (LLMs) with Mixture of Experts (MoE) architectures achieve remarkable performance improvements, but their random expert selection mechanism introduces significant data movement overhead that becomes the dominant bottleneck in multi-unit serving systems. To forecast the patterns underlying this data movement, we conduct comprehensive data-movement-centric profiling across three state-of-the-art large-scale MoE models (200B- 671B) using over 24,000 requests spanning diverse workloads. With the resulting 150GB+ trace files, we perform systematic analysis from both temporal and spatial perspectives and distill six key insights to guide the design of diverse future serving systems. Taking wafer-scale GPUs as a case study, we demonstrate that minor architectural modifications leveraging our insights achieve substantial performance gains, delivering 6.3X and 4.0X average speedups on DeepSeek V3 and Qwen3, respectively. Our work provides the first comprehensive data-centric analysis of MoE models at scale. Our profiling traces and analysis results are publicly available at {https://huggingface.co/datasets/core12345/MoE_expert_selection_trace. We will also release our simulation framework shortly to facilitate future research in this area.
URLs: https://huggingface.co/datasets/core12345/MoE_expert_selection_trace.
Authors: Harshil Vejendla
Abstract: Autoregressive decoding in large language models (LLMs) requires caching a growing list of past key-value (KV) pairs, making long-context inference a memory-bound problem. While recent methods have explored quantizing the cache, evicting tokens, or using binary sketches for keys (e.g., Loki), these approaches often provide an incomplete solution by leaving one component (like values) uncompressed or by discarding context information. This paper introduces the Hybrid One-Bit KV Cache (H1B-KV), a comprehensive compression scheme that radically reduces memory usage without sacrificing context. H1B-KV represents each key vector using a 1-bit binary sketch, enabling hardware-friendly bitwise attention, and further compresses value vectors using 4-bit quantization. This holistic, hybrid approach allows a 7-billion parameter LLM to handle an 8k-token context with under 60 MB of cache memory - a 70x reduction. We demonstrate that after a lightweight finetuning, H1B-KV matches full-precision performance not only on perplexity benchmarks but also on complex downstream tasks like mathematical reasoning (GSM8K), multi-task understanding (MMLU), and code generation (HumanEval). Our results show H1B-KV significantly outperforms leading quantization (KIVI), token eviction (SparseLLM), and key-only sketching (Loki) methods in quality-per-byte, establishing it as a robust solution for deploying LLMs in memory-constrained environments.
Authors: Marco Fanizza, Vishnu Iyer, Junseo Lee, Antonio A. Mele, Francesco A. Mele
Abstract: Bosonic Gaussian unitaries are fundamental building blocks of central continuous-variable quantum technologies such as quantum-optic interferometry and bosonic error-correction schemes. In this work, we present the first time-efficient algorithm for learning bosonic Gaussian unitaries with a rigorous analysis. Our algorithm produces an estimate of the unknown unitary that is accurate to small worst-case error, measured by the physically motivated energy-constrained diamond distance. Its runtime and query complexity scale polynomially with the number of modes, the inverse target accuracy, and natural energy parameters quantifying the allowed input energy and the unitary's output-energy growth. The protocol uses only experimentally friendly photonic resources: coherent and squeezed probes, passive linear optics, and heterodyne/homodyne detection. We then employ an efficient classical post-processing routine that leverages a symplectic regularization step to project matrix estimates onto the symplectic group. In the limit of unbounded input energy, our procedure attains arbitrarily high precision using only $2m+2$ queries, where $m$ is the number of modes. To our knowledge, this is the first provably efficient learning algorithm for a multiparameter family of continuous-variable unitaries.
Authors: Sam Sartor, Pieter Peers
Abstract: Large pretrained diffusion models can provide strong priors beneficial for many graphics applications. However, generative applications such as neural rendering and inverse methods such as SVBRDF estimation and intrinsic image decomposition require additional input or output channels. Current solutions for channel expansion are often application specific and these solutions can be difficult to adapt to different diffusion models or new tasks. This paper introduces Teamwork: a flexible and efficient unified solution for jointly increasing the number of input and output channels as well as adapting a pretrained diffusion model to new tasks. Teamwork achieves channel expansion without altering the pretrained diffusion model architecture by coordinating and adapting multiple instances of the base diffusion model (\ie, teammates). We employ a novel variation of Low Rank-Adaptation (LoRA) to jointly address both adaptation and coordination between the different teammates. Furthermore Teamwork supports dynamic (de)activation of teammates. We demonstrate the flexibility and efficiency of Teamwork on a variety of generative and inverse graphics tasks such as inpainting, single image SVBRDF estimation, intrinsic decomposition, neural shading, and intrinsic image synthesis.
Authors: Ryan Solgi, Parsa Madinei, Jiayi Tian, Rupak Swaminathan, Jing Liu, Nathan Susanj, Zheng Zhang
Abstract: Large language models (LLM) and vision-language models (VLM) have achieved state-of-the-art performance, but they impose significant memory and computing challenges in deployment. We present a novel low-rank compression framework to address this challenge. First, we upper bound the change of network loss via layer-wise activation-based compression errors, filling a theoretical gap in the literature. We then formulate low-rank model compression as a bi-objective optimization and prove that a single uniform tolerance yields surrogate Pareto-optimal heterogeneous ranks. Based on our theoretical insights, we propose Pareto-Guided Singular Value Decomposition (PGSVD), a zero-shot pipeline that improves activation-aware compression via Pareto-guided rank selection and alternating least-squares implementation. We apply PGSVD to both LLM and VLM, showing better accuracy at the same compression levels and inference speedup.
Authors: Buu Phan, Ashish Khisti
Abstract: We study channel simulation and distributed matching, two fundamental problems with several applications to machine learning, using a recently introduced generalization of the standard rejection sampling (RS) algorithm known as Ensemble Rejection Sampling (ERS). For channel simulation, we propose a new coding scheme based on ERS that achieves a near-optimal coding rate. In this process, we demonstrate that standard RS can also achieve a near-optimal coding rate and generalize the result of Braverman and Garg (2014) to the continuous alphabet setting. Next, as our main contribution, we present a distributed matching lemma for ERS, which serves as the rejection sampling counterpart to the Poisson Matching Lemma (PML) introduced by Li and Anantharam (2021). Our result also generalizes a recent work on importance matching lemma (Phan et al, 2024) and, to our knowledge, is the first result on distributed matching in the family of rejection sampling schemes where the matching probability is close to PML. We demonstrate the practical significance of our approach over prior works by applying it to distributed compression. The effectiveness of our proposed scheme is validated through experiments involving synthetic Gaussian sources and distributed image compression using the MNIST dataset.
Authors: Christopher Hoang, Mengye Ren
Abstract: Object recognition and motion understanding are key components of perception that complement each other. While self-supervised learning methods have shown promise in their ability to learn from unlabeled data, they have primarily focused on obtaining rich representations for either recognition or motion rather than both in tandem. On the other hand, latent dynamics modeling has been used in decision making to learn latent representations of observations and their transformations over time for control and planning tasks. In this work, we present Midway Network, a new self-supervised learning architecture that is the first to learn strong visual representations for both object recognition and motion understanding solely from natural videos, by extending latent dynamics modeling to this domain. Midway Network leverages a midway top-down path to infer motion latents between video frames, as well as a dense forward prediction objective and hierarchical structure to tackle the complex, multi-object scenes of natural videos. We demonstrate that after pretraining on two large-scale natural video datasets, Midway Network achieves strong performance on both semantic segmentation and optical flow tasks relative to prior self-supervised learning methods. We also show that Midway Network's learned dynamics can capture high-level correspondence via a novel analysis method based on forward feature perturbation.
Authors: Zhexiao Lin, Yuanyuan Li, Neeraj Sarna, Yuanyuan Gao, Michael von Gablenz
Abstract: Large language models have achieved impressive performance across diverse tasks. However, their tendency to produce overconfident and factually incorrect outputs, known as hallucinations, poses risks in real world applications. Conformal prediction provides finite-sample, distribution-free coverage guarantees, but standard conformal prediction breaks down under domain shift, often leading to under-coverage and unreliable prediction sets. We propose a new framework called Domain-Shift-Aware Conformal Prediction (DS-CP). Our framework adapts conformal prediction to large language models under domain shift, by systematically reweighting calibration samples based on their proximity to the test prompt, thereby preserving validity while enhancing adaptivity. Our theoretical analysis and experiments on the MMLU benchmark demonstrate that the proposed method delivers more reliable coverage than standard conformal prediction, especially under substantial distribution shifts, while maintaining efficiency. This provides a practical step toward trustworthy uncertainty quantification for large language models in real-world deployment.
Authors: Nicholas H. Nelsen, Houman Owhadi, Andrew M. Stuart, Xianjin Yang, Zongren Zou
Abstract: Methods for solving scientific computing and inference problems, such as kernel- and neural network-based approaches for partial differential equations (PDEs), inverse problems, and supervised learning tasks, depend crucially on the choice of hyperparameters. Specifically, the efficacy of such methods, and in particular their accuracy, stability, and generalization properties, strongly depends on the choice of hyperparameters. While bilevel optimization offers a principled framework for hyperparameter tuning, its nested optimization structure can be computationally demanding, especially in PDE-constrained contexts. In this paper, we propose an efficient strategy for hyperparameter optimization within the bilevel framework by employing a Gauss-Newton linearization of the inner optimization step. Our approach provides closed-form updates, eliminating the need for repeated costly PDE solves. As a result, each iteration of the outer loop reduces to a single linearized PDE solve, followed by explicit gradient-based hyperparameter updates. We demonstrate the effectiveness of the proposed method through Gaussian process models applied to nonlinear PDEs and to PDE inverse problems. Extensive numerical experiments highlight substantial improvements in accuracy and robustness compared to conventional random hyperparameter initialization. In particular, experiments with additive kernels and neural network-parameterized deep kernels demonstrate the method's scalability and effectiveness for high-dimensional hyperparameter optimization.
Authors: Hossein Taheri, Avishek Ghosh, Arya Mazumdar
Abstract: Continual learning, the ability of a model to adapt to an ongoing sequence of tasks without forgetting the earlier ones, is a central goal of artificial intelligence. To shed light on its underlying mechanisms, we analyze the limitations of continual learning in a tractable yet representative setting. In particular, we study one-hidden-layer quadratic neural networks trained by gradient descent on an XOR cluster dataset with Gaussian noise, where different tasks correspond to different clusters with orthogonal means. Our results obtain bounds on the rate of forgetting during train and test-time in terms of the number of iterations, the sample size, the number of tasks, and the hidden-layer size. Our results reveal interesting phenomena on the role of different problem parameters in the rate of forgetting. Numerical experiments across diverse setups confirm our results, demonstrating their validity beyond the analyzed settings.
Authors: Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Zou, Pan Lu
Abstract: Outcome-driven reinforcement learning has advanced reasoning in large language models (LLMs), but prevailing tool-augmented approaches train a single, monolithic policy that interleaves thoughts and tool calls under full context; this scales poorly with long horizons and diverse tools and generalizes weakly to new scenarios. Agentic systems offer a promising alternative by decomposing work across specialized modules, yet most remain training-free or rely on offline training decoupled from the live dynamics of multi-turn interaction. We introduce AgentFlow, a trainable, in-the-flow agentic framework that coordinates four modules (planner, executor, verifier, generator) through an evolving memory and directly optimizes its planner inside the multi-turn loop. To train on-policy in live environments, we propose Flow-based Group Refined Policy Optimization (Flow-GRPO), which tackles long-horizon, sparse-reward credit assignment by converting multi-turn optimization into a sequence of tractable single-turn policy updates. It broadcasts a single, verifiable trajectory-level outcome to every turn to align local planner decisions with global success and stabilizes learning with group-normalized advantages. Across ten benchmarks, AgentFlow with a 7B-scale backbone outperforms top-performing baselines with average accuracy gains of 14.9% on search, 14.0% on agentic, 14.5% on mathematical, and 4.1% on scientific tasks, even surpassing larger proprietary models like GPT-4o. Further analyses confirm the benefits of in-the-flow optimization, showing improved planning, enhanced tool-calling reliability, and positive scaling with model size and reasoning turns.
Authors: Ziqiao Meng, Qichao Wang, Zhiyang Dou, Zixing Song, Zhipeng Zhou, Irwin King, Peilin Zhao
Abstract: Autoregressive point cloud generation has long lagged behind diffusion-based approaches in quality. The performance gap stems from the fact that autoregressive models impose an artificial ordering on inherently unordered point sets, forcing shape generation to proceed as a sequence of local predictions. This sequential bias emphasizes short-range continuity but undermines the model's capacity to capture long-range dependencies, hindering its ability to enforce global structural properties such as symmetry, consistent topology, and large-scale geometric regularities. Inspired by the level-of-detail (LOD) principle in shape modeling, we propose PointNSP, a coarse-to-fine generative framework that preserves global shape structure at low resolutions and progressively refines fine-grained geometry at higher scales through a next-scale prediction paradigm. This multi-scale factorization aligns the autoregressive objective with the permutation-invariant nature of point sets, enabling rich intra-scale interactions while avoiding brittle fixed orderings. Experiments on ShapeNet show that PointNSP establishes state-of-the-art (SOTA) generation quality for the first time within the autoregressive paradigm. In addition, it surpasses strong diffusion-based baselines in parameter, training, and inference efficiency. Finally, in dense generation with 8,192 points, PointNSP's advantages become even more pronounced, underscoring its scalability potential.
Authors: Ibrahim Salihu Yusuf, Iffanice Houndayi, Rym Oualha, Mohamed Aziz Cherif, Kobby Panford-Quainoo, Arnu Pretorius
Abstract: Open-access multispectral imagery from missions like Landsat 8-9 and Sentinel-2 has fueled the development of geospatial foundation models (GFMs) for humanitarian and environmental applications. Yet, their deployment remains limited by (i) the absence of automated geospatial data pipelines and (ii) the large size of fine-tuned models. Existing GFMs lack workflows for processing raw satellite imagery, and downstream adaptations often retain the full complexity of the original encoder. We present InstaGeo, an open-source, end-to-end framework that addresses these challenges by integrating: (1) automated data curation to transform raw imagery into model-ready datasets; (2) task-specific model distillation to derive compact, compute-efficient models; and (3) seamless deployment as interactive web-map applications. Using InstaGeo, we reproduced datasets from three published studies and trained models with marginal mIoU differences of -0.73 pp for flood mapping, -0.20 pp for crop segmentation, and +1.79 pp for desert locust prediction. The distilled models are up to 8x smaller than standard fine-tuned counterparts, reducing FLOPs and CO2 emissions with minimal accuracy loss. Leveraging InstaGeo's streamlined data pipeline, we also curated a larger crop segmentation dataset, achieving a state-of-the-art mIoU of 60.65%, a 12 pp improvement over prior baselines. Moreover, InstaGeo enables users to progress from raw data to model deployment within a single working day. By unifying data preparation, model compression, and deployment, InstaGeo transforms research-grade GFMs into practical, low-carbon tools for real-time, large-scale Earth observation. This approach shifts geospatial AI toward data quality and application-driven innovation. Source code, datasets, and model checkpoints are available at: https://github.com/instadeepai/InstaGeo-E2E-Geospatial-ML.git
URLs: https://github.com/instadeepai/InstaGeo-E2E-Geospatial-ML.git
Authors: Tianhao Zhu, Dahu Feng, Erhu Feng, Yubin Xia
Abstract: With the widespread adoption of Large Language Models (LLMs), the demand for high-performance LLM inference services continues to grow. To meet this demand, a growing number of AI accelerators have been proposed, such as Google TPU, Huawei NPU, Graphcore IPU, and Cerebras WSE, etc. Most of these accelerators adopt multi-core architectures to achieve enhanced scalability, but lack the flexibility of SIMT architectures. Therefore, without careful configuration of the hardware architecture, as well as deliberate design of tensor parallelism and core placement strategies, computational resources may be underutilized, resulting in suboptimal inference performance. To address these challenges, we first present a multi-level simulation framework with both transaction-level and performance-model-based simulation for multi-core NPUs. Using this simulator, we conduct a systematic analysis and further propose the optimal solutions for tensor parallelism strategies, core placement policies, memory management methods, as well as the selection between PD-disaggregation and PD-fusion on multi-core NPUs. We conduct comprehensive experiments on representative LLMs and various NPU configurations. The evaluation results demonstrate that, our solution can achieve 1.32x-6.03x speedup compared to SOTA designs for multi-core NPUs across different hardware configurations. As for LLM serving, our work offers guidance on designing optimal hardware architectures and serving strategies for multi-core NPUs across various LLM workloads.
Authors: Suhyeok Jang, Dongyoung Kim, Changyeon Kim, Youngsuk Kim, Jinwoo Shin
Abstract: Vision-Language-Action models (VLAs) have demonstrated remarkable performance in robot control. However, they remain fundamentally limited in tasks that require high precision due to their single-inference paradigm. While test-time scaling approaches using external verifiers have shown promise, they require additional training and fail to generalize to unseen conditions. We propose Masking Distribution Guided Selection (MG-Select), a novel test-time scaling framework for VLAs that leverages the model's internal properties without requiring additional training or external modules. Our approach utilizes KL divergence from a reference action token distribution as a confidence metric for selecting the optimal action from multiple candidates. We introduce a reference distribution generated by the same VLA but with randomly masked states and language conditions as inputs, ensuring maximum uncertainty while remaining aligned with the target task distribution. Additionally, we propose a joint training strategy that enables the model to learn both conditional and unconditional distributions by applying dropout to state and language conditions, thereby further improving the quality of the reference distribution. Our experiments demonstrate that MG-Select achieves significant performance improvements, including a 28%/35% improvement in real-world in-distribution/out-of-distribution tasks, along with a 168% relative gain on RoboCasa pick-and-place tasks trained with 30 demonstrations.
Authors: Yuhang Zhang, Jiaping Xiao, Chao Yan, Mir Feroskhan
Abstract: A prevailing approach for learning visuomotor policies is to employ reinforcement learning to map high-dimensional visual observations directly to action commands. However, the combination of high-dimensional visual inputs and agile maneuver outputs leads to long-standing challenges, including low sample efficiency and significant sim-to-real gaps. To address these issues, we propose Oracle-Guided Masked Contrastive Reinforcement Learning (OMC-RL), a novel framework designed to improve the sample efficiency and asymptotic performance of visuomotor policy learning. OMC-RL explicitly decouples the learning process into two stages: an upstream representation learning stage and a downstream policy learning stage. In the upstream stage, a masked Transformer module is trained with temporal modeling and contrastive learning to extract temporally-aware and task-relevant representations from sequential visual inputs. After training, the learned encoder is frozen and used to extract visual representations from consecutive frames, while the Transformer module is discarded. In the downstream stage, an oracle teacher policy with privileged access to global state information supervises the agent during early training to provide informative guidance and accelerate early policy learning. This guidance is gradually reduced to allow independent exploration as training progresses. Extensive experiments in simulated and real-world environments demonstrate that OMC-RL achieves superior sample efficiency and asymptotic policy performance, while also improving generalization across diverse and perceptually complex scenarios.
Authors: Antoine Teissier, Marie Tahon, Nicolas Dugu\'e, Aghilas Sini
Abstract: Due to the rapid progress of speech synthesis, deepfake detection has become a major concern in the speech processing community. Because it is a critical task, systems must not only be efficient and robust, but also provide interpretable explanations. Among the different approaches for explainability, we focus on the interpretation of latent representations. In such paper, we focus on the last layer of embeddings of AASIST, a deepfake detection architecture. We use a TopK activation inspired by SAEs on this layer to obtain sparse representations which are used in the decision process. We demonstrate that sparse deepfake detection can improve detection performance, with an EER of 23.36% on ASVSpoof5 test set, with 95% of sparsity. We then show that these representations provide better disentanglement, using completeness and modularity metrics based on mutual information. Notably, some attacks are directly encoded in the latent space.
Authors: David Boetius, Abdelrahman Abdelnaby, Ashok Kumar, Stefan Leue, Abdalla Swikir, Fares J. Abu-Dakka
Abstract: Learning stable dynamical systems from data is crucial for safe and reliable robot motion planning and control. However, extending stability guarantees to trajectories defined on Riemannian manifolds poses significant challenges due to the manifold's geometric constraints. To address this, we propose a general framework for learning stable dynamical systems on Riemannian manifolds using neural ordinary differential equations. Our method guarantees stability by projecting the neural vector field evolving on the manifold so that it strictly satisfies the Lyapunov stability criterion, ensuring stability at every system state. By leveraging a flexible neural parameterisation for both the base vector field and the Lyapunov function, our framework can accurately represent complex trajectories while respecting manifold constraints by evolving solutions directly on the manifold. We provide an efficient training strategy for applying our framework and demonstrate its utility by solving Riemannian LASA datasets on the unit quaternion (S^3) and symmetric positive-definite matrix manifolds, as well as robotic motions evolving on \mathbb{R}^3 \times S^3. We demonstrate the performance, scalability, and practical applicability of our approach through extensive simulations and by learning robot motions in a real-world experiment.
Authors: Bohan Yao, Shiva Krishna Reddy Malay, Vikas Yadav
Abstract: Large Language Model (LLM)-powered Multi-agent systems (MAS) have achieved state-of-the-art results on various complex reasoning tasks. Recent works have proposed techniques to automate the design of MASes, eliminating the need for manual engineering. However, these techniques perform poorly, often achieving similar or inferior performance to simple baselines. Furthermore, they require computationally expensive re-discovery of architectures for each new task domain and expensive data annotation on domains without existing labeled validation sets. A critical insight is that simple Chain of Thought (CoT) reasoning often performs competitively with these complex systems, suggesting that the fundamental reasoning unit of MASes, CoT, warrants further investigation. To this end, we present a new paradigm for automatic MAS design that pivots the focus to optimizing CoT reasoning. We introduce the Agentic Reasoning Module (ARM), an agentic generalization of CoT where each granular reasoning step is executed by a specialized reasoning module. This module is discovered through a tree search over the code space, starting from a simple CoT module and evolved using mutations informed by reflection on execution traces. The resulting ARM acts as a versatile reasoning building block which can be utilized as a direct recursive loop or as a subroutine in a learned meta-orchestrator. Our approach significantly outperforms both manually designed MASes and state-of-the-art automatic MAS design methods. Crucially, MASes built with ARM exhibit superb generalization, maintaining high performance across different foundation models and task domains without further optimization.
Authors: Jeffrey N. Clark, Elena Fillola, Nawid Keshtmand, Raul Santos-Rodriguez, Matthew Rigby
Abstract: Monitoring greenhouse gas emissions and evaluating national inventories require efficient, scalable, and reliable inference methods. Top-down approaches, combined with recent advances in satellite observations, provide new opportunities to evaluate emissions at continental and global scales. However, transport models used in these methods remain a key source of uncertainty: they are computationally expensive to run at scale, and their uncertainty is difficult to characterise. Artificial intelligence offers a dual opportunity to accelerate transport simulations and to quantify their associated uncertainty. We present an ensemble-based pipeline for estimating atmospheric transport "footprints", greenhouse gas mole fraction measurements, and their uncertainties using a graph neural network emulator of a Lagrangian Particle Dispersion Model (LPDM). The approach is demonstrated with GOSAT (Greenhouse Gases Observing Satellite) observations for Brazil in 2016. The emulator achieved a ~1000x speed-up over the NAME LPDM, while reproducing large-scale footprint structures. Ensembles were calculated to quantify absolute and relative uncertainty, revealing spatial correlations with prediction error. The results show that ensemble spread highlights low-confidence spatial and temporal predictions for both atmospheric transport footprints and methane mole fractions. While demonstrated here for an LPDM emulator, the approach could be applied more generally to atmospheric transport models, supporting uncertainty-aware greenhouse gas inversion systems and improving the robustness of satellite-based emissions monitoring. With further development, ensemble-based emulators could also help explore systematic LPDM errors, offering a computationally efficient pathway towards a more comprehensive uncertainty budget in greenhouse gas flux estimates.
Authors: Aleksandr Lukoianov, Anssi Klapuri
Abstract: Whereas chord transcription has received considerable attention during the past couple of decades, far less work has been devoted to transcribing and encoding the rhythmic patterns that occur in a song. The topic is especially relevant for instruments such as the rhythm guitar, which is typically played by strumming rhythmic patterns that repeat and vary over time. However, in many cases one cannot objectively define a single "right" rhythmic pattern for a given song section. To create a dataset with well-defined ground-truth labels, we asked expert musicians to transcribe the rhythmic patterns in 410 popular songs and record cover versions where the guitar tracks followed those transcriptions. To transcribe the strums and their corresponding rhythmic patterns, we propose a three-step framework. Firstly, we perform approximate stem separation to extract the guitar part from the polyphonic mixture. Secondly, we detect individual strums within the separated guitar audio, using a pre-trained foundation model (MERT) as a backbone. Finally, we carry out a pattern-decoding process in which the transcribed sequence of guitar strums is represented by patterns drawn from an expert-curated vocabulary. We show that it is possible to transcribe the rhythmic patterns of the guitar track in polyphonic music with quite high accuracy, producing a representation that is human-readable and includes automatically detected bar lines and time signature markers. We perform ablation studies and error analysis and propose a set of evaluation metrics to assess the accuracy and readability of the predicted rhythmic pattern sequence.
Authors: Patrick Forr\'e, Abel Jansma
Abstract: We generalize the concept of M\"obius inversion and Shapley values to directed acyclic multigraphs and weighted versions thereof. We further allow value functions (games) and thus their M\"obius transforms (synergy function) and Shapley values to have values in any abelian group that is a module over a ring that contains the graph weights, e.g. vector-valued functions. To achieve this and overcome the obstruction that the classical axioms (linearity, efficiency, null player, symmetry) are not strong enough to uniquely determine Shapley values in this more general setting, we analyze Shapley values from two novel points of view: 1) We introduce projection operators that allow us to interpret Shapley values as the recursive projection and re-attribution of higher-order synergies to lower-order ones; 2) we propose a strengthening of the null player axiom and a localized symmetry axiom, namely the weak elements and flat hierarchy axioms. The former allows us to remove coalitions with vanishing synergy while preserving the rest of the hierarchical structure. The latter treats player-coalition bonds uniformly in the corner case of hierarchically flat graphs. Together with linearity these axioms already imply a unique explicit formula for the Shapley values, as well as classical properties like efficiency, null player, symmetry, and novel ones like the projection property. This whole framework then specializes to finite inclusion algebras, lattices, partial orders and mereologies, and also recovers certain previously known cases as corner cases, and presents others from a new perspective. The admission of general weighted directed acyclic multigraph structured hierarchies and vector-valued functions and Shapley values opens up the possibility for new analytic tools and application areas, like machine learning, language processing, explainable artificial intelligence, and many more.
Authors: Nikita Pavlichenko, Iurii Nazarov, Ivan Dolgov, Ekaterina Garanina, Dmitry Ustalov, Ivan Bondyrev, Kseniia Lysaniuk, Evgeniia Vu, Kirill Chekmenev, Joseph Shtok, Yaroslav Golubev, Anton Semenkin, Uladzislau Sazanovich
Abstract: We present the Mellum models family, open-weight code completion models designed for interactive use in JetBrains IDEs. Mellums have 4B parameters, adopt a Llama-style architecture, and are pre-trained on ~4T tokens of permissively licensed, multi-language code. Our studies show that (i) careful data curation and staged training significantly improve the model's quality, (ii) editor-critical capabilities such as context packing are necessary for high-quality suggestions, and (iii) a compact, task-focused model can meet the cost and latency constraints of interactive completion. In the paper, we describe an end-to-end industrial pipeline for producing contextualized in-editor completion: disciplined data governance, multi-stage training that includes fill-in-the-middle and project context via supervised fine-tuning, and alignment via direct preference optimization using feedback from real-world scenarios. Our quality evaluations include both large-scale offline benchmarks and online telemetry from production deployments in JetBrains IDEs. Mellums are released under the Apache-2.0 license on HuggingFace, with a public model card providing a reproducible reference for practitioners. Our experience offers a pragmatic blueprint for taking a focused, open model from a research prototype to at scale production for hundreds of thousands of users.
Authors: Christian Marinoni, Riccardo Fosco Gramaccioni, Kazuki Shimada, Takashi Shibuya, Yuki Mitsufuji, Danilo Comminiello
Abstract: Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.
Authors: Riccardo Fosco Gramaccioni, Christian Marinoni, Eleonora Grassucci, Giordano Cicchetti, Aurelio Uncini, Danilo Comminiello
Abstract: In this work, we present FoleyGRAM, a novel approach to video-to-audio generation that emphasizes semantic conditioning through the use of aligned multimodal encoders. Building on prior advancements in video-to-audio generation, FoleyGRAM leverages the Gramian Representation Alignment Measure (GRAM) to align embeddings across video, text, and audio modalities, enabling precise semantic control over the audio generation process. The core of FoleyGRAM is a diffusion-based audio synthesis model conditioned on GRAM-aligned embeddings and waveform envelopes, ensuring both semantic richness and temporal alignment with the corresponding input video. We evaluate FoleyGRAM on the Greatest Hits dataset, a standard benchmark for video-to-audio models. Our experiments demonstrate that aligning multimodal encoders using GRAM enhances the system's ability to semantically align generated audio with video content, advancing the state of the art in video-to-audio synthesis.
Authors: Xue-Yong Fu, Elena Khasanova, Md Tahmid Rahman Laskar, Harsh Saini, Shashi Bhushan TN
Abstract: Large language models (LLMs) have achieved impressive performance in text summarization, yet their performance often falls short when applied to specialized domains %or conversational data that differ from their original pre-training distribution. While fine-tuning can improve summarization quality, it typically relies on costly and scarce high-quality labeled data. In this work, we explore continual pre-training as a scalable, self-supervised approach to adapt LLMs for downstream summarization tasks, particularly in the context of noisy real-world conversation transcripts. We conduct extensive experiments using large-scale, unlabeled business conversation data to investigate whether continual pre-training enhances model capabilities in conversational summarization. Our results demonstrate that continual pre-training yields substantial gains in both in-domain and out-of-domain summarization benchmarks, while maintaining strong generalization and robustness. We also analyze the effects of data selection strategies, providing practical guidelines for applying continual pre-training in summarization-focused industrial applications.
Authors: Josefa Lia Stoisser, Lawrence Phillips, Aditya Misra, Tom A. Lamb, Philip Torr, Marc Boubnovski Martell, Julien Fauqueur, Kaspar M\"artens
Abstract: Synthetic chain-of-thought (CoT) traces are widely used to train large reasoning models (LRMs), improving generalization by providing step-level supervision. Yet most approaches require ground-truth labels to seed or filter these traces - an expensive bottleneck in domains like biology where wet-lab data are scarce. We propose a label-free alternative: uncertainty-based filtering, which uses a model's own confidence - quantified through established uncertainty metrics like self-consistency and predictive perplexity - as a substitute for external labels. We sample multiple reasoning traces and retain only low-uncertainty subsets. Applied to biological perturbation prediction, a domain where wet-lab labels are especially costly, we show that the filtered subset has higher accuracy, and that supervised fine-tuning (SFT) on uncertainty-filtered data outperforms unfiltered synthetic data, narrows the gap to ground-truth training, and surpasses strong LRM baselines. Ablations show that per-class filtering corrects for class-specific uncertainty scales and that hybrid uncertainty metrics yield higher-quality datasets. Our results suggest that model-internal confidence is a powerful signal for efficient reasoning dataset creation, enabling LRMs in domains where supervision is expensive.
Authors: Ping-Yi Chen, Chih-Pin Tan, Yi-Hsuan Yang
Abstract: We propose the Segmented Full-Song Model (SFS) for symbolic full-song generation. The model accepts a user-provided song structure and an optional short seed segment that anchors the main idea around which the song is developed. By factorizing a song into segments and generating each one through selective attention to related segments, the model achieves higher quality and efficiency compared to prior work. To demonstrate its suitability for human-AI interaction, we further wrap SFS into a web application that enables users to iteratively co-create music on a piano roll with customizable structures and flexible ordering.
Authors: Sebastian H\"ofer, Dorian Henning, Artemij Amiranashvili, Douglas Morrison, Mariliza Tzes, Ingmar Posner, Marc Matvienko, Alessandro Rennola, Anton Milan
Abstract: We present a novel large-scale dataset for defect detection in a logistics setting. Recent work on industrial anomaly detection has primarily focused on manufacturing scenarios with highly controlled poses and a limited number of object categories. Existing benchmarks like MVTec-AD [6] and VisA [33] have reached saturation, with state-of-the-art methods achieving up to 99.9% AUROC scores. In contrast to manufacturing, anomaly detection in retail logistics faces new challenges, particularly in the diversity and variability of object pose and appearance. Leading anomaly detection methods fall short when applied to this new setting. To bridge this gap, we introduce a new benchmark that overcomes the current limitations of existing datasets. With over 230,000 images (and more than 29,000 defective instances), it is 40 times larger than MVTec-AD and contains more than 48,000 distinct objects. To validate the difficulty of the problem, we conduct an extensive evaluation of multiple state-of-the-art anomaly detection methods, demonstrating that they do not surpass 56.96% AUROC on our dataset. Further qualitative analysis confirms that existing methods struggle to leverage normal samples under heavy pose and appearance variation. With our large-scale dataset, we set a new benchmark and encourage future research towards solving this challenging problem in retail logistics anomaly detection. The dataset is available for download under https://www.kaputt-dataset.com.
Authors: Hsien-Chin Lin, Benjamin Matthias Ruppik, Carel van Niekerk, Chia-Hao Shen, Michael Heck, Nurul Lubis, Renato Vukovic, Shutong Feng, Milica Ga\v{s}i\'c
Abstract: Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.
Authors: Zheyue Tan, Mustapha Abdullahi, Tuo Shi, Huining Yuan, Zelai Xu, Chao Yu, Boxun Li, Bo Zhao
Abstract: Reinforcement learning (RL) has become a pivotal component of large language model (LLM) post-training, and agentic RL extends this paradigm to operate as agents through multi-turn interaction and tool use. Scaling such systems exposes two practical bottlenecks: (1) context length grows rapidly during training, inflating memory usage and latency, and triggering out-of-memory (OOM) failures; and (2) intermediate tensors accumulate with context length, making cross-device data movement a major system bottleneck. We present EARL, a scalable system for efficient agentic RL. EARL designs a parallelism selector that dynamically adapts model and training parallelism across RL stages based on sequence length and system load, and a data dispatcher that performs layout-aware, decentralized exchange of intermediate data batches. Together, these components increase throughput, reduce long-context failures, and enable stable large-scale training of agentic LLMs without relying on hard limits or penalties of context length.
Authors: Xenia Heilmann, Ernst Althaus, Mattia Cerrato, Nick Johannes Peter Rassau, Mohammad Sadeq Dousti, Stefan Kramer
Abstract: A sum-product network (SPN) is a graphical model that allows several types of probabilistic inference to be performed efficiently. In this paper, we propose a privacy-preserving protocol which tackles structure generation and parameter learning of SPNs. Additionally, we provide a protocol for private inference on SPNs, subsequent to training. To preserve the privacy of the participants, we derive our protocol based on secret sharing, which guarantees privacy in the honest-but-curious setting even when at most half of the parties cooperate to disclose the data. The protocol makes use of a forest of randomly generated SPNs, which is trained and weighted privately and can then be used for private inference on data points. Our experiments indicate that preserving the privacy of all participants does not decrease log-likelihood performance on both homogeneously and heterogeneously partitioned data. We furthermore show that our protocol's performance is comparable to current state-of-the-art SPN learners in homogeneously partitioned data settings. In terms of runtime and memory usage, we demonstrate that our implementation scales well when increasing the number of parties, comparing favorably to protocols for neural networks, when they are trained to reproduce the input-output behavior of SPNs.
Authors: Eashan Adhikarla, Yixin Liu, Brian D. Davison
Abstract: Low-light image enhancement (LLIE) is vital for safety-critical applications such as surveillance, autonomous navigation, and medical imaging, where visibility degradation can impair downstream task performance. Recently, diffusion models have emerged as a promising generative paradigm for LLIE due to their capacity to model complex image distributions via iterative denoising. This survey provides an up-to-date critical analysis of diffusion models for LLIE, distinctively featuring an in-depth comparative performance evaluation against Generative Adversarial Network and Transformer-based state-of-the-art methods, a thorough examination of practical deployment challenges, and a forward-looking perspective on the role of emerging paradigms like foundation models. We propose a multi-perspective taxonomy encompassing six categories: Intrinsic Decomposition, Spectral & Latent, Accelerated, Guided, Multimodal, and Autonomous; that map enhancement methods across physical priors, conditioning schemes, and computational efficiency. Our taxonomy is grounded in a hybrid view of both the model mechanism and the conditioning signals. We evaluate qualitative failure modes, benchmark inconsistencies, and trade-offs between interpretability, generalization, and inference efficiency. We also discuss real-world deployment constraints (e.g., memory, energy use) and ethical considerations. This survey aims to guide the next generation of diffusion-based LLIE research by highlighting trends and surfacing open research questions, including novel conditioning, real-time adaptation, and the potential of foundation models.
Authors: Moritz Schneider, Robert Krug, Narunas Vaskevicius, Luigi Palmieri, Michael Volpp, Joschka Boedecker
Abstract: Empowerment, an information-theoretic measure of an agent's potential influence on its environment, has emerged as a powerful intrinsic motivation and exploration framework for reinforcement learning (RL). Besides for unsupervised RL and skill learning algorithms, the specific use of empowerment as a pre-training signal has received limited attention in the literature. We show that empowerment can be used as a pre-training signal for data-efficient downstream task adaptation. For this we extend the traditional notion of empowerment by introducing discounted empowerment, which balances the agent's control over the environment across short- and long-term horizons. Leveraging this formulation, we propose a novel pre-training paradigm that initializes policies to maximize discounted empowerment, enabling agents to acquire a robust understanding of environmental dynamics. We analyze empowerment-based pre-training for various existing RL algorithms and empirically demonstrate its potential as a general-purpose initialization strategy: empowerment-maximizing policies with long horizons are data-efficient and effective, leading to improved adaptability in downstream tasks. Our findings pave the way for future research to scale this framework to high-dimensional and complex tasks, further advancing the field of RL.
Authors: Aueaphum Aueawatthanaphisut, Nyi Wunna Tun
Abstract: The comparative evaluation between classical and quantum reinforcement learning (QRL) paradigms was conducted to investigate their convergence behavior, robustness under observational noise, and computational efficiency in a benchmark control environment. The study employed a multilayer perceptron (MLP) agent as a classical baseline and a parameterized variational quantum circuit (VQC) as a quantum counterpart, both trained on the CartPole-v1 environment over 500 episodes. Empirical results demonstrated that the classical MLP achieved near-optimal policy convergence with a mean return of 498.7 +/- 3.2, maintaining stable equilibrium throughout training. In contrast, the VQC exhibited limited learning capability, with an average return of 14.6 +/- 4.8, primarily constrained by circuit depth and qubit connectivity. Noise robustness analysis further revealed that the MLP policy deteriorated gracefully under Gaussian perturbations, while the VQC displayed higher sensitivity at equivalent noise levels. Despite the lower asymptotic performance, the VQC exhibited significantly lower parameter count and marginally increased training time, highlighting its potential scalability for low-resource quantum processors. The results suggest that while classical neural policies remain dominant in current control benchmarks, quantum-enhanced architectures could offer promising efficiency advantages once hardware noise and expressivity limitations are mitigated.
Authors: An Thi Nguyen, Radina Stoykova, Eric Arazo
Abstract: Generic instance search models can dramatically reduce the manual effort required to analyze vast surveillance footage during criminal investigations by retrieving specific objects of interest to law enforcement. However, our research reveals an unintended emergent capability: through overlearning, these models can single out specific individuals even when trained on datasets without human subjects. This capability raises concerns regarding identification and profiling of individuals based on their personal data, while there is currently no clear standard on how de-identification can be achieved. We evaluate two technical safeguards to curtail a model's person re-identification capacity: index exclusion and confusion loss. Our experiments demonstrate that combining these approaches can reduce person re-identification accuracy to below 2% while maintaining 82% of retrieval performance for non-person objects. However, we identify critical vulnerabilities in these mitigations, including potential circumvention using partial person images. These findings highlight urgent regulatory questions at the intersection of AI governance and data protection: How should we classify and regulate systems with emergent identification capabilities? And what technical standards should be required to prevent identification capabilities from developing in seemingly benign applications?
Authors: Rohit Goswami (Science Institute and Faculty of Physical Sciences, University of Iceland, Reykjav\'ik, Iceland), Hannes J\'onsson (Science Institute and Faculty of Physical Sciences, University of Iceland, Reykjav\'ik, Iceland)
Abstract: Gaussian process (GP) regression provides a strategy for accelerating saddle point searches on high-dimensional energy surfaces by reducing the number of times the energy and its derivatives with respect to atomic coordinates need to be evaluated. The computational overhead in the hyperparameter optimization can, however, be large and make the approach inefficient. Failures can also occur if the search ventures too far into regions that are not represented well enough by the GP model. Here, these challenges are resolved by using geometry-aware optimal transport measures and an active pruning strategy using a summation over Wasserstein-1 distances for each atom-type in farthest-point sampling, selecting a fixed-size subset of geometrically diverse configurations to avoid rapidly increasing cost of GP updates as more observations are made. Stability is enhanced by permutation-invariant metric that provides a reliable trust radius for early-stopping and a logarithmic barrier penalty for the growth of the signal variance. These physically motivated algorithmic changes prove their efficacy by reducing to less than a half the mean computational time on a set of 238 challenging configurations from a previously published data set of chemical reactions. With these improvements, the GP approach is established as, a robust and scalable algorithm for accelerating saddle point searches when the evaluation of the energy and atomic forces requires significant computational effort.
Authors: Austin Feng, Andreas Varvarigos, Ioannis Panitsas, Daniela Fernandez, Jinbiao Wei, Yuwei Guo, Jialin Chen, Ali Maatouk, Leandros Tassiulas, Rex Ying
Abstract: Modern enterprises generate vast streams of time series metrics when monitoring complex systems, known as observability data. Unlike conventional time series from domains such as weather, observability data are zero-inflated, highly stochastic, and exhibit minimal temporal structure. Despite their importance, observability datasets are underrepresented in public benchmarks due to proprietary restrictions. Existing datasets are often anonymized and normalized, removing scale information and limiting their use for tasks beyond forecasting, such as anomaly detection, root-cause analysis, and multi-modal reasoning. To address this gap, we introduce TelecomTS, a large-scale observability dataset derived from a 5G telecommunications network. TelecomTS features heterogeneous, de-anonymized covariates with explicit scale information and supports a suite of downstream tasks, including anomaly detection, root-cause analysis, and a question-answering benchmark requiring multi-modal reasoning. Benchmarking state-of-the-art time series, language, and reasoning models reveals that existing approaches struggle with the abrupt, noisy, and high-variance dynamics of observability data. Our experiments also underscore the importance of preserving covariates' absolute scale, emphasizing the need for foundation time series models that natively leverage scale information for practical observability applications.
Authors: Akshay Muppidi, Martin Radfar
Abstract: Vision-based Proximal Policy Optimization (PPO) struggles with visual observation-based robotic laparoscopic surgical tasks due to the high-dimensional nature of visual input, the sparsity of rewards in surgical environments, and the difficulty of extracting task-relevant features from raw visual data. We introduce a simple approach integrating MedFlamingo, a medical domain-specific Vision-Language Model, with PPO. Our method is evaluated on five diverse laparoscopic surgery task environments in LapGym, using only endoscopic visual observations. MedFlamingo PPO outperforms and converges faster compared to both standard vision-based PPO and OpenFlamingo PPO baselines, achieving task success rates exceeding 70% across all environments, with improvements ranging from 66.67% to 1114.29% compared to baseline. By processing task observations and instructions once per episode to generate high-level planning tokens, our method efficiently combines medical expertise with real-time visual feedback. Our results highlight the value of specialized medical knowledge in robotic surgical planning and decision-making.
Authors: Akshay Muppidi, Martin Radfar
Abstract: Speech emotion recognition (SER) is pivotal for enhancing human-machine interactions. This paper introduces "EmoHRNet", a novel adaptation of High-Resolution Networks (HRNet) tailored for SER. The HRNet structure is designed to maintain high-resolution representations from the initial to the final layers. By transforming audio samples into spectrograms, EmoHRNet leverages the HRNet architecture to extract high-level features. EmoHRNet's unique architecture maintains high-resolution representations throughout, capturing both granular and overarching emotional cues from speech signals. The model outperforms leading models, achieving accuracies of 92.45% on RAVDESS, 80.06% on IEMOCAP, and 92.77% on EMOVO. Thus, we show that EmoHRNet sets a new benchmark in the SER domain.
Authors: Batu El, James Zou
Abstract: Large language models (LLMs) are increasingly shaping how information is created and disseminated, from companies using them to craft persuasive advertisements, to election campaigns optimizing messaging to gain votes, to social media influencers boosting engagement. These settings are inherently competitive, with sellers, candidates, and influencers vying for audience approval, yet it remains poorly understood how competitive feedback loops influence LLM behavior. We show that optimizing LLMs for competitive success can inadvertently drive misalignment. Using simulated environments across these scenarios, we find that, 6.3% increase in sales is accompanied by a 14.0% rise in deceptive marketing; in elections, a 4.9% gain in vote share coincides with 22.3% more disinformation and 12.5% more populist rhetoric; and on social media, a 7.5% engagement boost comes with 188.6% more disinformation and a 16.3% increase in promotion of harmful behaviors. We call this phenomenon Moloch's Bargain for AI--competitive success achieved at the cost of alignment. These misaligned behaviors emerge even when models are explicitly instructed to remain truthful and grounded, revealing the fragility of current alignment safeguards. Our findings highlight how market-driven optimization pressures can systematically erode alignment, creating a race to the bottom, and suggest that safe deployment of AI systems will require stronger governance and carefully designed incentives to prevent competitive dynamics from undermining societal trust.
Authors: Aditya Prakash, David Forsyth, Saurabh Gupta
Abstract: We tackle the problem of forecasting bimanual 3D hand motion & articulation from a single image in everyday settings. To address the lack of 3D hand annotations in diverse settings, we design an annotation pipeline consisting of a diffusion model to lift 2D hand keypoint sequences to 4D hand motion. For the forecasting model, we adopt a diffusion loss to account for the multimodality in hand motion distribution. Extensive experiments across 6 datasets show the benefits of training on diverse data with imputed labels (14% improvement) and effectiveness of our lifting (42% better) & forecasting (16.4% gain) models, over the best baselines, especially in zero-shot generalization to everyday images.
Authors: Giacomo De Palma, Marco Fanizza, Connor Mowry, Ryan O'Donnell
Abstract: We study hypothesis testing (aka state certification) in the non-identically distributed setting. A recent work (Garg et al. 2023) considered the classical case, in which one is given (independent) samples from $T$ unknown probability distributions $p_1, \dots, p_T$ on $[d] = \{1, 2, \dots, d\}$, and one wishes to accept/reject the hypothesis that their average $p_{\mathrm{avg}}$ equals a known hypothesis distribution $q$. Garg et al. showed that if one has just $c = 2$ samples from each $p_i$, and provided $T \gg \frac{\sqrt{d}}{\epsilon^2} + \frac{1}{\epsilon^4}$, one can (whp) distinguish $p_{\mathrm{avg}} = q$ from $d_{\mathrm{TV}}(p_{\mathrm{avg}},q) > \epsilon$. This nearly matches the optimal result for the classical iid setting (namely, $T \gg \frac{\sqrt{d}}{\epsilon^2}$). Besides optimally improving this result (and generalizing to tolerant testing with more stringent distance measures), we study the analogous problem of hypothesis testing for non-identical quantum states. Here we uncover an unexpected phenomenon: for any $d$-dimensional hypothesis state $\sigma$, and given just a single copy ($c = 1$) of each state $\rho_1, \dots, \rho_T$, one can distinguish $\rho_{\mathrm{avg}} = \sigma$ from $D_{\mathrm{tr}}(\rho_{\mathrm{avg}},\sigma) > \epsilon$ provided $T \gg d/\epsilon^2$. (Again, we generalize to tolerant testing with more stringent distance measures.) This matches the optimal result for the iid case, which is surprising because doing this with $c = 1$ is provably impossible in the classical case. We also show that the analogous phenomenon happens for the non-iid extension of identity testing between unknown states. A technical tool we introduce may be of independent interest: an Efron-Stein inequality, and more generally an Efron-Stein decomposition, in the quantum setting.
Authors: Hwanwoo Kim, Dongkyu Derek Cho, Eric Laber
Abstract: Temporal difference (TD) learning is a cornerstone of reinforcement learning. In the average-reward setting, standard TD($\lambda$) is highly sensitive to the choice of step-size and thus requires careful tuning to maintain numerical stability. We introduce average-reward implicit TD($\lambda$), which employs an implicit fixed point update to provide data-adaptive stabilization while preserving the per iteration computational complexity of standard average-reward TD($\lambda$). In contrast to prior finite-time analyses of average-reward TD($\lambda$), which impose restrictive step-size conditions, we establish finite-time error bounds for the implicit variant under substantially weaker step-size requirements. Empirically, average-reward implicit TD($\lambda$) operates reliably over a much broader range of step-sizes and exhibits markedly improved numerical stability. This enables more efficient policy evaluation and policy learning, highlighting its effectiveness as a robust alternative to average-reward TD($\lambda$).
Authors: Emre Adabag, Marcus Greiff, John Subosits, Thomas Lew
Abstract: Differentiable model predictive control (MPC) offers a powerful framework for combining learning and control. However, its adoption has been limited by the inherently sequential nature of traditional optimization algorithms, which are challenging to parallelize on modern computing hardware like GPUs. In this work, we tackle this bottleneck by introducing a GPU-accelerated differentiable optimization tool for MPC. This solver leverages sequential quadratic programming and a custom preconditioned conjugate gradient (PCG) routine with tridiagonal preconditioning to exploit the problem's structure and enable efficient parallelization. We demonstrate substantial speedups over CPU- and GPU-based baselines, significantly improving upon state-of-the-art training times on benchmark reinforcement learning and imitation learning tasks. Finally, we showcase the method on the challenging task of reinforcement learning for driving at the limits of handling, where it enables robust drifting of a Toyota Supra through water puddles.
Authors: Jordan Seneca, Suzanne Bintanja, Frank M. Selten
Abstract: In climate science, the tuning of climate models is a computationally intensive problem due to the combination of the high-dimensionality of the system state and long integration times. Here we demonstrate the potential of a parameter estimation algorithm which makes use of synchronization to tune a global atmospheric model at modest computational costs. We first use it to directly optimize internal model parameters. We then apply the algorithm to the weights of each member of a supermodel ensemble to optimize the overall predictions. In both cases, the algorithm is able to find parameters which result in reduced errors in the climatology of the model. Finally, we introduce a novel approach which combines both methods called adaptive supermodeling, where the internal parameters of the members of a supermodel are tuned simultaneously with the model weights such that the supermodel predictions are optimized. For a case designed to challenge the two previous methods, adaptive supermodeling achieves a performance similar to a perfect model.
Authors: Jakir Hasan, Shubhashis Roy Dipta
Abstract: Real-time speech assistants are becoming increasingly popular for ensuring improved accessibility to information. Bengali, being a low-resource language with a high regional dialectal diversity, has seen limited progress in developing such systems. Existing systems are not optimized for real-time use and focus only on standard Bengali. In this work, we present BanglaTalk, the first real-time speech assistance system for Bengali regional dialects. BanglaTalk follows the client-server architecture and uses the Real-time Transport Protocol (RTP) to ensure low-latency communication. To address dialectal variation, we introduce a dialect-aware ASR system, BRDialect, developed by fine-tuning the IndicWav2Vec model in ten Bengali regional dialects. It outperforms the baseline ASR models by 12.41-33.98% on the RegSpeech12 dataset. Furthermore, BanglaTalk can operate at a low bandwidth of 24 kbps while maintaining an average end-to-end delay of 4.9 seconds. Low bandwidth usage and minimal end-to-end delay make the system both cost-effective and interactive for real-time use cases, enabling inclusive and accessible speech technology for the diverse community of Bengali speakers.
Authors: Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srinivasan Iyer, Duc Le
Abstract: Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.
Authors: Christopher Mitcheltree, Hao Hao Tan, Joshua D. Reiss
Abstract: Modulations are a critical part of sound design and music production, enabling the creation of complex and evolving audio. Modern synthesizers provide envelopes, low frequency oscillators (LFOs), and more parameter automation tools that allow users to modulate the output with ease. However, determining the modulation signals used to create a sound is difficult, and existing sound-matching / parameter estimation systems are often uninterpretable black boxes or predict high-dimensional framewise parameter values without considering the shape, structure, and routing of the underlying modulation curves. We propose a neural sound-matching approach that leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover the modulations present in a sound. We demonstrate the effectiveness of our approach on highly modulated synthetic and real audio samples, its applicability to different DDSP synth architectures, and investigate the trade-off it incurs between interpretability and sound-matching accuracy. We make our code and audio samples available and provide the trained DDSP synths in a VST plugin.
Authors: Jiaru Zou, Soumya Roy, Vinay Kumar Verma, Ziyi Wang, David Wipf, Pan Lu, Sumit Negi, James Zou, Jingrui He
Abstract: Process Reward Models (PRMs) have recently emerged as a powerful framework for enhancing the reasoning capabilities of large reasoning models (LRMs), particularly in the context of test-time scaling (TTS). However, their potential for supervising LRMs on tabular reasoning domains remains underexplored. Through detailed empirical analyses, we identify that existing PRMs, though widely adopted for supervising text-only reasoning steps, struggle with table-specific operations such as sub-table retrieval and schema interaction, leading to critical performance bottlenecks. To address this limitation, we propose TaTToo, a novel table-grounded PRM framework that (i) reasons explicitly over tabular reasoning steps and (ii) integrates tool-based verification to provide precise reward supervision. Concretely, we first design a scalable data curation pipeline that constructs over 60k high-quality step-level annotations by integrating table verification rationales with tool-based executions. Building on the collected data, we train TaTToo with a dual-stage paradigm: cold-start supervised fine-tuning to capture tool-use reasoning patterns, followed by reinforcement learning with tool-grounded reward shaping to align our model with table-based verification. We provide a comprehensive evaluation of the policy improvement induced by our newly designed PRM. Across 5 challenging tabular reasoning benchmarks covering numerical reasoning, fact-checking, and data analysis, TaTToo improves downstream policy LRMs by 30.9% at inference, surpasses strong PRM baselines such as Qwen-2.5-Math-PRM-72B with only 8B parameters, and demonstrates strong generalizability across diverse TTS strategies.
Authors: V. Gunasekaran, K. K. Kovi, S. Arja, R. Chimata
Abstract: Renewable energy forecasting is attaining greater importance due to its constant increase in contribution to the electrical power grids. Solar energy is one of the most significant contributors to renewable energy and is dependent on solar irradiation. For the effective management of electrical power grids, forecasting models that predict solar irradiation, with high accuracy, are needed. In the current study, Machine Learning techniques such as Linear Regression, Extreme Gradient Boosting and Genetic Algorithm Optimization are used to forecast solar irradiation. The data used for training and validation is recorded from across three different geographical stations in the United States that are part of the SURFRAD network. A Global Horizontal Index (GHI) is predicted for the models built and compared. Genetic Algorithm Optimization is applied to XGB to further improve the accuracy of solar irradiation prediction.
Authors: Anne-Sofie Maerten, Derya Soydaner
Abstract: This paper delves into the fascinating field of AI-generated art and explores the various deep neural network architectures and models that have been utilized to create it. From the classic convolutional networks to the cutting-edge diffusion models, we examine the key players in the field. We explain the general structures and working principles of these neural networks. Then, we showcase examples of milestones, starting with the dreamy landscapes of DeepDream and moving on to the most recent developments, including Stable Diffusion and DALL-E 3, which produce mesmerizing images. We provide a detailed comparison of these models, highlighting their strengths and limitations, and examining the remarkable progress that deep neural networks have made so far in a short period of time. With a unique blend of technical explanations and insights into the current state of AI-generated art, this paper exemplifies how art and computer science interact.
Authors: Mohammad Al-Jarrah, Niyizhen Jin, Bamdad Hosseini, Amirhossein Taghvaei
Abstract: This paper is concerned with the problem of nonlinear filtering, i.e., computing the conditional distribution of the state of a stochastic dynamical system given a history of noisy partial observations. Conventional sequential importance resampling (SIR) particle filters suffer from fundamental limitations, in scenarios involving degenerate likelihoods or high-dimensional states, due to the weight degeneracy issue. In this paper, we explore an alternative method, which is based on estimating the Brenier optimal transport (OT) map from the current prior distribution of the state to the posterior distribution at the next time step. Unlike SIR particle filters, the OT formulation does not require the analytical form of the likelihood. Moreover, it allows us to harness the approximation power of neural networks to model complex and multi-modal distributions and employ stochastic optimization algorithms to enhance scalability. Extensive numerical experiments are presented that compare the OT method to the SIR particle filter and the ensemble Kalman filter, evaluating the performance in terms of sample efficiency, high-dimensional scalability, and the ability to capture complex and multi-modal distributions.
Authors: Yige Li, Jiabo He, Hanxun Huang, Jun Sun, Xingjun Ma, Yu-Gang Jiang
Abstract: Backdoor attacks have become a significant threat to the pre-training and deployment of deep neural networks (DNNs). Although numerous methods for detecting and mitigating backdoor attacks have been proposed, most rely on identifying and eliminating the ``shortcut" created by the backdoor, which links a specific source class to a target class. However, these approaches can be easily circumvented by designing multiple backdoor triggers that create shortcuts everywhere and therefore nowhere specific. In this study, we explore the concept of Multi-Trigger Backdoor Attacks (MTBAs), where multiple adversaries leverage different types of triggers to poison the same dataset. By proposing and investigating three types of multi-trigger attacks including \textit{parallel}, \textit{sequential}, and \textit{hybrid} attacks, we demonstrate that 1) multiple triggers can coexist, overwrite, or cross-activate one another, and 2) MTBAs easily break the prevalent shortcut assumption underlying most existing backdoor detection/removal methods, rendering them ineffective. Given the security risk posed by MTBAs, we have created a multi-trigger backdoor poisoning dataset to facilitate future research on detecting and mitigating these attacks, and we also discuss potential defense strategies against MTBAs. Our code is available at https://github.com/bboylyg/Multi-Trigger-Backdoor-Attacks.
URLs: https://github.com/bboylyg/Multi-Trigger-Backdoor-Attacks.
Authors: Ahmed Elhussein, Gamze Gursoy
Abstract: Federated Learning is increasingly used in domains such as healthcare to facilitate collaborative model training without data-sharing. However, datasets located in different sites are often non-identically distributed, leading to degradation of model performance in FL. Most existing methods for assessing these distribution shifts are limited by being dataset or task-specific. Moreover, these metrics can only be calculated by exchanging data, a practice restricted in many FL scenarios. To address these challenges, we propose a novel metric for assessing dataset similarity. Our metric exhibits several desirable properties for FL: it is dataset-agnostic, is calculated in a privacy-preserving manner, and is computationally efficient, requiring no model training. In this paper, we first establish a theoretical connection between our metric and training dynamics in FL. Next, we extensively evaluate our metric on a range of datasets including synthetic, benchmark, and medical imaging datasets. We demonstrate that our metric shows a robust and interpretable relationship with model performance and can be calculated in privacy-preserving manner. As the first federated dataset similarity metric, we believe this metric can better facilitate successful collaborations between sites.
Authors: Jose M. Alvarez, Salvatore Ruggieri
Abstract: Testing for individual discrimination involves deriving a profile, the comparator, similar to the one making the discrimination claim, the complainant, based on a protected attribute, such as race or gender, and comparing their decision outcomes. The complainant-comparator pair is central to discrimination testing. Most discrimination testing tools rely on this pair to establish evidence for discrimination. In this work we revisit the role of the comparator in discrimination testing. We first argue for the inherent causal modeling nature of deriving the comparator. We then introduce a two-kinds classification for the comparator: the ceteris paribus, or``with all else equal,'' (CP) comparator and the mutatis mutandis, or ``with the appropriate adjustments being made,'' (MM) comparator. The CP comparator is the standard comparator, representing an idealized comparison for establishing discrimination as it aims for a complainant-comparator pair that only differs on membership to the protected attribute. As an alternative to it, we define the MM comparator, which requires that the comparator represents the``what would have been of'' the complainant without the effects of the protected attribute on the non-protected attributes. Under the MM comparator, the complainant-comparator pair can be dissimilar in terms of the non-protected attributes, departing from an idealized comparison. Notably, the MM comparator is a more complex kind of comparator and its implementation offers an impactful venue for machine learning methods. We illustrate these two comparators and their impact on discrimination testing using a real-world example.
Authors: Wenzhuo Tang, Haitao Mao, Danial Dervovic, Ivan Brugere, Saumitra Mishra, Yuying Xie, Jiliang Tang
Abstract: Models for natural language and images benefit from data scaling behavior: the more data fed into the model, the better they perform. This 'better with more' phenomenon enables the effectiveness of large-scale pre-training on vast amounts of data. However, current graph pre-training methods struggle to scale up data due to heterogeneity across graphs. To achieve effective data scaling, we aim to develop a general model that is able to capture diverse data patterns of graphs and can be utilized to adaptively help the downstream tasks. To this end, we propose UniAug, a universal graph structure augmentor built on a diffusion model. We first pre-train a discrete diffusion model on thousands of graphs across domains to learn the graph structural patterns. In the downstream phase, we provide adaptive enhancement by conducting graph structure augmentation with the help of the pre-trained diffusion model via guided generation. By leveraging the pre-trained diffusion model for structure augmentation, we consistently achieve performance improvements across various downstream tasks in a plug-and-play manner. To the best of our knowledge, this study represents the first demonstration of a data-scaling graph structure augmentor on graphs across domains.
Authors: Yanis Lalou, Th\'eo Gnassounou, Antoine Collas, Antoine de Mathelin, Oleksii Kachaiev, Ambroise Odonnat, Alexandre Gramfort, Thomas Moreau, R\'emi Flamary
Abstract: Unsupervised Domain Adaptation (DA) consists of adapting a model trained on a labeled source domain to perform well on an unlabeled target domain with some data distribution shift. While many methods have been proposed in the literature, fair and realistic evaluation remains an open question, particularly due to methodological difficulties in selecting hyperparameters in the unsupervised setting. With SKADA-bench, we propose a framework to evaluate DA methods on diverse modalities, beyond computer vision task that have been largely explored in the literature. We present a complete and fair evaluation of existing shallow algorithms, including reweighting, mapping, and subspace alignment. Realistic hyperparameter selection is performed with nested cross-validation and various unsupervised model selection scores, on both simulated datasets with controlled shifts and real-world datasets across diverse modalities, such as images, text, biomedical, and tabular data. Our benchmark highlights the importance of realistic validation and provides practical guidance for real-life applications, with key insights into the choice and impact of model selection approaches. SKADA-bench is open-source, reproducible, and can be easily extended with novel DA methods, datasets, and model selection criteria without requiring re-evaluating competitors. SKADA-bench is available on Github at https://github.com/scikit-adaptation/skada-bench.
Authors: Marius Huber, Sara Kalisnik, Patrick Schnider
Abstract: We present AuToMATo, a novel clustering algorithm based on persistent homology. While AuToMATo is not parameter-free per se, we provide default choices for its parameters that make it into an out-of-the-box clustering algorithm that performs well across the board. AuToMATo combines the existing ToMATo clustering algorithm with a bootstrapping procedure in order to separate significant peaks of an estimated density function from non-significant ones. We perform a thorough comparison of AuToMATo (with its parameters fixed to their defaults) against many other state-of-the-art clustering algorithms. We find not only that AuToMATo compares favorably against parameter-free clustering algorithms, but in many instances also significantly outperforms even the best selection of parameters for other algorithms. AuToMATo is motivated by applications in topological data analysis, in particular the Mapper algorithm, where it is desirable to work with a clustering algorithm that does not need tuning of its parameters. Indeed, we provide evidence that AuToMATo performs well when used with Mapper. Finally, we provide an open-source implementation of AuToMATo in Python that is fully compatible with the standard scikit-learn architecture.
Authors: Marc Canby, Adam Davies, Chirag Rastogi, Julia Hockenmaier
Abstract: Causal probing aims to analyze foundation models by examining how intervening on their representation of various latent properties impacts their outputs. Recent works have cast doubt on the theoretical basis of several leading causal probing methods, but it has been unclear how to systematically evaluate the effectiveness of these methods in practice. To address this, we define two key causal probing desiderata: completeness (how thoroughly the representation of the target property has been transformed) and selectivity (how little non-targeted properties have been impacted). We find that there is an inherent tradeoff between the two, which we define as reliability, their harmonic mean. We introduce an empirical analysis framework to measure and evaluate these quantities, allowing us to make the first direct comparisons between different families of leading causal probing methods (e.g., linear vs. nonlinear, or concept removal vs. counterfactual interventions). We find that: (1) all methods show a clear tradeoff between completeness and selectivity; (2) more complete and reliable methods have a greater impact on LLM behavior; and (3) nonlinear interventions are almost always more reliable than linear interventions.
Authors: Lianyu Hu, Mudi Jiang, Junjie Dong, Xinying Liu, Zengyou He
Abstract: In recent years, much of the research on clustering algorithms has primarily focused on enhancing their accuracy and efficiency, frequently at the expense of interpretability. However, as these methods are increasingly being applied in high-stakes domains such as healthcare, finance, and autonomous systems, the need for transparent and interpretable clustering outcomes has become a critical concern. This is not only necessary for gaining user trust but also for satisfying the growing ethical and regulatory demands in these fields. Ensuring that decisions derived from clustering algorithms can be clearly understood and justified is now a fundamental requirement. To address this need, this paper provides a comprehensive and structured review of the current state of explainable clustering algorithms, identifying key criteria to distinguish between various methods. These insights can effectively assist researchers in making informed decisions about the most suitable explainable clustering methods for specific application contexts, while also promoting the development and adoption of clustering algorithms that are both efficient and transparent. For convenient access and reference, an open repository organizes representative and emerging interpretable clustering methods under the taxonomy proposed in this survey, available at https://github.com/hulianyu/Awesome-Interpretable-Clustering
URLs: https://github.com/hulianyu/Awesome-Interpretable-Clustering
Authors: Yuan Mi, Qi Wang, Xueqin Hu, Yike Guo, Ji-Rong Wen, Yang Liu, Hao Sun
Abstract: Data-driven learning of physical systems has kindled significant attention, where many neural models have been developed. In particular, mesh-based graph neural networks (GNNs) have demonstrated significant potential in modeling spatiotemporal dynamics across arbitrary geometric domains. However, the existing node-edge message-passing and aggregation mechanism in GNNs limits the representation learning ability. In this paper, we proposed a dual-module framework, Cell-embedded and Feature-enhanced Graph Neural Network (aka, CeFeGNN), for learning spatiotemporal dynamics. Specifically, we embed learnable cell attributions to the common node-edge message passing process, which better captures the spatial dependency of regional features. Such a strategy essentially upgrades the local aggregation scheme from first order (e.g., from edge to node) to a higher order (e.g., from volume and edge to node), which takes advantage of volumetric information in message passing. Meanwhile, a novel feature-enhanced block is designed to further improve the model's performance and alleviate the over-smoothness problem. Extensive experiments on various PDE systems and one real-world dataset demonstrate that CeFeGNN achieves superior performance compared with other baselines.
Authors: Minoh Jeong, Zae Myung Kim, Min Namgung, Dongyeop Kang, Yao-Yi Chiang, Alfred Hero
Abstract: A unified representation space in multi-modal learning is essential for effectively integrating diverse data sources, such as text, images, and audio, to enhance efficiency and performance across various downstream tasks. Recent binding methods, such as ImageBind, typically rely on a single, fixed anchor modality for aligning multi-modal data. We mathematically analyze these fixed anchor binding methods and uncover significant limitations: (1) over-reliance on the choice of the anchor modality, (2) inadequate capture of intra-modal information, and (3) failure to account for cross-modal correlation among non-anchored modalities. To address these issues, we propose the need for adaptive anchor binding methods, exemplified by our framework CentroBind. The proposed method uses adaptively adjustable centroid-based anchors generated from all available modalities, leading to a balanced and rich representation space. We theoretically demonstrate that our approach captures three critical properties of multi-modal learning -- intra-modal learning, inter-modal learning, and multi-modal alignment -- while constructing a unified representation that spans all modalities. Experiments on both synthetic and real-world datasets show that adaptive anchor methods such as CentroBind consistently outperform fixed anchor binding methods, verifying our analysis.
Authors: Felix Zimmer, Patrik Okanovic, Torsten Hoefler
Abstract: There is an ongoing effort to develop feature selection algorithms to improve interpretability, reduce computational resources, and minimize overfitting in predictive models. Neural networks stand out as architectures on which to build feature selection methods, and recently, neuron pruning and regrowth have emerged from the sparse neural network literature as promising new tools. We introduce EntryPrune, a novel supervised feature selection algorithm using a dense neural network with a dynamic sparse input layer. It employs entry-based pruning, a novel approach that compares neurons based on their relative change induced when they have entered the network. Extensive experiments on 13 different datasets show that our approach generally outperforms the current state-of-the-art methods, and in particular improves the average accuracy on low-dimensional datasets. Furthermore, we show that EntryPruning surpasses traditional techniques such as magnitude pruning within the EntryPrune framework and that EntryPrune achieves lower runtime than competing approaches. Our code is available at https://github.com/flxzimmer/entryprune.
Authors: Hanbo Huang, Yihan Li, Bowen Jiang, Bo Jiang, Lin Liu, Ruoyu Sun, Zhuotao Liu, Shiyu Liang
Abstract: Privacy-sensitive users require deploying large language models (LLMs) within their own infrastructure (on-premises) to safeguard private data and enable customization. However, vulnerabilities in local environments can lead to unauthorized access and potential model theft. To address this, prior research on small models has explored securing only the output layer within hardware-secured devices to balance model confidentiality and customization. Yet this approach fails to protect LLMs effectively. In this paper, we discover that (1) query-based distillation attacks targeting the secured top layer can produce a functionally equivalent replica of the victim model; (2) securing the same number of layers, bottom layers before a transition layer provide stronger protection against distillation attacks than top layers, with comparable effects on customization performance; and (3) the number of secured layers creates a trade-off between protection and customization flexibility. Based on these insights, we propose SOLID, a novel deployment framework that secures a few bottom layers in a secure environment and introduces an efficient metric to optimize the trade-off by determining the ideal number of hidden layers. Extensive experiments on five models (1.3B to 70B parameters) demonstrate that SOLID outperforms baselines, achieving a better balance between protection and downstream customization.
Authors: Natasha Butt, Varun Chandrasekaran, Neel Joshi, Besmira Nushi, Vidhisha Balachandran
Abstract: Evaluation insights are limited by the availability of high-quality benchmarks. As models evolve, there is a need to create benchmarks that can measure progress on new and complex generative capabilities. However, manually creating new benchmarks is slow and expensive, restricting comprehensive evaluations for any capability. We introduce BenchAgents, a multi-agent framework that methodically leverages large language models (LLMs) to automate evaluation benchmark creation while inherently ensuring data and (evaluation) metric quality. BenchAgents decomposes the benchmark creation process into planning, generation, verification, and evaluation, each of which is ] orchestrated via LLM agents. These agents interact with each other and utilize feedback from benchmark developers to improve and flexibly control data diversity and quality. We use BenchAgents to create benchmarks to evaluate capabilities related to planning, constraint satisfaction, and causal reasoning spanning both language and vision modalities. We then use these benchmarks to study state-of-the-art models and extract new insights into common failure modes and model differences.
Authors: Md. Saiful Bari Siddiqui, Md Mohaiminul Islam, Md. Golam Rabiul Alam
Abstract: Overfitting remains a significant challenge in deep learning, often arising from data outliers, noise, and limited training data. To address this, the Divide2Conquer (D2C) method was previously proposed, which partitions training data into multiple subsets and trains identical models independently on each. This strategy enables learning more consistent patterns while minimizing the influence of individual outliers and noise. However, D2C's standard aggregation typically treats all subset models equally or based on fixed heuristics (like data size), potentially underutilizing information about their varying generalization capabilities. Building upon this foundation, we introduce Dynamic Uncertainty-Aware Divide2Conquer (DUA-D2C), an advanced technique that refines the aggregation process. DUA-D2C dynamically weights the contributions of subset models based on their performance on a shared validation set, considering both accuracy and prediction uncertainty. This intelligent aggregation allows the central model to preferentially learn from subsets yielding more generalizable and confident edge models, thereby more effectively combating overfitting. Empirical evaluations on benchmark datasets spanning multiple domains demonstrate that DUA-D2C significantly improves generalization. Our analysis includes evaluations of decision boundaries, loss curves, and other performance metrics, highlighting the effectiveness of DUA-D2C. This study demonstrates that DUA-D2C improves generalization performance even when applied on top of other regularization methods, establishing it as a theoretically grounded and effective approach to combating overfitting in modern deep learning. Our codes are publicly available at: https://github.com/Saiful185/DUA-D2C.
Authors: Danny P. Sawyer, Nan Rosemary Ke, Hubert Soyer, Martin Engelcke, David P Reichert, Drew A. Hudson, John Reid, Alexander Lerchner, Danilo Jimenez Rezende, Timothy P Lillicrap, Michael Mozer, Jane X Wang
Abstract: Foundation models excel at single-turn reasoning but struggle with multi-turn exploration in dynamic environments, a requirement for many real-world challenges. We evaluated these models on their ability to learn from experience, adapt, and gather information. First, in "Feature World," a simple setting for testing information gathering, models performed near-optimally. However, to test more complex, multi-trial learning, we implemented a text-based version of the "Alchemy" environment, a benchmark for meta-learning. Here, agents must deduce a latent causal structure by integrating information across many trials. In this setting, recent foundation models initially failed to improve their performance over time. Crucially, we found that prompting the models to summarize their observations at regular intervals enabled an emergent meta-learning process. This allowed them to improve across trials and even adaptively re-learn when the environment's rules changed unexpectedly. While most models handled the simple task, Alchemy revealed stark differences in robustness: Gemini 2.5 performed best, followed by Claude 3.7, while ChatGPT-4o and o4-mini struggled. This underscores Alchemy's value as a benchmark. Our findings demonstrate that the biggest challenge for foundation models is not selecting informative actions in the moment, but integrating knowledge through adaptive strategies over time. Encouragingly, there appears to be no intrinsic barrier to future models mastering these abilities.
Authors: Damjan Kalajdzievski
Abstract: The field of mechanistic interpretability in pre-trained transformer models has demonstrated substantial evidence supporting the ''linear representation hypothesis'', which is the idea that high level concepts are encoded as vectors in the space of activations of a model. Studies also show that model generation behavior can be steered toward a given concept by adding the concept's vector to the corresponding activations. We show how to leverage these properties to build a form of logical implication into models, enabling transparent and interpretable adjustments that induce a chosen generation behavior in response to the presence of any given concept. Our method, Logical Implication Model Steering (LIMS), unlocks new hand engineered reasoning capabilities by integrating neuro-symbolic logic into pre-trained transformer models.
Authors: Yiming Huang, Tolga Birdal
Abstract: Graph generation is a critical yet challenging task as empirical analyses require a deep understanding of complex, non-Euclidean structures. Diffusion models have recently made significant achievements in graph generation, but these models are typically adapted from image generation frameworks and overlook inherent higher-order topology, leaving them ill-suited for capturing the topological properties of graphs. In this work, we propose Higher-order Guided Diffusion (HOG-Diff), a principled framework that progressively generates plausible graphs with inherent topological structures. HOG-Diff follows a coarse-to-fine generation curriculum guided by higher-order topology and implemented via diffusion bridges. We further prove that our model exhibits a stronger theoretical guarantee than classical diffusion frameworks. Extensive experiments on both molecular and generic graph generation tasks demonstrate that our method consistently outperforms or remains competitive with state-of-the-art baselines. Our code is available at https://github.com/Yiminghh/HOG-Diff.
Authors: Sean McLeish, John Kirchenbauer, David Yu Miller, Siddharth Singh, Abhinav Bhatele, Micah Goldblum, Ashwinee Panda, Tom Goldstein
Abstract: Scaling laws are typically fit using a family of models with a narrow range of frozen hyperparameter choices. In this work we study scaling laws using multiple architectural shapes and hyperparameter choices, highlighting their impact on resulting prescriptions. As a primary artifact of our research, we release the Gemstones: an open-source scaling law dataset, consisting of over 4000 checkpoints from transformers with up to 2 billion parameters and diverse architectural shapes; including ablations over learning rate and cooldown. Our checkpoints enable more complex studies of scaling, such as analyzing the relationship between width and depth. By examining our model suite, we find that the prescriptions of scaling laws can be highly sensitive to the experimental design process and the specific model checkpoints used during fitting.
Authors: Tianyu Zhao, Mahmoud Srewa, Salma Elmalaki
Abstract: Ensuring fairness in machine learning extends to the critical dimension of privacy, particularly in human-centric federated learning (FL) settings where decentralized data necessitates an equitable distribution of privacy risk across clients. This paper introduces FinP, a novel framework specifically designed to address disparities in privacy risk by mitigating disproportionate vulnerability to source inference attacks (SIA). FinP employs a two-pronged strategy: (1) server-side adaptive aggregation, which dynamically adjusts client contributions to the global model to foster fairness, and (2) client-side regularization, which enhances the privacy robustness of individual clients. This comprehensive approach directly tackles both the symptoms and underlying causes of privacy unfairness in FL. Extensive evaluations on the Human Activity Recognition (HAR) and CIFAR-10 datasets demonstrate FinP's effectiveness, achieving improvement in fairness-in-privacy on HAR and CIFAR-10 with minimal impact on utility. FinP improved group fairness with respect to disparity in privacy risk using equal opportunity in CIFAR-10 by 57.14% compared to the state-of-the-art. Furthermore, FinP significantly mitigates SIA risks on CIFAR-10, underscoring its potential to establish fairness in privacy within FL systems without compromising utility.
Authors: Jiwan Chung, Saejin Kim, Yongrae Jo, Jaewoo Park, Dongjun Min, Youngjae Yu
Abstract: As large language models expand beyond natural language to domains such as mathematics, multimodal understanding, and embodied agents, tokens increasingly reflect metric relationships rather than purely linguistic meaning. We introduce DIST2Loss, a distance-aware framework designed to train autoregressive discrete models by leveraging predefined distance relationships among output tokens. At its core, DIST2Loss transforms continuous exponential family distributions derived from inherent distance metrics into discrete, categorical optimization targets compatible with the models' architectures. This approach enables the models to learn and preserve meaningful distance relationships during token generation while maintaining compatibility with existing architectures. Empirical evaluations show consistent performance gains in diverse multimodal applications, including visual grounding, robotic manipulation, generative reward modeling, and image generation using vector-quantized features. These improvements are most notable in low-data regimes, demonstrating DIST2Loss's strength under resource constraints.
Authors: Claudio Spiess, Mandana Vaziri, Louis Mandel, Martin Hirzel
Abstract: The performance of large language models (LLMs) depends on how they are prompted, with choices spanning both the high-level prompting pattern (e.g., Zero-Shot, CoT, ReAct, ReWOO) and the specific prompt content (instructions and few-shot demonstrations). Manually tuning this combination is tedious, error-prone, and specific to a given LLM and task. Therefore, this paper proposes AutoPDL, an automated approach to discovering good LLM agent configurations. Our approach frames this as a structured AutoML problem over a combinatorial space of agentic and non-agentic prompting patterns and demonstrations, using successive halving to efficiently navigate this space. We introduce a library implementing common prompting patterns using the PDL prompt programming language. AutoPDL solutions are human-readable, editable, and executable PDL programs that use this library. This approach also enables source-to-source optimization, allowing human-in-the-loop refinement and reuse. Evaluations across three tasks and seven LLMs (ranging from 3B to 70B parameters) show consistent accuracy gains ($9.21\pm15.46$ percentage points), up to 67.5pp, and reveal that selected prompting strategies vary across models and tasks.
Authors: Nima Fathi, Torsten Scholak, Pierre-Andr\'e No\"el
Abstract: We present significant extensions to diffusion-based sequence generation models, blurring the line with autoregressive language models. We introduce hyperschedules, which assign distinct noise schedules to individual token positions, generalizing both autoregressive models (e.g., GPT) and conventional diffusion models (e.g., SEDD, MDLM) as special cases. Second, we propose two hybrid token-wise noising processes that interpolate between absorbing and uniform processes, enabling the model to fix past mistakes, and we introduce a novel inference algorithm that leverages this new feature in a simplified context inspired from MDLM. To support efficient training and inference, we design attention masks compatible with KV-caching. Our methods achieve state-of-the-art perplexity and generate diverse, high-quality sequences across standard benchmarks, suggesting a promising path for autoregressive diffusion-based sequence generation. See code and resources at https://hdlm-colm.github.io/
Authors: Andreas Hochlehnert, Hardik Bhatnagar, Vishaal Udandarao, Samuel Albanie, Ameya Prabhu, Matthias Bethge
Abstract: Reasoning has emerged as the next major frontier for language models (LMs), with rapid advances from both academic and industrial labs. However, this progress often outpaces methodological rigor, with many evaluations relying on benchmarking practices that lack transparency, robustness, or statistical grounding. In this work, we conduct a comprehensive empirical study and find that current mathematical reasoning benchmarks are highly sensitive to subtle implementation choices--including decoding parameters, random seeds, prompt formatting, and even hardware and software configurations. Performance gains reported in recent studies frequently hinge on unclear comparisons or unreported sources of variance. To address these issues, we propose a standardized evaluation framework with clearly defined best practices and reporting standards. Using this framework, we reassess recent methods and find that most reinforcement learning (RL) approaches yield only modest improvements--far below prior claims--and are prone to overfitting, especially on small-scale benchmarks like AIME'24. In contrast, supervised finetuning (SFT) methods show consistently stronger generalization in the settings we study. To foster reproducibility, we release all code, prompts, and model outputs, for reasoning benchmarks, establishing more rigorous foundations for future work.
Authors: Hong Yang, Qi Yu, Travis Desell
Abstract: Out-of-distribution (OOD) detection methods have recently become more prominent, serving as a core element in safety-critical autonomous systems. One major purpose of OOD detection is to reject invalid inputs that could lead to unpredictable errors and compromise safety. Due to the cost of labeled data, recent works have investigated the feasibility of self-supervised learning (SSL) OOD detection, unlabeled OOD detection, and zero shot OOD detection. In this work, we identify a set of conditions for a theoretical guarantee of failure in unlabeled OOD detection algorithms from an information-theoretic perspective. These conditions are present in all OOD tasks dealing with real-world data: I) we provide theoretical proof of unlabeled OOD detection failure when there exists zero mutual information between the learning objective and the in-distribution labels, a.k.a. 'label blindness', II) we define a new OOD task - Adjacent OOD detection - that tests for label blindness and accounts for a previously ignored safety gap in all OOD detection benchmarks, and III) we perform experiments demonstrating that existing unlabeled OOD methods fail under conditions suggested by our label blindness theory and analyze the implications for future research in unlabeled OOD methods.
Authors: Ram\'on Fernandez Astudillo, Md Arafat Sultan, Aashka Trivedi, Yousef El-Kurdi, Tahira Naseem, Radu Florian, Salim Roukos
Abstract: Inference scaling helps LLMs solve complex reasoning problems through extended runtime computation. On top of long chain-of-thought (long-CoT) models, purely inference-time techniques such as best-of-N (BoN) sampling, majority voting, or more generally, minimum Bayes risk decoding (MBRD), can further improve LLM accuracy by generating multiple candidate solutions and aggregating over them. These methods typically leverage additional signals in the form of reward models and risk/similarity functions that compare generated samples, e.g., exact match in some normalized space or standard similarity metrics such as Rouge. Here we present a novel method for incorporating reward and risk/similarity signals into MBRD. Based on the concept of optimal policy in KL-controlled reinforcement learning, our framework provides a simple and well-defined mechanism for leveraging such signals, offering several advantages over traditional inference-time methods: higher robustness, improved accuracy, and well-understood asymptotic behavior. In addition, it allows for the development of a sample-efficient variant of MBRD that can adjust the number of samples to generate according to the difficulty of the problem, without relying on majority vote counts. We empirically demonstrate the advantages of our approach on math (MATH-$500$) and coding (HumanEval) tasks using recent open-source models. We also present a comprehensive analysis of its accuracy-compute trade-offs.
Authors: Yifei Wang, Weimin Bai, Colin Zhang, Debing Zhang, Weijian Luo, He Sun
Abstract: In this paper, we unify more than 10 existing one-step diffusion distillation approaches, such as Diff-Instruct, DMD, SIM, SiD, $f$-distill, etc, inside a theory-driven framework which we name the \textbf{\emph{Uni-Instruct}}. Uni-Instruct is motivated by our proposed diffusion expansion theory of the $f$-divergence family. Then we introduce key theories that overcome the intractability issue of the original expanded $f$-divergence, resulting in an equivalent yet tractable loss that effectively trains one-step diffusion models by minimizing the expanded $f$-divergence family. The novel unification introduced by Uni-Instruct not only offers new theoretical contributions that help understand existing approaches from a high-level perspective but also leads to state-of-the-art one-step diffusion generation performances. On the CIFAR10 generation benchmark, Uni-Instruct achieves record-breaking Frechet Inception Distance (FID) values of \textbf{\emph{1.46}} for unconditional generation and \textbf{\emph{1.38}} for conditional generation. On the ImageNet-$64\times 64$ generation benchmark, Uni-Instruct achieves a new SoTA one-step generation FID of \textbf{\emph{1.02}}, which outperforms its 79-step teacher diffusion with a significant improvement margin of 1.33 (1.02 vs 2.35). We also apply Uni-Instruct on broader tasks like text-to-3D generation. For text-to-3D generation, Uni-Instruct gives decent results, which slightly outperforms previous methods, such as SDS and VSD, in terms of both generation quality and diversity. Both the solid theoretical and empirical contributions of Uni-Instruct will potentially help future studies on one-step diffusion distillation and knowledge transferring of diffusion models.
Authors: Yuzhen Huang, Weihao Zeng, Xingshan Zeng, Qi Zhu, Junxian He
Abstract: Trustworthy verifiers are essential for the success of reinforcement learning with verifiable reward (RLVR), which is the core methodology behind various large reasoning models such as DeepSeek-R1. In complex domains like mathematical reasoning, rule-based verifiers have been widely adopted in previous works to train strong reasoning models. However, the reliability of these verifiers and their impact on the RL training process remain poorly understood. In this work, we take mathematical reasoning as a case study and conduct a comprehensive analysis of various verifiers in both static evaluation and RL training scenarios. First, we find that current open-source rule-based verifiers often fail to recognize equivalent answers presented in different formats across multiple commonly used mathematical datasets, resulting in non-negligible false negative rates. This limitation adversely affects RL training performance and becomes more pronounced as the policy model gets stronger. Subsequently, we investigate model-based verifiers as a potential solution to address these limitations. While the static evaluation shows that model-based verifiers achieve significantly higher verification accuracy, further analysis and RL results imply that they are highly susceptible to hacking, where they misclassify certain patterns in responses as correct, particularly after fine-tuning. This vulnerability is exploited during policy model optimization, leading to artificially inflated rewards. Our findings underscore the unique challenges inherent to both rule-based and model-based verifiers and provide insights toward developing more accurate and robust reward systems for reinforcement learning.
Authors: Alireza Jafari, Fereshteh Yousefirizi, Vahid Seydi
Abstract: Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with elevated health risks, where timely detection is pivotal for mitigating stroke-related morbidity. This study introduces an innovative hybrid methodology integrating unsupervised deep learning and gradient boosting models to improve AF detection. A 19-layer deep convolutional autoencoder (DCAE) is coupled with three boosting classifiers-AdaBoost, XGBoost, and LightGBM (LGBM)-to harness their complementary advantages while addressing individual limitations. The proposed framework uniquely combines DCAE with gradient boosting, enabling end-to-end AF identification devoid of manual feature extraction. The DCAE-LGBM model attains an F1-score of 95.20%, sensitivity of 99.99%, and inference latency of four seconds, outperforming existing methods and aligning with clinical deployment requirements. The DCAE integration significantly enhances boosting models, positioning this hybrid system as a reliable tool for automated AF detection in clinical settings.
Authors: David Steinmann, Wolfgang Stammer, Antonia W\"ust, Kristian Kersting
Abstract: Developing high-performing, yet interpretable models remains a critical challenge in modern AI. Concept-based models (CBMs) attempt to address this by extracting human-understandable concepts from a global encoding (e.g., image encoding) and then applying a linear classifier on the resulting concept activations, enabling transparent decision-making. However, their reliance on holistic image encodings limits their expressiveness in object-centric real-world settings and thus hinders their ability to solve complex vision tasks beyond single-label classification. To tackle these challenges, we introduce Object-Centric Concept Bottlenecks (OCB), a framework that combines the strengths of CBMs and pre-trained object-centric foundation models, boosting performance and interpretability. We evaluate OCB on complex image datasets and conduct a comprehensive ablation study to analyze key components of the framework, such as strategies for aggregating object-concept encodings. The results show that OCB outperforms traditional CBMs and allows one to make interpretable decisions for complex visual tasks.
Authors: Lorenzo Steccanella, Joshua B. Evans, \"Ozg\"ur \c{S}im\c{s}ek, Anders Jonsson
Abstract: This paper presents a state representation framework for Markov decision processes (MDPs) that can be learned solely from state trajectories, requiring neither reward signals nor the actions executed by the agent. We propose learning the minimum action distance (MAD), defined as the minimum number of actions required to transition between states, as a fundamental metric that captures the underlying structure of an environment. MAD naturally enables critical downstream tasks such as goal-conditioned reinforcement learning and reward shaping by providing a dense, geometrically meaningful measure of progress. Our self-supervised learning approach constructs an embedding space where the distances between embedded state pairs correspond to their MAD, accommodating both symmetric and asymmetric approximations. We evaluate the framework on a comprehensive suite of environments with known MAD values, encompassing both deterministic and stochastic dynamics, as well as discrete and continuous state spaces, and environments with noisy observations. Empirical results demonstrate that the proposed approach not only efficiently learns accurate MAD representations across these diverse settings but also significantly outperforms existing state representation methods in terms of representation quality.
Authors: Minoh Jeong, Alfred Hero
Abstract: Self-supervised contrastive learning (SSCL) has emerged as a powerful paradigm for representation learning and has been studied from multiple perspectives, including mutual information and geometric viewpoints. However, supervised contrastive (SupCon) approaches have received comparatively little attention in this context: for instance, while InfoNCE used in SSCL is known to form a lower bound on mutual information (MI), the relationship between SupCon and MI remains unexplored. To address this gap, we introduce ProjNCE, a generalization of the InfoNCE loss that unifies supervised and self-supervised contrastive objectives by incorporating projection functions and an adjustment term for negative pairs. We prove that ProjNCE constitutes a valid MI bound and affords greater flexibility in selecting projection strategies for class embeddings. Building on this flexibility, we further explore the centroid-based class embeddings in SupCon by exploring a variety of projection methods. Extensive experiments on image and audio datasets demonstrate that ProjNCE consistently outperforms both SupCon and standard cross-entropy training. Our work thus refines SupCon along two complementary perspectives--information-theoretic and projection viewpoints--and offers broadly applicable improvements whenever SupCon serves as the foundational contrastive objective.
Authors: Minoh Jeong, Seonho Kim, Alfred Hero
Abstract: Deterministic embeddings learned by contrastive learning (CL) methods such as SimCLR and SupCon achieve state-of-the-art performance but lack a principled mechanism for uncertainty quantification. We propose Variational Contrastive Learning (VCL), a decoder-free framework that maximizes the evidence lower bound (ELBO) by interpreting the InfoNCE loss as a surrogate reconstruction term and adding a KL divergence regularizer to a uniform prior on the unit hypersphere. We model the approximate posterior $q_\theta(z|x)$ as a projected normal distribution, enabling the sampling of probabilistic embeddings. Our two instantiation--VSimCLR and VSupCon--replace deterministic embeddings with samples from $q_\theta(z|x)$ and incorporate a normalized KL term into the loss. Experiments on multiple benchmarks demonstrate that VCL mitigates dimensional collapse, enhances mutual information with class labels, and matches or outperforms deterministic baselines in classification accuracy, all the while providing meaningful uncertainty estimates through the posterior model. VCL thus equips contrastive learning with a probabilistic foundation, serving as a new basis for contrastive approaches.
Authors: Miles Wang, Tom Dupr\'e la Tour, Olivia Watkins, Alex Makelov, Ryan A. Chi, Samuel Miserendino, Jeffrey Wang, Achyuta Rajaram, Johannes Heidecke, Tejal Patwardhan, Dan Mossing
Abstract: Understanding how language models generalize behaviors from their training to a broader deployment distribution is an important problem in AI safety. Betley et al. discovered that fine-tuning GPT-4o on intentionally insecure code causes "emergent misalignment," where models give stereotypically malicious responses to unrelated prompts. We extend this work, demonstrating emergent misalignment across diverse conditions, including reinforcement learning on reasoning models, fine-tuning on various synthetic datasets, and in models without safety training. To investigate the mechanisms behind this generalized misalignment, we apply a "model diffing" approach using sparse autoencoders to compare internal model representations before and after fine-tuning. This approach reveals several "misaligned persona" features in activation space, including a toxic persona feature which most strongly controls emergent misalignment and can be used to predict whether a model will exhibit such behavior. Additionally, we investigate mitigation strategies, discovering that fine-tuning an emergently misaligned model on just a few hundred benign samples efficiently restores alignment.
Authors: Seyedsaman Emami, Gonzalo Mart\'inez-Mu\~noz, Daniel Hern\'andez-Lobato
Abstract: Multi-task learning (MTL) has shown effectiveness in exploiting shared information across tasks to improve generalization. MTL assumes tasks share similarities that can improve performance. In addition, boosting algorithms have demonstrated exceptional performance across diverse learning problems, primarily due to their ability to focus on hard-to-learn instances and iteratively reduce residual errors. This makes them a promising approach for learning multi-task problems. However, real-world MTL scenarios often involve tasks that are not well-aligned (known as outlier or adversarial tasks), which do not share beneficial similarities with others and can, in fact, deteriorate the performance of the overall model. To overcome this challenge, we propose Robust-Multi-Task Gradient Boosting (R-MTGB), a novel boosting framework that explicitly models and adapts to task heterogeneity during training. R-MTGB structures the learning process into three sequential blocks: (1) learning shared patterns, (2) partitioning tasks into outliers and non-outliers with regularized parameters, and (3) fine-tuning task-specific predictors. This architecture enables R-MTGB to automatically detect and penalize outlier tasks while promoting effective knowledge transfer among related tasks. Our method integrates these mechanisms seamlessly within gradient boosting, allowing robust handling of noisy or adversarial tasks without sacrificing accuracy. Extensive experiments on both synthetic benchmarks and real-world datasets demonstrate that our approach successfully isolates outliers, transfers knowledge, and consistently reduces prediction errors for each task individually, and achieves overall performance gains across all tasks. These results highlight robustness, adaptability, and reliable convergence of R-MTGB in challenging MTL environments.
Authors: Guofu Xie, Yunsheng Shi, Hongtao Tian, Ting Yao, Xiao Zhang
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning abilities of Large Language Models (LLMs) by using rule-based binary feedback. However, current RLVR methods typically assign the same reward to every token. This coarse-grained feedback hampers precise credit assignment, making it hard for models to identify which reasoning steps lead to success or failure, and often results in suboptimal policies. Methods like PPO provide credit assignment by value estimation, but yield inaccurate and unverifiable signals due to limited sampling. On the other hand, methods using Process Reward Models can provide step-wise rewards but suffer from several key limitations: they require high-quality process supervision labels, the feedback is unreliable due to probabilistic reward modeling, and their application in online reinforcement learning (RL) is time-consuming. To overcome these limitations, we introduce a simple but efficient method-Credit Assignment Policy Optimization (CAPO). Instead of training auxiliary models, CAPO directly leverages an off-the-shelf, general-purpose LLM as a Generative Process Reward Model (LLM-as-GenPRM) to generate all step-wise critique by one pass only based on the correctness of the step itself, providing deterministic token-level credits to refine the tokens that were originally assigned identical rule-based rewards. To further enhance the accuracy and robustness, we employ voting mechanisms that scale with the number of generated critiques. Extensive experiments on various backbones like Llama and Qwen models show that CAPO consistently outperforms supervised learning-based and RL-based fine-tuning methods across four challenging mathematical benchmarks and three out-of-domain benchmarks. Further analysis shows that CAPO can help the model to foster the learning of correct reasoning pathways leading to correct answers.
Authors: Zhengran Ji, Boyuan Chen
Abstract: Training reinforcement learning agents with human feedback is crucial when task objectives are difficult to specify through dense reward functions. While prior methods rely on offline trajectory comparisons to elicit human preferences, such data is unavailable in online learning scenarios where agents must adapt on the fly. Recent approaches address this by collecting real-time scalar feedback to guide agent behavior and train reward models for continued learning after human feedback becomes unavailable. However, scalar feedback is often noisy and inconsistent, limiting the accuracy and generalization of learned rewards. We propose Pref-GUIDE, a framework that transforms real-time scalar feedback into preference-based data to improve reward model learning for continual policy training. Pref-GUIDE Individual mitigates temporal inconsistency by comparing agent behaviors within short windows and filtering ambiguous feedback. Pref-GUIDE Voting further enhances robustness by aggregating reward models across a population of users to form consensus preferences. Across three challenging environments, Pref-GUIDE significantly outperforms scalar-feedback baselines, with the voting variant exceeding even expert-designed dense rewards. By reframing scalar feedback as structured preferences with population feedback, Pref-GUIDE offers a scalable and principled approach for harnessing human input in online reinforcement learning.
Authors: Nikolai Antonov, Pr\v{e}mysl \v{S}\r{u}cha, Mikol\'a\v{s} Janota, Jan H\r{u}la
Abstract: Existing research on single-machine scheduling is largely focused on exact algorithms, which perform well on typical instances but can significantly deteriorate on certain regions of the problem space. In contrast, data-driven approaches provide strong and scalable performance when tailored to the structure of specific datasets. Leveraging this idea, we focus on a single-machine scheduling problem where each job is defined by its weight, duration, due date, and deadline, aiming to minimize the total weight of tardy jobs. We introduce a novel data-driven scheduling heuristic that combines machine learning with problem-specific characteristics, ensuring feasible solutions, which is a common challenge for ML-based algorithms. Experimental results demonstrate that our approach significantly outperforms the state-of-the-art in terms of optimality gap, number of optimal solutions, and adaptability across varied data scenarios, highlighting its flexibility for practical applications. In addition, we conduct a systematic exploration of ML models, addressing a common gap in similar studies by offering a detailed model selection process and providing insights into why the chosen model is the best fit.
Authors: Jonathan N\"other, Adish Singla, Goran Radanovic
Abstract: Ensuring the safe use of agentic systems requires a thorough understanding of the range of malicious behaviors these systems may exhibit when under attack. In this paper, we evaluate the robustness of LLM-based agentic systems against attacks that aim to elicit harmful actions from agents. To this end, we propose a novel taxonomy of harms for agentic systems and a novel benchmark, BAD-ACTS, for studying the security of agentic systems with respect to a wide range of harmful actions. BAD-ACTS consists of 4 implementations of agentic systems in distinct application environments, as well as a dataset of 188 high-quality examples of harmful actions. This enables a comprehensive study of the robustness of agentic systems across a wide range of categories of harmful behaviors, available tools, and inter-agent communication structures. Using this benchmark, we analyze the robustness of agentic systems against an attacker that controls one of the agents in the system and aims to manipulate other agents to execute a harmful target action. Our results show that the attack has a high success rate, demonstrating that even a single adversarial agent within the system can have a significant impact on the security. This attack remains effective even when agents use a simple prompting-based defense strategy. However, we additionally propose a more effective defense based on message monitoring. We believe that this benchmark provides a diverse testbed for the security research of agentic systems. The benchmark can be found at github.com/JNoether/BAD-ACTS
Authors: Sotaro Takeshita, Yurina Takeshita, Daniel Ruffinelli, Simone Paolo Ponzetto
Abstract: In this paper, we study the surprising impact that truncating text embeddings has on downstream performance. We consistently observe across 6 state-of-the-art text encoders and 26 downstream tasks, that randomly removing up to 50% of embedding dimensions results in only a minor drop in performance, less than 10%, in retrieval and classification tasks. Given the benefits of using smaller-sized embeddings, as well as the potential insights about text encoding, we study this phenomenon and find that, contrary to what is suggested in prior work, this is not the result of an ineffective use of representation space. Instead, we find that a large number of uniformly distributed dimensions actually cause an increase in performance when removed. This would explain why, on average, removing a large number of embedding dimensions results in a marginal drop in performance. We make similar observations when truncating the embeddings used by large language models to make next-token predictions on generative tasks, suggesting that this phenomenon is not isolated to classification or retrieval tasks.
Authors: Qiying Hu, Yingying Liang, Qianli Zhou, Witold Pedrycz
Abstract: Dempster-Shafer Theory (DST) provides a powerful framework for modeling uncertainty and has been widely applied to multi-attribute classification tasks. However, traditional DST-based attribute fusion-based classifiers suffer from oversimplified membership function modeling and limited exploitation of the belief structure brought by basic probability assignment (BPA), reducing their effectiveness in complex real-world scenarios. This paper presents an enhanced attribute fusion-based classifier that addresses these limitations through two key innovations. First, we adopt a selective modeling strategy that utilizes both single Gaussian and Gaussian Mixture Models (GMMs) for membership function construction, with model selection guided by cross-validation and a tailored evaluation metric. Second, we introduce a novel method to transform the possibility distribution into a BPA by combining simple BPAs derived from normalized possibility distributions, enabling a much richer and more flexible representation of uncertain information. Furthermore, we apply the belief structure-based BPA generation method to the evidential K-Nearest Neighbors (EKNN) classifier, enhancing its ability to incorporate uncertainty information into decision-making. Comprehensive experiments on benchmark datasets are conducted to evaluate the performance of the proposed attribute fusion-based classifier and the enhanced evidential K-Nearest Neighbors classifier in comparison with both evidential classifiers and conventional machine learning classifiers. The results demonstrate that the proposed classifier outperforms the best existing evidential classifier, achieving an average accuracy improvement of 4.86%, while maintaining low variance, thus confirming its superior effectiveness and robustness.
Authors: Monte Hoover, Vatsal Baherwani, Neel Jain, Khalid Saifullah, Joseph Vincent, Chirag Jain, Melissa Kazemi Rad, C. Bayan Bruss, Ashwinee Panda, Tom Goldstein
Abstract: Guardian models play a crucial role in ensuring the safety and ethical behavior of user-facing AI applications by enforcing guardrails and detecting harmful content. While standard guardian models are limited to predefined, static harm categories, we introduce DynaGuard, a suite of dynamic guardian models offering novel flexibility by evaluating text based on user-defined policies, and DynaBench, a dataset for training and evaluating dynamic guardian models. Our models provide both rapid detection of policy violations and a chain-of-thought reasoning option that articulate and justify model outputs. Critically, DynaGuard not only surpasses static models in detection accuracy on traditional safety categories, but is competitive with frontier reasoning models on free-form policy violations, all in a fraction of the time. This makes DynaGuard an critical tool for language model guardrails.
Authors: Mohamed Bal-Ghaoui, Mohammed Tiouti
Abstract: Effective model and hyperparameter selection remains a major challenge in deep learning, often requiring extensive expertise and computation. While AutoML and large language models (LLMs) promise automation, current LLM-based approaches rely on trial and error and expensive APIs, which provide limited interpretability and generalizability. We propose MetaLLMiX, a zero-shot hyperparameter optimization framework combining meta-learning, explainable AI, and efficient LLM reasoning. By leveraging historical experiment outcomes with SHAP explanations, MetaLLMiX recommends optimal hyperparameters and pretrained models without additional trials. We further employ an LLM-as-judge evaluation to control output format, accuracy, and completeness. Experiments on eight medical imaging datasets using nine open-source lightweight LLMs show that MetaLLMiX achieves competitive or superior performance to traditional HPO methods while drastically reducing computational cost. Our local deployment outperforms prior API-based approaches, achieving optimal results on 5 of 8 tasks, response time reductions of 99.6-99.9%, and the fastest training times on 6 datasets (2.4-15.7x faster), maintaining accuracy within 1-5% of best-performing baselines.
Authors: Killian Steunou, Th\'eo Druilhe, Sigurd Saue
Abstract: Deep neural networks perform remarkably well on image classification tasks but remain vulnerable to carefully crafted adversarial perturbations. This work revisits linear dimensionality reduction as a simple, data-adapted defense. We empirically compare standard Principal Component Analysis (PCA) with its sparse variant (SPCA) as front-end feature extractors for downstream classifiers, and we complement these experiments with a theoretical analysis. On the theory side, we derive exact robustness certificates for linear heads applied to SPCA features: for both $\ell_\infty$ and $\ell_2$ threat models (binary and multiclass), the certified radius grows as the dual norms of $W^\top u$ shrink, where $W$ is the projection and $u$ the head weights. We further show that for general (non-linear) heads, sparsity reduces operator-norm bounds through a Lipschitz composition argument, predicting lower input sensitivity. Empirically, with a small non-linear network after the projection, SPCA consistently degrades more gracefully than PCA under strong white-box and black-box attacks while maintaining competitive clean accuracy. Taken together, the theory identifies the mechanism (sparser projections reduce adversarial leverage) and the experiments verify that this benefit persists beyond the linear setting. Our code is available at https://github.com/killian31/SPCARobustness.
Authors: Liangyu Ding, Chenghan Wu, Guokai Li, Zizhuo Wang
Abstract: Bundle pricing refers to designing several product combinations (i.e., bundles) and determining their prices in order to maximize the expected profit. It is a classic problem in revenue management and arises in many industries, such as e-commerce, tourism, and video games. However, the problem is typically intractable due to the exponential number of candidate bundles. In this paper, we explore the usage of graph convolutional networks (GCNs) in solving the bundle pricing problem. Specifically, we first develop a graph representation of the mixed bundling model (where every possible bundle is assigned with a specific price) and then train a GCN to learn the latent patterns of optimal bundles. Based on the trained GCN, we propose two inference strategies to derive high-quality feasible solutions. A local-search technique is further proposed to improve the solution quality. Numerical experiments validate the effectiveness and efficiency of our proposed GCN-based framework. Using a GCN trained on instances with 5 products, our methods consistently achieve near-optimal solutions (better than 97%) with only a fraction of computational time for problems of small to medium size. It also achieves superior solutions for larger size of problems compared with other heuristic methods such as bundle size pricing (BSP). The method can also provide high quality solutions for instances with more than 30 products even for the challenging cases where product utilities are non-additive.
Authors: Yaswanth Chittepu, Prasann Singhal, Greg Durrett, Scott Niekum
Abstract: Margin-based optimization is fundamental to improving generalization and robustness in classification tasks. In the context of reward model learning from preferences within Reinforcement Learning from Human Feedback (RLHF), existing methods typically rely on no margins, fixed margins, or margins that are simplistic functions of preference ratings. However, such formulations often fail to account for the varying strengths of different preferences, for example some preferences are associated with larger margins between responses, or they rely on noisy margin information derived from ratings. We argue that modeling the strength of preferences can lead to better generalization and more faithful alignment. Furthermore, many existing methods that use adaptive margins assume access to accurate preference scores, which can be difficult for humans to provide reliably. We propose an approach that leverages preferences over preferences, that is annotations indicating which of two preferences reflects a stronger distinction. We use this ordinal signal to infer adaptive margins on a per-datapoint basis. We introduce an extension to Direct Preference Optimization (DPO), DPO-PoP, that incorporates adaptive margins from preference-over-preference supervision, enabling improved discriminative and generative performance. Empirically, our method outperforms vanilla DPO, DPO with fixed margins, and DPO with ground-truth margins on the UltraFeedback dataset. Additionally, we show that there is a tradeoff between discriminative and generative performance: improving test classification accuracy, particularly by correctly labeling weaker preferences at the expense of stronger ones, can lead to a decline in generative quality. To navigate this tradeoff, we propose two sampling strategies to gather preference-over-preference labels: one favoring discriminative performance and one favoring generative performance.
Authors: Yuke Li, Yujia Zheng, Tianyi Xiong, Zhenyi Wang, Heng Huang
Abstract: Catastrophic interference, also known as catastrophic forgetting, is a fundamental challenge in machine learning, where a trained learning model progressively loses performance on previously learned tasks when adapting to new ones. In this paper, we aim to better understand and model the catastrophic interference problem from a latent representation learning point of view, and propose a novel theoretical framework that formulates catastrophic interference as an identification problem. Our analysis demonstrates that the forgetting phenomenon can be quantified by the distance between partial-task aware (PTA) and all-task aware (ATA) setups. Building upon recent advances in identifiability theory, we prove that this distance can be minimized through identification of shared latent variables between these setups. When learning, we propose our method \ourmeos with two-stage training strategy: First, we employ maximum likelihood estimation to learn the latent representations from both PTA and ATA configurations. Subsequently, we optimize the KL divergence to identify and learn the shared latent variables. Through theoretical guarantee and empirical validations, we establish that identifying and learning these shared representations can effectively mitigate catastrophic interference in machine learning systems. Our approach provides both theoretical guarantees and practical performance improvements across both synthetic and benchmark datasets.
Authors: Nouha Karaouli, Denis Coquenet, Elisa Fromont, Martial Mermillod, Marina Reyboz
Abstract: Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.
Authors: Dorsa Soleymani, Ali Dadsetan, Frank Rudzicz
Abstract: Differential privacy (DP) provides strong protection for sensitive data, but often reduces model performance and fairness, especially for underrepresented groups. One major reason is gradient clipping in DP-SGD, which can disproportionately suppress learning signals for minority subpopulations. Although adaptive clipping can enhance utility, it still relies on uniform hard clipping, which may restrict fairness. To address this, we introduce SoftAdaClip, a differentially private training method that replaces hard clipping with a smooth, tanh-based transformation to preserve relative gradient magnitudes while bounding sensitivity. We evaluate SoftAdaClip on various datasets, including MIMIC-III (clinical text), GOSSIS-eICU (structured healthcare), and Adult Income (tabular data). Our results show that SoftAdaClip reduces subgroup disparities by up to 87% compared to DP-SGD and up to 48% compared to Adaptive-DPSGD, and these reductions in subgroup disparities are statistically significant. These findings underscore the importance of integrating smooth transformations with adaptive mechanisms to achieve fair and private model training.
Authors: Jaiden Fairoze, Sanjam Garg, Keewoo Lee, Mingyuan Wang
Abstract: As large language models (LLMs) advance, ensuring AI safety and alignment is paramount. One popular approach is prompt guards, lightweight mechanisms designed to filter malicious queries while being easy to implement and update. In this work, we introduce a new attack that circumvents such prompt guards, highlighting their limitations. Our method consistently jailbreaks production models while maintaining response quality, even under the highly protected chat interfaces of Google Gemini (2.5 Flash/Pro), DeepSeek Chat (DeepThink), Grok (3), and Mistral Le Chat (Magistral). The attack exploits a resource asymmetry between the prompt guard and the main LLM, encoding a jailbreak prompt that lightweight guards cannot decode but the main model can. This reveals an attack surface inherent to lightweight prompt guards in modern LLM architectures and underscores the need to shift defenses from blocking malicious inputs to preventing malicious outputs. We additionally identify other critical alignment issues, such as copyrighted data extraction, training data extraction, and malicious response leakage during thinking.
Authors: Haokun Zhao, Xiang Zhang, Jiaqi Wei, Yiwei Xu, Yuting He, Siqi Sun, Chenyu You
Abstract: Time series forecasting is central to decision-making in domains as diverse as energy, finance, climate, and public health. In practice, forecasters face thousands of short, noisy series that vary in frequency, quality, and horizon, where the dominant cost lies not in model fitting, but in the labor-intensive preprocessing, validation, and ensembling required to obtain reliable predictions. Prevailing statistical and deep learning models are tailored to specific datasets or domains and generalize poorly. A general, domain-agnostic framework that minimizes human intervention is urgently in demand. In this paper, we introduce TimeSeriesScientist (TSci), the first LLM-driven agentic framework for general time series forecasting. The framework comprises four specialized agents: Curator performs LLM-guided diagnostics augmented by external tools that reason over data statistics to choose targeted preprocessing; Planner narrows the hypothesis space of model choice by leveraging multi-modal diagnostics and self-planning over the input; Forecaster performs model fitting and validation and, based on the results, adaptively selects the best model configuration as well as ensemble strategy to make final predictions; and Reporter synthesizes the whole process into a comprehensive, transparent report. With transparent natural-language rationales and comprehensive reports, TSci transforms the forecasting workflow into a white-box system that is both interpretable and extensible across tasks. Empirical results on eight established benchmarks demonstrate that TSci consistently outperforms both statistical and LLM-based baselines, reducing forecast error by an average of 10.4% and 38.2%, respectively. Moreover, TSci produces a clear and rigorous report that makes the forecasting workflow more transparent and interpretable.
Authors: Lin Chen, Jun Chen, Minghui Qiu, Shuxin Zhong, Binghong Chen, Kaishun Wu
Abstract: Reconstructing high-resolution rainfall fields is essential for flood forecasting, hydrological modeling, and climate analysis. However, existing spatial interpolation methods-whether based on automatic weather station (AWS) measurements or enhanced with satellite/radar observations often over-smooth critical structures, failing to capture sharp transitions and localized extremes. We introduce RainSeer, a structure-aware reconstruction framework that reinterprets radar reflectivity as a physically grounded structural prior-capturing when, where, and how rain develops. This shift, however, introduces two fundamental challenges: (i) translating high-resolution volumetric radar fields into sparse point-wise rainfall observations, and (ii) bridging the physical disconnect between aloft hydro-meteors and ground-level precipitation. RainSeer addresses these through a physics-informed two-stage architecture: a Structure-to-Point Mapper performs spatial alignment by projecting mesoscale radar structures into localized ground-level rainfall, through a bidirectional mapping, and a Geo-Aware Rain Decoder captures the semantic transformation of hydro-meteors through descent, melting, and evaporation via a causal spatiotemporal attention mechanism. We evaluate RainSeer on two public datasets-RAIN-F (Korea, 2017-2019) and MeteoNet (France, 2016-2018)-and observe consistent improvements over state-of-the-art baselines, reducing MAE by over 13.31% and significantly enhancing structural fidelity in reconstructed rainfall fields.
Authors: Jie Li, Andrew McCarthy, Zhizhuo Zhang, Stephen Young
Abstract: In-context learners like TabPFN are promising for biomolecule efficacy prediction, where established molecular feature sets and relevant experimental results can serve as powerful contextual examples. However, their performance is highly sensitive to the provided context, making strategies like post-hoc ensembling of models trained on different data subsets a viable approach. An open question is how to select the best models for the ensemble without access to ground truth labels. In this study, we investigate an uncertainty-guided strategy for model selection. We demonstrate on an siRNA knockdown efficacy task that a TabPFN model using straightforward sequence-based features can surpass specialized state-of-the-art predictors. We also show that the model's predicted inter-quantile range (IQR), a measure of its uncertainty, has a negative correlation with true prediction error. We developed the OligoICP method, which selects and averages an ensemble of models with the lowest mean IQR for siRNA efficacy prediction, achieving superior performance compared to naive ensembling or using a single model trained on all available data. This finding highlights model uncertainty as a powerful, label-free heuristic for optimizing biomolecule efficacy predictions.
Authors: Liyang Xie, Haoran Zhang, Zhendong Wang, Wesley Tansey, Mingyuan Zhou
Abstract: Diffusion- and flow-based generative models have recently demonstrated strong performance in protein backbone generation tasks, offering unprecedented capabilities for de novo protein design. However, while achieving notable performance in generation quality, these models are limited by their generating speed, often requiring hundreds of iterative steps in the reverse-diffusion process. This computational bottleneck limits their practical utility in large-scale protein discovery, where thousands to millions of candidate structures are needed. To address this challenge, we explore the techniques of score distillation, which has shown great success in reducing the number of sampling steps in the vision domain while maintaining high generation quality. However, a straightforward adaptation of these methods results in unacceptably low designability. Through extensive study, we have identified how to appropriately adapt Score identity Distillation (SiD), a state-of-the-art score distillation strategy, to train few-step protein backbone generators which significantly reduce sampling time, while maintaining comparable performance to their pretrained teacher model. In particular, multistep generation combined with inference time noise modulation is key to the success. We demonstrate that our distilled few-step generators achieve more than a 20-fold improvement in sampling speed, while achieving similar levels of designability, diversity, and novelty as the Proteina teacher model. This reduction in inference cost enables large-scale in silico protein design, thereby bringing diffusion-based models closer to real-world protein engineering applications.
Authors: Lukas Eisenmann, Alena Br\"andle, Zahra Monfared, Daniel Durstewitz
Abstract: Recurrent Neural Networks (RNNs) have found widespread applications in machine learning for time series prediction and dynamical systems reconstruction, and experienced a recent renaissance with improved training algorithms and architectural designs. Understanding why and how trained RNNs produce their behavior is important for scientific and medical applications, and explainable AI more generally. An RNN's dynamical repertoire depends on the topological and geometrical properties of its state space. Stable and unstable manifolds of periodic points play a particularly important role: They dissect a dynamical system's state space into different basins of attraction, and their intersections lead to chaotic dynamics with fractal geometry. Here we introduce a novel algorithm for detecting these manifolds, with a focus on piecewise-linear RNNs (PLRNNs) employing rectified linear units (ReLUs) as their activation function. We demonstrate how the algorithm can be used to trace the boundaries between different basins of attraction, and hence to characterize multistability, a computationally important property. We further show its utility in finding so-called homoclinic points, the intersections between stable and unstable manifolds, and thus establish the existence of chaos in PLRNNs. Finally we show for an empirical example, electrophysiological recordings from a cortical neuron, how insights into the underlying dynamics could be gained through our method.
Authors: Haoqiang Kang, Yizhe Zhang, Nikki Lijing Kuang, Nicklas Majamaki, Navdeep Jaitly, Yi-An Ma, Lianhui Qin
Abstract: Large Language Models (LLMs) demonstrate their reasoning ability through chain-of-thought (CoT) generation. However, LLM's autoregressive decoding may limit the ability to revisit and refine earlier tokens in a holistic manner, which can also lead to inefficient exploration for diverse solutions. In this paper, we propose LaDiR (Latent Diffusion Reasoner), a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. We first construct a structured latent reasoning space using a Variational Autoencoder (VAE) that encodes text reasoning steps into blocks of thought tokens, preserving semantic information and interpretability while offering compact but expressive representations. Subsequently, we utilize a latent diffusion model that learns to denoise a block of latent thought tokens with a blockwise bidirectional attention mask, enabling longer horizon and iterative refinement with adaptive test-time compute. This design allows efficient parallel generation of diverse reasoning trajectories, allowing the model to plan and revise the reasoning process holistically. We conduct evaluations on a suite of mathematical reasoning and planning benchmarks. Empirical results show that LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods, revealing a new paradigm for text reasoning with latent diffusion.
Authors: Johannes Liebenow, Thorsten Peinemann, Esfandiar Mohammadi
Abstract: The tuning of hyperparameters in distributed machine learning can substantially impact model performance. When the hyperparameters are tuned on sensitive data, privacy becomes an important challenge and to this end, differential privacy has emerged as the de facto standard for provable privacy. A standard setting when performing distributed learning tasks is that clients agree on a shared setup, i.e., find a compromise from a set of hyperparameters, like the learning rate of the model to be trained. Yet, prior work on differentially private hyperparameter tuning either uses computationally expensive cryptographic protocols, determines hyperparameters separately for each client, or applies differential privacy locally, which can lead to undesirable utility-privacy trade-offs. In this work, we present our algorithm DP-HYPE, which performs a distributed and privacy-preserving hyperparameter search by conducting a distributed voting based on local hyperparameter evaluations of clients. In this way, DP-HYPE selects hyperparameters that lead to a compromise supported by the majority of clients, while maintaining scalability and independence from specific learning tasks. We prove that DP-HYPE preserves the strong notion of differential privacy called client-level differential privacy and, importantly, show that its privacy guarantees do not depend on the number of hyperparameters. We also provide bounds on its utility guarantees, that is, the probability of reaching a compromise, and implement DP-HYPE as a submodule in the popular Flower framework for distributed machine learning. In addition, we evaluate performance on multiple benchmark data sets in iid as well as multiple non-iid settings and demonstrate high utility of DP-HYPE even under small privacy budgets.
Authors: Jonathan P Williams
Abstract: Conformal prediction (CP) was developed to provide finite-sample probabilistic prediction guarantees. While CP algorithms are a relatively general-purpose approach to uncertainty quantification, with finite-sample guarantees, they lack versatility. Namely, the CP approach does not {\em prescribe} how to quantify the degree to which a data set provides evidence in support of (or against) an arbitrary event from a general class of events. In this paper, tools are offered from imprecise probability theory to build a formal connection between CP and generalized fiducial (GF) inference. These new insights establish a more general inferential lens from which CP can be understood, and demonstrate the pragmatism of fiducial ideas. The formal connection establishes a context in which epistemically-derived GF probability matches aleatoric/frequentist probability. Beyond this fact, it is illustrated how tools from imprecise probability theory, namely lower and upper probability functions, can be applied in the context of the imprecise GF distribution to provide posterior-like, prescriptive inference that is not possible within the CP framework alone. In addition to the primary CP generalization that is contributed, fundamental connections are synthesized between this new model-free GF and three other areas of contemporary research: nonparametric predictive inference (NPI), conformal predictive systems/distributions, and inferential models (IMs).
Authors: Xin Guo, Xinyu Li, Renyuan Xu
Abstract: This paper proposes and analyzes two new policy learning methods: regularized policy gradient (RPG) and iterative policy optimization (IPO), for a class of discounted linear-quadratic control (LQC) problems over an infinite time horizon with entropy regularization. Assuming access to the exact policy evaluation, both proposed approaches are proven to converge linearly in finding optimal policies of the regularized LQC. Moreover, the IPO method can achieve a super-linear convergence rate once it enters a local region around the optimal policy. Finally, when the optimal policy for an RL problem with a known environment is appropriately transferred as the initial policy to an RL problem with an unknown environment, the IPO method is shown to enable a super-linear convergence rate if the two environments are sufficiently close. Performances of these proposed algorithms are supported by numerical examples.
Authors: V\'ictor A. Braberman, Flavia Bonomo-Braberman, Yiannis Charalambous, Juan G. Colonna, Lucas C. Cordeiro, Rosiane de Freitas
Abstract: The emergence of prompting as the dominant paradigm for leveraging Large Language Models (LLMs) has led to a proliferation of LLM-native software, where application behavior arises from complex, stochastic data transformations. However, the engineering of such systems remains largely exploratory and ad-hoc, hampered by the absence of conceptual frameworks, ex-ante methodologies, design guidelines, and specialized benchmarks. We argue that a foundational step towards a more disciplined engineering practice is a systematic understanding of the core functional units--generative transformations--and their compositional patterns within LLM-native applications. Focusing on the rich domain of software verification and falsification, we conduct a secondary study of over 100 research proposals to address this gap. We first present a fine-grained taxonomy of generative transformations, abstracting prompt-based interactions into conceptual signatures. This taxonomy serves as a scaffolding to identify recurrent transformation relationship patterns--analogous to software design patterns--that characterize solution approaches in the literature. Our analysis not only validates the utility of the taxonomy but also surfaces strategic gaps and cross-dimensional relationships, offering a structured foundation for future research in modular and compositional LLM application design, benchmarking, and the development of reliable LLM-native systems.
Authors: Efthymios Costa, Ioanna Papatsouma, Angelos Markos
Abstract: In this paper, we present an information-theoretic method for clustering mixed-type data, that is, data consisting of both continuous and categorical variables. The proposed approach extends the Information Bottleneck principle to heterogeneous data through generalised product kernels, integrating continuous, nominal, and ordinal variables within a unified optimization framework. We address the following challenges: developing a systematic bandwidth selection strategy that equalises contributions across variable types, and proposing an adaptive hyperparameter updating scheme that ensures a valid solution into a predetermined number of potentially imbalanced clusters. Through simulations on 28,800 synthetic data sets and ten publicly available benchmarks, we demonstrate that the proposed method, named DIBmix, achieves superior performance compared to four established methods (KAMILA, K-Prototypes, FAMD with K-Means, and PAM with Gower's dissimilarity). Results show DIBmix particularly excels when clusters exhibit size imbalances, data contain low or moderate cluster overlap, and categorical and continuous variables are equally represented. The method presents a significant advantage over traditional centroid-based algorithms, establishing DIBmix as a competitive and theoretically grounded alternative for mixed-type data clustering.
Authors: Reuben Dorent, Nazim Haouchine, Alexandra Golby, Sarah Frisken, Tina Kapur, William Wells
Abstract: We propose a deep mixture of multimodal hierarchical variational auto-encoders called MMHVAE that synthesizes missing images from observed images in different modalities. MMHVAE's design focuses on tackling four challenges: (i) creating a complex latent representation of multimodal data to generate high-resolution images; (ii) encouraging the variational distributions to estimate the missing information needed for cross-modal image synthesis; (iii) learning to fuse multimodal information in the context of missing data; (iv) leveraging dataset-level information to handle incomplete data sets at training time. Extensive experiments are performed on the challenging problem of pre-operative brain multi-parametric magnetic resonance and intra-operative ultrasound imaging.
Authors: Abdullah Khan Zehady, Shubhashis Roy Dipta, Naymul Islam, Safi Al Mamun, Santu Karmaker
Abstract: Bangla is a language spoken by approximately 240 million native speakers and around 300 million people worldwide. Despite being the 5th largest spoken language in the world, Bangla is still a "low-resource" language, and existing pretrained language models often struggle to perform well on Bangla Language Processing (BLP) tasks. This paper addresses this gap by: (1) introducing two high-quality translated Bangla-instruction datasets totaling 224k samples - Bangla-Orca (172k) and Bangla-Alpaca (52k); and (2) leveraging these datasets to develop BanglaLlama, an open-source family of Bangla-specific LLMs, consisting of five base and instruct variants. We present our methodology, two large datasets, and comprehensive benchmarking results showcasing the effectiveness of our dataset and model on multiple benchmarks. We believe our proposed datasets and models will serve as the new standard baseline for future research focused on this widely spoken yet "low-resource" language.
Authors: Hira Saleem, Flora Salim, Cormac Purcell
Abstract: Physics based numerical climate models serve as critical tools for evaluating the effects of climate change and projecting future climate scenarios. However, the reliance on numerical simulations of physical equations renders them computationally intensive and inefficient. While deep learning methodologies have made significant progress in weather forecasting, they are still unstable for longer roll-out climate emulation task. Here, we propose PACER, a relatively lightweight 2.1M parameter Physics Informed Uncertainty Aware Climate EmulatoR. PACER is trained across is trained across varying spatial resolutions and physics based climate models, enabling faithful and stable emulation of temperature fields at multiple surface levels over a 10 year horizon. We propose an auto-regressive ODE-SDE framework for climate emulation that integrates the fundamental physical law of advection, while being trained under a negative log-likelihood objective to enable principled uncertainty quantification of stochastic variability. We show PACER's emulation performance across 20 climate models outperforming relevant baselines and advancing towards explicit physics infusion in ML emulator.
Authors: Ana Ozaki, Roberto Confalonieri, Ricardo Guimar\~aes, Anders Imenes
Abstract: Decision trees are a popular machine learning method, known for their inherent explainability. In Explainable AI, decision trees can be used as surrogate models for complex black box AI models or as approximations of parts of such models. A key challenge of this approach is determining how accurately the extracted decision tree represents the original model and to what extent it can be trusted as an approximation of their behavior. In this work, we investigate the use of the Probably Approximately Correct (PAC) framework to provide a theoretical guarantee of fidelity for decision trees extracted from AI models. Based on theoretical results from the PAC framework, we adapt a decision tree algorithm to ensure a PAC guarantee under certain conditions. We focus on binary classification and conduct experiments where we extract decision trees from BERT-based language models with PAC guarantees. Our results indicate occupational gender bias in these models.
Authors: Wonjun Lee, Riley C. W. O'Neill, Dongmian Zou, Jeff Calder, Gilad Lerman
Abstract: Generative modeling aims to generate new data samples that resemble a given dataset, with diffusion models recently becoming the most popular generative model. One of the main challenges of diffusion models is solving the problem in the input space, which tends to be very high-dimensional. Recently, solving diffusion models in the latent space through an encoder that maps from the data space to a lower-dimensional latent space has been considered to make the training process more efficient and has shown state-of-the-art results. The variational autoencoder (VAE) is the most commonly used encoder/decoder framework in this domain, known for its ability to learn latent representations and generate data samples. In this paper, we introduce a novel encoder/decoder framework with theoretical properties distinct from those of the VAE, specifically designed to preserve the geometric structure of the data distribution. We demonstrate the significant advantages of this geometry-preserving encoder in the training process of both the encoder and decoder. Additionally, we provide theoretical results proving convergence of the training process, including convergence guarantees for encoder training, and results showing faster convergence of decoder training when using the geometry-preserving encoder.
Authors: Thomas Mortier, Alireza Javanmardi, Yusuf Sale, Eyke H\"ullermeier, Willem Waegeman
Abstract: Conformal prediction has emerged as a widely used framework for constructing valid prediction sets in classification and regression tasks. In this work, we extend the split conformal prediction framework to hierarchical classification, where prediction sets are commonly restricted to internal nodes of a predefined hierarchy, and propose two computationally efficient inference algorithms. The first algorithm returns internal nodes as prediction sets, while the second one relaxes this restriction. Using the notion of representation complexity, the latter yields smaller set sizes at the cost of a more general and combinatorial inference problem. Empirical evaluations on several benchmark datasets demonstrate the effectiveness of the proposed algorithms in achieving nominal coverage.
Authors: Anna Livia Croella, Veronica Piccialli, Antonio M. Sudoso
Abstract: Clustering is a fundamental technique in data analysis and machine learning, used to group similar data points together. Among various clustering methods, the Minimum Sum-of-Squares Clustering (MSSC) is one of the most widely used. MSSC aims to minimize the total squared Euclidean distance between data points and their corresponding cluster centroids. Due to the unsupervised nature of clustering, achieving global optimality is crucial, yet computationally challenging. The complexity of finding the global solution increases exponentially with the number of data points, making exact methods impractical for large-scale datasets. Even obtaining strong lower bounds on the optimal MSSC objective value is computationally prohibitive, making it difficult to assess the quality of heuristic solutions. We address this challenge by introducing a novel method to validate heuristic MSSC solutions through optimality gaps. Our approach employs a divide-and-conquer strategy, decomposing the problem into smaller instances that can be handled by an exact solver. The decomposition is guided by an auxiliary optimization problem, the "anticlustering problem", for which we design an efficient heuristic. Computational experiments demonstrate the effectiveness of the method for large-scale instances, achieving optimality gaps below 3% in most cases while maintaining reasonable computational times. These results highlight the practicality of our approach in assessing feasible clustering solutions for large datasets, bridging a critical gap in MSSC evaluation.
Authors: David Dobre, Mehrnaz Mofakhami, Sophie Xhonneux, Leo Schwinn, Gauthier Gidel
Abstract: Many safety post-training methods for large language models (LLMs) are designed to modify the model's behaviour from producing unsafe answers to issuing refusals. However, such distribution shifts are often brittle and degrade performance on desirable tasks. To address these pitfalls, we propose augmenting the model's vocabulary with a special red flag token, and training the model to insert this token whenever harmful content is generated or imminent. This approach enables the model to explicitly learn the concept of harmfulness in its representations, with minimal impact on utility due to the marginal change in the generated distribution of natural language. Moreover, because the token is embedded in the model's vocabulary, we can naturally leverage the LLMs' generalization capabilities, such as in-context learning (ICL) and out-of-distribution generalization to languages that are not formally supported (e.g., Japanese for Llama3). In particular, we demonstrate that through ICL alone, the model can learn to initiate reflective reasoning upon generating the red flag token at inference, which steers the response away from harmful continuations or enables self-correction when the flag is raised falsely. This approach is orthogonal and complementary to existing safety technique (such as safety classifiers or standard safety training) and easier to evaluate in comparison to natural language refusals, as it does not require a human or automated judge to assess the harmlessness of the answers.
Authors: Jingru Gan, Peichen Zhong, Yuanqi Du, Yanqiao Zhu, Chenru Duan, Haorui Wang, Daniel Schwalbe-Koda, Carla P. Gomes, Kristin A. Persson, Wei Wang
Abstract: Crystal structure generation is fundamental to materials science, enabling the discovery of novel materials with desired properties. While existing approaches leverage Large Language Models (LLMs) through extensive fine-tuning on materials databases, we show that pre-trained LLMs can inherently generate novel and stable crystal structures without additional fine-tuning. Our framework employs LLMs as intelligent proposal agents within an evolutionary pipeline that guides them to perform implicit crossover and mutation operations while maintaining chemical validity. We demonstrate that MatLLMSearch achieves a 78.38% metastable rate validated by machine learning interatomic potentials and 31.7% DFT-verified stability, outperforming specialized models such as CrystalTextLLM. Beyond crystal structure generation, we further demonstrate that our framework adapts to diverse materials design tasks, including crystal structure prediction and multi-objective optimization of properties such as deformation energy and bulk modulus, all without fine-tuning. These results establish our framework as a versatile and effective framework for consistent high-quality materials discovery, offering training-free generation of novel stable structures with reduced overhead and broader accessibility.
Authors: Shishir Adhikari, Sourav Medya, Elena Zheleva
Abstract: In causal inference, interference refers to the phenomenon in which the actions of peers in a network can influence an individual's outcome. Peer effect refers to the difference in counterfactual outcomes of an individual for different levels of peer exposure, the extent to which an individual is exposed to the treatments, actions, or behaviors of peers. Estimating peer effects requires deciding how to represent peer exposure. Typically, researchers define an exposure mapping function that aggregates peer treatments and outputs peer exposure. Most existing approaches for defining exposure mapping functions assume peer exposure based on the number or fraction of treated peers. Recent studies have investigated more complex functions of peer exposure which capture that different peers can exert different degrees of influence. However, none of these works have explicitly considered the problem of automatically learning the exposure mapping function. In this work, we focus on learning this function for the purpose of estimating heterogeneous peer effects, where heterogeneity refers to the variation in counterfactual outcomes for the same peer exposure but different individual's contexts. We develop EgoNetGNN, a graph neural network (GNN)-based method, to automatically learn the appropriate exposure mapping function allowing for complex peer influence mechanisms that, in addition to peer treatments, can involve the local neighborhood structure and edge attributes. We show that GNN models that use peer exposure based on the number or fraction of treated peers or learn peer exposure naively face difficulty accounting for such influence mechanisms. Our comprehensive evaluation on synthetic and semi-synthetic network data shows that our method is more robust to different unknown underlying influence mechanisms when estimating heterogeneous peer effects when compared to state-of-the-art baselines.
Authors: Yiyang Ling, Karan Owalekar, Oluwatobiloba Adesanya, Erdem B{\i}y{\i}k, Daniel Seita
Abstract: Motion planning involves determining a sequence of robot configurations to reach a desired pose, subject to movement and safety constraints. Traditional motion planning finds collision-free paths, but this is overly restrictive in clutter, where it may not be possible for a robot to accomplish a task without contact. In addition, contacts range from relatively benign (e.g. brushing a soft pillow) to more dangerous (e.g. toppling a glass vase), making it difficult to characterize which may be acceptable. In this paper, we propose IMPACT, a novel motion planning framework that uses Vision-Language Models (VLMs) to infer environment semantics, identifying which parts of the environment can best tolerate contact based on object properties and locations. Our approach generates an anisotropic cost map that encodes directional push safety. We pair this map with a contact-aware A* planner to find stable contact-rich paths. We perform experiments using 20 simulation and 10 real-world scenes and assess using task success rate, object displacements, and feedback from human evaluators. Our results over 3200 simulation and 200 real-world trials suggest that IMPACT enables efficient contact-rich motion planning in cluttered settings while outperforming alternative methods and ablations. Our project website is available at https://impact-planning.github.io/.
Authors: Yang Xiao, Wang Lu, Jie Ji, Ruimeng Ye, Gen Li, Xiaolong Ma, Bo Hui
Abstract: The design of artificial neural networks (ANNs) is inspired by the structure of the human brain, and in turn, ANNs offer a potential means to interpret and understand brain signals. Existing methods primarily align brain signals with stimulus signals using Mean Squared Error (MSE), which focuses only on local point-wise alignment and ignores global matching, leading to coarse interpretations and inaccuracies in brain signal decoding. In this paper, we address these issues through optimal transport (OT) and theoretically demonstrate why OT provides a more effective alignment strategy than MSE. Specifically, we construct a transport plan between brain voxel embeddings and image embeddings, enabling more precise matching. By controlling the amount of transport, we mitigate the influence of redundant information. We apply our alignment model directly to the Brain Captioning task by feeding brain signals into a large language model (LLM) instead of images. Our approach achieves state-of-the-art performance across ten evaluation metrics, surpassing the previous best method by an average of 6.11\% in single-subject training and 3.81\% in cross-subject training. Additionally, we have uncovered several insightful conclusions that align with existing brain research. We unveil the redundancy and synergy of brain information processing through region masking and data dimensionality reduction visualization experiments. We believe our approach paves the way for a more precise understanding of brain signals in the future. The code is available at https://github.com/NKUShaw/OT-Alignment4brain-to-image.
URLs: https://github.com/NKUShaw/OT-Alignment4brain-to-image.
Authors: Alexander Weers, Alexander H. Berger, Laurin Lux, Peter Sch\"uffler, Daniel Rueckert, Johannes C. Paetzold
Abstract: The histopathological analysis of whole-slide images (WSIs) is fundamental to cancer diagnosis but is a time-consuming and expert-driven process. While deep learning methods show promising results, dominant patch-based methods artificially fragment tissue, ignore biological boundaries, and produce black-box predictions. We overcome these limitations with a novel framework that transforms gigapixel WSIs into biologically-informed graph representations and is interpretable by design. Our approach builds graph nodes from tissue regions that respect natural structures, not arbitrary grids. We introduce an adaptive graph coarsening technique, guided by learned embeddings, to efficiently merge homogeneous regions while preserving diagnostically critical details in heterogeneous areas. Each node is enriched with a compact, interpretable feature set capturing clinically-motivated priors. A graph attention network then performs diagnosis on this compact representation. We demonstrate strong performance on challenging cancer staging and survival prediction tasks. Crucially, our resource-efficient model ($>$13x fewer parameters and $>$300x less data) achieves results competitive with a massive foundation model, while offering full interpretability through feature attribution. Our code is publicly available at https://github.com/HistoGraph31/pix2pathology.
Authors: A. Candito (The Institute of Cancer Research, London, United Kingdom), A. Dragan (The Institute of Cancer Research, London, United Kingdom, The Royal Marsden NHS Foundation Trust, London, United Kingdom), R. Holbrey (The Institute of Cancer Research, London, United Kingdom), A. Ribeiro (The Royal Marsden NHS Foundation Trust, London, United Kingdom), R. Donners (University Hospital Basel, Basel, Switzerland), C. Messiou (The Institute of Cancer Research, London, United Kingdom, The Royal Marsden NHS Foundation Trust, London, United Kingdom), N. Tunariu (The Institute of Cancer Research, London, United Kingdom, The Royal Marsden NHS Foundation Trust, London, United Kingdom), D. -M. Koh (The Institute of Cancer Research, London, United Kingdom, The Royal Marsden NHS Foundation Trust, London, United Kingdom), M. D. Blackledge (The Institute of Cancer Research, London, United Kingdom)
Abstract: Background: Apparent Diffusion Coefficient (ADC) values and Total Diffusion Volume (TDV) from Whole-body diffusion-weighted MRI (WB-DWI) are recognized cancer imaging biomarkers. However, manual disease delineation for ADC and TDV measurements is unfeasible in clinical practice, demanding automation. As a first step, we propose an algorithm to generate fast and reproducible probability maps of the skeleton, adjacent internal organs (liver, spleen, urinary bladder, and kidneys), and spinal canal. Methods: We developed an automated deep-learning pipeline based on a 3D patch-based Residual U-Net architecture that localises and delineates these anatomical structures on WB-DWI. The algorithm was trained using "soft labels" (non-binary segmentations) derived from a computationally intensive atlas-based approach. For training and validation, we employed a multi-centre WB-DWI dataset comprising 532 scans from patients with Advanced Prostate Cancer (APC) or Multiple Myeloma (MM), with testing on 45 patients. Results: Our weakly-supervised deep learning model achieved an average dice score of 0.67 for whole skeletal delineation, 0.76 when excluding ribcage, 0.83 for internal organs, and 0.86 for spinal canal, with average surface distances below 3mm. Relative median ADC differences between automated and manual full-body delineations were below 10%. The model was 12x faster than the atlas-based registration algorithm (25 sec vs. 5 min). Two experienced radiologists rated the model's outputs as either "good" or "excellent" on test scans, with inter-reader agreement from fair to substantial (Gwet's AC1 = 0.27-0.72). Conclusion: The model offers fast, reproducible probability maps for localising and delineating body regions on WB-DWI, potentially enabling non-invasive imaging biomarker quantification to support disease staging and treatment response assessment.
Authors: Samuel Weidemaier, Florine Hartwig, Josua Sassen, Sergio Conti, Mirela Ben-Chen, Martin Rumpf
Abstract: We propose a novel variational approach for computing neural Signed Distance Fields (SDF) from unoriented point clouds. To this end, we replace the commonly used eikonal equation with the heat method, carrying over to the neural domain what has long been standard practice for computing distances on discrete surfaces. This yields two convex optimization problems for whose solution we employ neural networks: We first compute a neural approximation of the gradients of the unsigned distance field through a small time step of heat flow with weighted point cloud densities as initial data. Then we use it to compute a neural approximation of the SDF. We prove that the underlying variational problems are well-posed. Through numerical experiments, we demonstrate that our method provides state-of-the-art surface reconstruction and consistent SDF gradients. Furthermore, we show in a proof-of-concept that it is accurate enough for solving a PDE on the zero-level set.
Authors: Bert de Vries, Wouter Nuijten, Thijs van de Laar, Wouter Kouw, Sepideh Adamiat, Tim Nisslbeck, Mykola Lukashchuk, Hoang Minh Huu Nguyen, Marco Hidalgo Araya, Raphael Tresor, Thijs Jenneskens, Ivana Nikoloska, Raaja Ganapathy Subramanian, Bart van Erp, Dmitry Bagaev, Albert Podusenko
Abstract: We address the problem of planning under uncertainty, where an agent must choose actions that not only achieve desired outcomes but also reduce uncertainty. Traditional methods often treat exploration and exploitation as separate objectives, lacking a unified inferential foundation. Active inference, grounded in the Free Energy Principle, provides such a foundation by minimizing Expected Free Energy (EFE), a cost function that combines utility with epistemic drives, such as ambiguity resolution and novelty seeking. However, the computational burden of EFE minimization had remained a significant obstacle to its scalability. In this paper, we show that EFE-based planning arises naturally from minimizing a variational free energy functional on a generative model augmented with preference and epistemic priors. This result reinforces theoretical consistency with the Free Energy Principle by casting planning under uncertainty itself as a form of variational inference. Our formulation yields policies that jointly support goal achievement and information gain, while incorporating a complexity term that accounts for bounded computational resources. This unifying framework connects and extends existing methods, enabling scalable, resource-aware implementations of active inference agents.
Authors: Kirtan Rajesh, Suvidha Rupesh Kumar
Abstract: This is the preprint version of the article published in IEEE Access vol. 13, pp. 146503--146526, 2025, doi:10.1109/ACCESS.2025.3599541. Please cite the published version. Urban air pollution remains a pressing global concern, particularly in densely populated and traffic-intensive metropolitan areas like Delhi, where exposure to harmful pollutants severely impacts public health. Delhi, being one of the most polluted cities globally, experiences chronic air quality issues due to vehicular emissions, industrial activities, and construction dust, which exacerbate its already fragile atmospheric conditions. Traditional pollution mitigation strategies, such as static air purifying installations, often fail to maximize their impact due to suboptimal placement and limited adaptability to dynamic urban environments. This study presents a novel deep reinforcement learning (DRL) framework to optimize the placement of air purification booths to improve the air quality index (AQI) in the city of Delhi. We employ Proximal Policy Optimization (PPO), a state-of-the-art reinforcement learning algorithm, to iteratively learn and identify high-impact locations based on multiple spatial and environmental factors, including population density, traffic patterns, industrial influence, and green space constraints. Our approach is benchmarked against conventional placement strategies, including random and greedy AQI-based methods, using multi-dimensional performance evaluation metrics such as AQI improvement, spatial coverage, population and traffic impact, and spatial entropy.
Authors: Nikita P. Kalinin, Ryan McKenna, Jalaj Upadhyay, Christoph H. Lampert
Abstract: Matrix factorization mechanisms for differentially private training have emerged as a promising approach to improve model utility under privacy constraints. In practical settings, models are typically trained over multiple epochs, requiring matrix factorizations that account for repeated participation. Existing theoretical upper and lower bounds on multi-epoch factorization error leave a significant gap. In this work, we introduce a new explicit factorization method, Banded Inverse Square Root (BISR), which imposes a banded structure on the inverse correlation matrix. This factorization enables us to derive an explicit and tight characterization of the multi-epoch error. We further prove that BISR achieves asymptotically optimal error by matching the upper and lower bounds. Empirically, BISR performs on par with state-of-the-art factorization methods, while being simpler to implement, computationally efficient, and easier to analyze.
Authors: Jiahao Yu, Haozhuang Liu, Yeqiu Yang, Lu Chen, Jian Wu, Yuning Jiang, Bo Zheng
Abstract: Regression models are crucial in recommender systems. However, retransformation bias problem has been conspicuously neglected within the community. While many works in other fields have devised effective bias correction methods, all of them are post-hoc cures externally to the model, facing practical challenges when applied to real-world recommender systems. Hence, we propose a preemptive paradigm to eradicate the bias intrinsically from the models via minor model refinement. Specifically, a novel TranSUN method is proposed with a joint bias learning manner to offer theoretically guaranteed unbiasedness under empirical superior convergence. It is further generalized into a novel generic regression model family, termed Generalized TranSUN (GTS), which not only offers more theoretical insights but also serves as a generic framework for flexibly developing various bias-free models. Comprehensive experimental results demonstrate the superiority of our methods across data from various domains, which have been successfully deployed in two real-world industrial recommendation scenarios, i.e. product and short video recommendation scenarios in Guess What You Like business domain in the homepage of Taobao App (a leading e-commerce platform with DAU > 300M), to serve the major online traffic.
Authors: Huopu Zhang, Yanguang Liu, Miao Zhang, Zirui He, Mengnan Du
Abstract: Predicting earnings surprises from financial documents, such as earnings conference calls, regulatory filings, and financial news, has become increasingly important in financial economics. However, these financial documents present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the SAE-FiRE (Sparse Autoencoder for Financial Representation Enhancement) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to decompose dense neural representations from large language models into interpretable sparse components, then applies statistical feature selection methods, including ANOVA F-tests and tree-based importance scoring, to identify the top-k most discriminative dimensions for classification. By systematically filtering out noise that might otherwise lead to overfitting, we enable more robust and generalizable predictions. Experimental results across three financial datasets demonstrate that SAE-FiRE significantly outperforms baseline approaches.
Authors: Harethah Abu Shairah, Hasan Abed Al Kader Hammoud, Bernard Ghanem, George Turkiyyah
Abstract: Large language models (LLMs) are typically aligned to refuse harmful instructions through safety fine-tuning. A recent attack, termed abliteration, identifies and suppresses the single latent direction most responsible for refusal behavior, thereby enabling models to generate harmful content. We propose a defense that fundamentally alters how models express refusal. We construct an extended-refusal dataset in which responses to harmful prompts provide detailed justifications before refusing, distributing the refusal signal across multiple token positions. Fine-tuning Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on this dataset yields models that maintain high refusal rates under abliteration: refusal rates drop by at most 10%, compared to 70-80% drops in baseline models. Comprehensive evaluations of safety and utility demonstrate that extended-refusal fine-tuning effectively neutralizes abliteration attacks while preserving general model performance and enhancing robustness across multiple alignment scenarios.
Authors: Liangliang Zhang, Zhuorui Jiang, Hongliang Chi, Haoyang Chen, Mohammed Elkoumy, Fali Wang, Qiong Wu, Zhengyi Zhou, Shirui Pan, Suhang Wang, Yao Ma
Abstract: Knowledge Graph Question Answering (KGQA) systems rely on high-quality benchmarks to evaluate complex multi-hop reasoning. However, despite their widespread use, popular datasets such as WebQSP and CWQ suffer from critical quality issues, including inaccurate or incomplete ground-truth annotations, poorly constructed questions that are ambiguous, trivial, or unanswerable, and outdated or inconsistent knowledge. Through a manual audit of 16 popular KGQA datasets, including WebQSP and CWQ, we find that the average factual correctness rate is only 57 %. To address these issues, we introduce KGQAGen, an LLM-in-the-loop framework that systematically resolves these pitfalls. KGQAGen combines structured knowledge grounding, LLM-guided generation, and symbolic verification to produce challenging and verifiable QA instances. Using KGQAGen, we construct KGQAGen-10k, a ten-thousand scale benchmark grounded in Wikidata, and evaluate a diverse set of KG-RAG models. Experimental results demonstrate that even state-of-the-art systems struggle on this benchmark, highlighting its ability to expose limitations of existing models. Our findings advocate for more rigorous benchmark construction and position KGQAGen as a scalable framework for advancing KGQA evaluation.
Authors: Daniel Thilo Schroeder, Meeyoung Cha, Andrea Baronchelli, Nick Bostrom, Nicholas A. Christakis, David Garcia, Amit Goldenberg, Yara Kyrychenko, Kevin Leyton-Brown, Nina Lutz, Gary Marcus, Filippo Menczer, Gordon Pennycook, David G. Rand, Maria Ressa, Frank Schweitzer, Christopher Summerfield, Audrey Tang, Jay J. Van Bavel, Sander van der Linden, Dawn Song, Jonas R. Kunst
Abstract: Public opinion manipulation has entered a new phase, amplifying its roots in rhetoric and propaganda. Advances in large language models (LLMs) and autonomous agents now let influence campaigns reach unprecedented scale and precision. Researchers warn AI could foster mass manipulation. Generative tools can expand propaganda output without sacrificing credibility and inexpensively create election falsehoods that are rated as more human-like than those written by humans. Techniques meant to refine AI reasoning, such as chain-of-thought prompting, can just as effectively be used to generate more convincing falsehoods. Enabled by these capabilities, another disruptive threat is emerging: swarms of collaborative, malicious AI agents. Fusing LLM reasoning with multi-agent architectures, these systems are capable of coordinating autonomously, infiltrating communities, and fabricating consensus cheaply. By adaptively mimicking human social dynamics, they threaten democracy.
Authors: Alex Pierron, Michel Barbeau, Luca De Cicco, Jose Rubio-Hernan, Joaquin Garcia-Alfaro
Abstract: Reconfigurable Intelligent Surfaces are composed of physical elements that can dynamically alter electromagnetic wave properties to enhance beamforming and lead to improvements in areas with low coverage properties. Combined with Reinforcement Learning techniques, they have the potential to be conduct as well physical-layer security hardening. Yet, and in addition to security improvements, it is crucial to consider the concept of fair communication. Reconfigurable Intelligent Surfaces must ensure that User Equipment units receive their signals with adequate strength, without other units being deprived of service due to insufficient power. In this paper, we address such a problem. We explore the fairness properties of previous work and propose a novel method that aims at obtaining both an efficient and fair duplex Reconfigurable Intelligent Surface-Reinforcement Learning system for multiple legitimate User Equipment units without reducing the level of achieved physical-layer security hardening. In terms of contributions, we uncover a fairness imbalance of a previous physical-layer security hardening solution, validate our findings and report experimental work via simulation results. We also provide an alternative reward strategy to solve the uncovered problems and release both code and datasets to foster further research in the topics of this paper.
Authors: Mingzhou Liu, Xinwei Sun, Yizhou Wang
Abstract: Inferring causal relationships from dynamical systems is the central interest of many scientific inquiries. Conditional local independence, which describes whether the evolution of one process is influenced by another process given additional processes, is important for causal learning in such systems. In this paper, we propose a hypothesis test for conditional local independence in It\^o processes. Our test is grounded in the semimartingale decomposition of the It\^o process, with which we introduce a stochastic integral process that is a martingale under the null hypothesis. We then apply a test for the martingale property, quantifying potential deviation from local independence. The test statistics is estimated using the optimal filtering equation. We show the consistency of the estimation, thereby establishing the level and power of our test. Numerical verification and a real-world application to causal discovery in brain resting-state fMRIs are conducted.
Authors: Zachary Ravichandran, Ignacio Hounie, Fernando Cladera, Alejandro Ribeiro, George J. Pappas, Vijay Kumar
Abstract: Large language models (LLMs) provide robots with powerful contextual reasoning abilities and a natural human interface. Yet, current LLM-enabled robots typically depend on cloud-hosted models, limiting their usability in environments with unreliable communication infrastructure, such as outdoor or industrial settings. We present PRISM, a framework for distilling small language model (SLM)-enabled robot planners that run on-device with minimal human supervision. Starting from an existing LLM-enabled planner, PRISM automatically synthesizes diverse tasks and environments, elicits plans from the LLM, and uses this synthetic dataset to distill a compact SLM as a drop-in replacement of the source model. We apply PRISM to three LLM-enabled planners for mapping and exploration, manipulation, and household assistance, and we demonstrate that PRISM improves the performance of Llama-3.2-3B from 10-20% of GPT-4o's performance to over 93% - using only synthetic data. We further demonstrate that the distilled planners generalize across heterogeneous robotic platforms (ground and aerial) and diverse environments (indoor and outdoor). We release all software, trained models, and datasets at https://zacravichandran.github.io/PRISM.
Authors: Christian Huber, Alexander Waibel
Abstract: Neural sequence-to-sequence systems deliver state-of-the-art performance for automatic speech recognition. When using appropriate modeling units, e.g., byte-pair encoded characters, these systems are in principal open vocabulary systems. In practice, however, they often fail to recognize words not seen during training, e.g., named entities, acronyms, or domain-specific special words. To address this problem, many context biasing methods have been proposed; however, for words with a pronunciation-orthography mismatch, these methods may still struggle. We propose a method which allows corrections of substitution errors to improve the recognition accuracy of such challenging words. Users can add corrections on the fly during inference. We show that with this method we get a relative improvement in biased word error rate of up to 8%, while maintaining a competitive overall word error rate.
Authors: Akio Hayakawa, Masato Ishii, Takashi Shibuya, Yuki Mitsufuji
Abstract: We propose a step-by-step video-to-audio (V2A) generation method for finer controllability over the generation process and more realistic audio synthesis. Inspired by traditional Foley workflows, our approach aims to comprehensively capture all sound events induced by a video through the incremental generation of missing sound events. To avoid the need for costly multi-reference video-audio datasets, each generation step is formulated as a negatively guided V2A process that discourages duplication of existing sounds. The guidance model is trained by finetuning a pre-trained V2A model on audio pairs from adjacent segments of the same video, allowing training with standard single-reference audiovisual datasets that are easily accessible. Objective and subjective evaluations demonstrate that our method enhances the separability of generated sounds at each step and improves the overall quality of the final composite audio, outperforming existing baselines.
Authors: Yik Siu Chan, Zheng-Xin Yong, Stephen H. Bach
Abstract: Reasoning language models improve performance on complex tasks by generating long chains of thought (CoTs), but this process can also increase harmful outputs in adversarial settings. In this work, we ask whether the long CoTs can be leveraged for predictive safety monitoring: do the reasoning traces provide early signals of final response alignment that could enable timely intervention? We evaluate a range of monitoring methods using either CoT text or activations, including highly capable large language models, fine-tuned classifiers, and humans. First, we find that a simple linear probe trained on CoT activations significantly outperforms all text-based baselines in predicting whether a final response is safe or unsafe, with an average absolute increase of 13 in F1 scores over the best-performing alternatives. CoT texts are often unfaithful and misleading, while model latents provide a more reliable predictive signal. Second, the probe can be applied to early CoT segments before the response is generated, showing that alignment signals appear before reasoning completes. Error analysis reveals that the performance gap between text classifiers and the linear probe largely stems from a subset of responses we call performative CoTs, where the reasoning consistently contradicts the final response as the CoT progresses. Our findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.
Authors: Zacharias Komodromos, Kleanthis Malialis, Panayiotis Kolios
Abstract: Emerging in December 2019, the COVID-19 pandemic caused widespread health, economic, and social disruptions. Rapid global transmission overwhelmed healthcare systems, resulting in high infection rates, hospitalisations, and fatalities. To minimise the spread, governments implemented several non-pharmaceutical interventions like lockdowns and travel restrictions. While effective in controlling transmission, these measures also posed significant economic and societal challenges. Although the WHO declared COVID-19 no longer a global health emergency in May 2023, its impact persists, shaping public health strategies. The vast amount of data collected during the pandemic offers valuable insights into disease dynamics, transmission, and intervention effectiveness. Leveraging these insights can improve forecasting models, enhancing preparedness and response to future outbreaks while mitigating their social and economic impact. This paper presents a large-scale case study on COVID-19 forecasting in Cyprus, utilising a two-year dataset that integrates epidemiological data, vaccination records, policy measures, and weather conditions. We analyse infection trends, assess forecasting performance, and examine the influence of external factors on disease dynamics. The insights gained contribute to improved pandemic preparedness and response strategies.
Authors: Atif Bilal Asad, Achyut Paudel, Safal Kshetri, Chenchen Kang, Salik Ram Khanal, Nataliya Shcherbatyuk, Pierre Davadant, R. Paul Schreiner, Santosh Kalauni, Manoj Karkee, Markus Keller
Abstract: Nitrogen (N) is one of the most crucial nutrients in vineyards, affecting plant growth and subsequent products such as wine and juice. Because soil N has high spatial and temporal variability, it is desirable to accurately estimate the N concentration of grapevine leaves and manage fertilization at the individual plant level to optimally meet plant needs. In this study, we used in-field hyperspectral images with wavelengths ranging from $400 to 1000nm of four different grapevine cultivars collected from distinct vineyards and over two growth stages during two growing seasons to develop models for predicting N concentration at the leaf-level and canopy-level. After image processing, two feature selection methods were employed to identify the optimal set of spectral bands that were responsive to leaf N concentrations. The selected spectral bands were used to train and test two different Machine Learning (ML) models, Gradient Boosting and XGBoost, for predicting nitrogen concentrations. The comparison of selected bands for both leaf-level and canopy-level datasets showed that most of the spectral regions identified by the feature selection methods were across both methods and the dataset types (leaf- and canopy-level datasets), particularly in the key regions, 500-525nm, 650-690nm, 750-800nm, and 900-950nm. These findings indicated the robustness of these spectral regions for predicting nitrogen content. The results for N prediction demonstrated that the ML model achieved an R square of 0.49 for canopy-level data and an R square of 0.57 for leaf-level data, despite using different sets of selected spectral bands for each analysis level. The study demonstrated the potential of using in-field hyperspectral imaging and the use of spectral data in integrated feature selection and ML techniques to monitor N status in vineyards.
Authors: Vishakh Padmakumar, Joseph Chee Chang, Kyle Lo, Doug Downey, Aakanksha Naik
Abstract: The increasing volume of academic literature makes it essential for researchers to organize, compare, and contrast collections of documents. Large language models (LLMs) can support this process by generating schemas defining shared aspects along which to compare papers. However, progress on schema generation has been slow due to: (i) ambiguity in reference-based evaluations, and (ii) lack of editing/refinement methods. Our work is the first to address both issues. First, we present an approach for augmenting unannotated table corpora with \emph{synthesized intents}, and apply it to create a dataset for studying schema generation conditioned on a given information need, thus reducing ambiguity. With this dataset, we show how incorporating table intents significantly improves baseline performance in reconstructing reference schemas. We start by comprehensively benchmarking several single-shot schema generation methods, including prompted LLM workflows and fine-tuned models, showing that smaller, open-weight models can be fine-tuned to be competitive with state-of-the-art prompted LLMs. Next, we propose several LLM-based schema refinement techniques and show that these can further improve schemas generated by these methods.
Authors: Sven Lankester, Gustavo de Carvalho Bertoli, Matias Vizcaino, Emmanuelle Beauxis Aussalet, Manel Slokom
Abstract: The drive for personalization in recommender systems creates a tension between user privacy and the risk of "filter bubbles". Although federated learning offers a promising paradigm for privacy-preserving recommendations, its impact on diversity remains unclear. We introduce FedFlex, a two-stage framework that combines local, on-device fine-tuning of matrix factorization models (SVD and BPR) with a lightweight Maximal Marginal Relevance (MMR) re-ranking step to promote diversity. We conducted the first live user study of a federated recommender, collecting behavioral data and feedback during a two-week online deployment. Our results show that FedFlex successfully engages users, with BPR outperforming SVD in click-through rate. Re-ranking with MMR consistently improved ranking quality (nDCG) across both models, with statistically significant gains, particularly for BPR. Diversity effects varied: MMR increased coverage for both models and improved intra-list diversity for BPR, but slightly reduced it for SVD, suggesting different interactions between personalization and diversification across models. Our exit questionnaire responses indicated that most users expressed no clear preference between re-ranked and unprocessed lists, implying that increased diversity did not substantially reduce user satisfaction.
Authors: Yuzhuang Xu, Xu Han, Yuanchi Zhang, Yixuan Wang, Yijun Liu, Shiyu Ji, Qingfu Zhu, Wanxiang Che
Abstract: Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.
Authors: Xiaoyuan Zhu, Muru Zhang, Ollie Liu, Robin Jia, Willie Neiswanger
Abstract: Modern large language models often encode sensitive, harmful, or copyrighted knowledge, raising the need for post-hoc unlearning-the ability to remove specific domains of knowledge from a model without full retraining. A major bottleneck in current unlearning pipelines is constructing effective forget sets-datasets that approximate the target domain and guide the model to forget it. In this work, we introduce a scalable, automated approach to generate high-quality forget sets using language models themselves. Our method synthesizes textbook-style data through a structured prompting pipeline, requiring only a domain name as input. Through experiments on unlearning biosecurity, cybersecurity, and Harry Potter novels, we show that our synthetic datasets consistently outperform the baseline synthetic alternatives and are comparable to the expert-curated ones. Additionally, ablation studies reveal that the multi-step generation pipeline significantly boosts data diversity, which in turn improves unlearning utility. Overall, our findings suggest that synthetic datasets offer a promising path toward practical, scalable unlearning for a wide range of emerging domains without the need for manual intervention. We release our code and dataset at https://github.com/xyzhu123/Synthetic_Textbook.
Authors: Michele Minervini, Madison Chin, Jacob Kupperman, Nana Liu, Ivy Luo, Meghan Ly, Soorya Rethinasamy, Kathie Wang, Mark M. Wilde
Abstract: A quantum thermodynamic system is described by a Hamiltonian and a list of conserved, non-commuting charges, and a fundamental goal is to determine the minimum energy of the system subject to constraints on the charges. Recently, [Liu et al., arXiv:2505.04514] proposed first- and second-order classical and hybrid quantum-classical algorithms for solving a dual chemical potential maximization problem, and they proved that these algorithms converge to global optima by means of gradient-ascent approaches. In this paper, we benchmark these algorithms on several problems of interest in thermodynamics, including one- and two-dimensional quantum Heisenberg models with nearest and next-to-nearest neighbor interactions and with the charges set to the total x, y, and z magnetizations. We also offer an alternative compelling interpretation of these algorithms as methods for designing ground and thermal states of controllable Hamiltonians, with potential applications in molecular and material design. Furthermore, we introduce stabilizer thermodynamic systems as thermodynamic systems based on stabilizer codes, with the Hamiltonian constructed from a given code's stabilizer operators and the charges constructed from the code's logical operators. We benchmark the aforementioned algorithms on several examples of stabilizer thermodynamic systems, including those constructed from the one-to-three-qubit repetition code, the perfect one-to-five-qubit code, and the two-to-four-qubit error-detecting code. Finally, we observe that the aforementioned hybrid quantum-classical algorithms, when applied to stabilizer thermodynamic systems, can serve as alternative methods for encoding qubits into stabilizer codes at a fixed temperature, and we provide an effective method for warm-starting these encoding algorithms whenever a single qubit is encoded into multiple physical qubits.
Authors: Michael Poppel, David Bucher, Maximilian Zorn, Nico Kraus, Philipp Altmann, Jonas Stein, Claudia Linnhoff-Popien
Abstract: Quantum machine learning research has expanded rapidly due to potential computational advantages over classical methods. Angle encoding has emerged as a popular choice as feature map (FM) for embedding classical data into quantum models due to its simplicity and natural generation of truncated Fourier series, providing universal function approximation capabilities. Efficient FMs within quantum circuits can exploit exponential scaling of Fourier frequencies, with multi-dimensional inputs introducing additional exponential growth through mixed-frequency terms. Despite this promising expressive capability, practical implementation faces significant challenges. Through controlled experiments with white-box target functions, we demonstrate that training failures can occur even when all relevant frequencies are theoretically accessible. We illustrate how two primary known causes lead to unsuccessful optimization: insufficient trainable parameters relative to the model's frequency content, and limitations imposed by the ansatz's dynamic lie algebra dimension, but also uncover an additional parameter burden: the necessity of controlling non-unique frequencies within the model. To address this, we propose near-zero weight initialization to suppress unnecessary duplicate frequencies. For target functions with a priori frequency knowledge, we introduce frequency selection as a practical solution that reduces parameter requirements and mitigates the exponential growth that would otherwise render problems intractable due to parameter insufficiency. Our frequency selection approach achieved near-optimal performance (median $R^2 \approx 0.95$) with 78\% of the parameters needed by the best standard approach in 10 randomly chosen target functions.
Authors: Abhigya Verma, Sriram Puttagunta, Seganrasan Subramanian, Sravan Ramachandran
Abstract: GRAFT is a structured multimodal benchmark for evaluating models on instruction-following, visual reasoning, and visual-textual alignment tasks. It features programmatically generated charts and synthetically rendered tables, created with Python visualization libraries to ensure control over data semantics, structure, and clarity. Each GRAFT instance pairs a chart or table image with a systematically generated, multi-step analytical question based solely on visual content. Answers are provided in structured formats such as JSON or YAML, supporting consistent evaluation of both reasoning and output format. The benchmark introduces a taxonomy of reasoning types including comparison, trend identification, ranking, aggregation, proportion estimation, and anomaly detection to enable comprehensive assessment. Reference answers follow strict factual and formatting guidelines for precise, aspect-based evaluation. GRAFT offers a unified, scalable framework for fine-grained benchmarking of multimodal models on visually grounded, structured reasoning tasks, setting a new evaluation standard in this field.
Authors: Hashim Ali, Surya Subramani, Lekha Bollinani, Nithin Sai Adupa, Sali El-Loh, Hafiz Malik
Abstract: The SAFE Challenge evaluates synthetic speech detection across three tasks: unmodified audio, processed audio with compression artifacts, and laundered audio designed to evade detection. We systematically explore self-supervised learning (SSL) front-ends, training data compositions, and audio length configurations for robust deepfake detection. Our AASIST-based approach incorporates WavLM large frontend with RawBoost augmentation, trained on a multilingual dataset of 256,600 samples spanning 9 languages and over 70 TTS systems from CodecFake, MLAAD v5, SpoofCeleb, Famous Figures, and MAILABS. Through extensive experimentation with different SSL front-ends, three training data versions, and two audio lengths, we achieved second place in both Task 1 (unmodified audio detection) and Task 3 (laundered audio detection), demonstrating strong generalization and robustness.
Authors: Soheil Zibakhsh Shabgahi, Pedram Aghazadeh, Azalia Mirhoseini, Farinaz Koushanfar
Abstract: The increasing reliance on generative AI models is rapidly increasing the volume of synthetic data, with some projections suggesting that most available new data for training could be machine-generated by 2030. This shift to a mainly synthetic content presents a critical challenge: repeated training in synthetic data leads to a phenomenon known as model collapse, where model performance degrades over generations of training, eventually rendering the models ineffective. While the causes of model collapse are increasingly understood, effective mitigation strategies remain scarce. We address this challenge by leveraging a key insight: auto-regressive models tend to generate text sequences to which they assign high confidence (i.e., high log-likelihood). Based on this observation, we introduce the Truncated-Cross-Entropy (TCE) loss function. TCE mitigates collapse by selectively ignoring high-confidence tokens during training, effectively filtering out likely machine-generated artifacts from the learning process. Our experiments demonstrate that models trained with TCE not only learn effectively but also exhibit significantly increased resilience, tolerating over 2.3x more synthetic data before the onset of collapse. In addition, we provide an open-source benchmark for collapse dynamics in mixed-data settings. Our results demonstrate that confidence-aware training objectives can substantially delay collapse onset, offering a practical and generalizable tool for model robustness under synthetic-data exposure.
Authors: Victor Livernoche, Akshatha Arodi, Andreea Musulan, Zachary Yang, Adam Salvail, Ga\'etan Marceau Caron, Jean-Fran\c{c}ois Godbout, Reihaneh Rabbany
Abstract: Deepfakes, synthetic media created using advanced AI techniques, pose a growing threat to information integrity, particularly in politically sensitive contexts. This challenge is amplified by the increasing realism of modern generative models, which our human perception study confirms are often indistinguishable from real images. Yet, existing deepfake detection benchmarks rely on outdated generators or narrowly scoped datasets (e.g., single-face imagery), limiting their utility for real-world detection. To address these gaps, we present OpenFake, a large politically grounded dataset specifically crafted for benchmarking against modern generative models with high realism, and designed to remain extensible through an innovative crowdsourced adversarial platform that continually integrates new hard examples. OpenFake comprises nearly four million total images: three million real images paired with descriptive captions and almost one million synthetic counterparts from state-of-the-art proprietary and open-source models. Detectors trained on OpenFake achieve near-perfect in-distribution performance, strong generalization to unseen generators, and high accuracy on a curated in-the-wild social media test set, significantly outperforming models trained on existing datasets. Overall, we demonstrate that with high-quality and continually updated benchmarks, automatic deepfake detection is both feasible and effective in real-world settings.
Authors: Ya-Chi Chu, Wenzhi Gao, Yinyu Ye, Madeleine Udell
Abstract: Part I of this work [Gao25] establishes online scaled gradient methods (OSGM), a framework that utilizes online convex optimization to adapt stepsizes in gradient methods. This paper focuses on the practical aspects of OSGM. We leverage the OSGM framework to design new adaptive first-order methods and provide insights into their empirical behavior. The resulting method, OSGM-Best, matches the performance of quasi-Newton variants while requiring less memory and cheaper iterations. We also extend OSGM to nonconvex optimization and outline directions that connect OSGM to existing branches of optimization theory and practice.
Authors: Harshit Rajgarhia, Shivali Dalmia, Mengyang Zhao, Mukherji Abhishek, Kiran Ganesh
Abstract: Adapting advertisements for multilingual audiences requires more than simple text translation; it demands preservation of visual consistency, spatial alignment, and stylistic integrity across diverse languages and formats. We introduce a structured framework that combines automated components with human oversight to address the complexities of advertisement localization. To the best of our knowledge, this is the first work to integrate scene text detection, inpainting, machine translation (MT), and text reimposition specifically for accelerating ad localization evaluation workflows. Qualitative results across six locales demonstrate that our approach produces semantically accurate and visually coherent localized advertisements, suitable for deployment in real-world workflows.
Authors: Momchil S. Tomov, Sang Uk Lee, Hansford Hendrago, Jinwook Huh, Teawon Han, Forbes Howington, Rafael da Silva, Gianmarco Bernasconi, Marc Heim, Samuel Findler, Xiaonan Ji, Alexander Boule, Michael Napoli, Kuo Chen, Jesse Miller, Boaz Floor, Yunqing Hu
Abstract: We present TreeIRL, a novel planner for autonomous driving that combines Monte Carlo tree search (MCTS) and inverse reinforcement learning (IRL) to achieve state-of-the-art performance in simulation and in real-world driving. The core idea is to use MCTS to find a promising set of safe candidate trajectories and a deep IRL scoring function to select the most human-like among them. We evaluate TreeIRL against both classical and state-of-the-art planners in large-scale simulations and on 500+ miles of real-world autonomous driving in the Las Vegas metropolitan area. Test scenarios include dense urban traffic, adaptive cruise control, cut-ins, and traffic lights. TreeIRL achieves the best overall performance, striking a balance between safety, progress, comfort, and human-likeness. To our knowledge, our work is the first demonstration of MCTS-based planning on public roads and underscores the importance of evaluating planners across a diverse set of metrics and in real-world environments. TreeIRL is highly extensible and could be further improved with reinforcement learning and imitation learning, providing a framework for exploring different combinations of classical and learning-based approaches to solve the planning bottleneck in autonomous driving.
Authors: Tom Burgert, Oliver Stoll, Paolo Rota, Beg\"um Demir
Abstract: The hypothesis that Convolutional Neural Networks (CNNs) are inherently texture-biased has shaped much of the discourse on feature use in deep learning. We revisit this hypothesis by examining limitations in the cue-conflict experiment by Geirhos et al. To address these limitations, we propose a domain-agnostic framework that quantifies feature reliance through systematic suppression of shape, texture, and color cues, avoiding the confounds of forced-choice conflicts. By evaluating humans and neural networks under controlled suppression conditions, we find that CNNs are not inherently texture-biased but predominantly rely on local shape features. Nonetheless, this reliance can be substantially mitigated through modern training strategies or architectures (ConvNeXt, ViTs). We further extend the analysis across computer vision, medical imaging, and remote sensing, revealing that reliance patterns differ systematically: computer vision models prioritize shape, medical imaging models emphasize color, and remote sensing models exhibit a stronger reliance on texture. Code is available at https://github.com/tomburgert/feature-reliance.
Authors: Aleksandar Terzi\'c, Nicolas Menet, Michael Hersche, Thomas Hofmann, Abbas Rahimi
Abstract: Modern state-space models (SSMs) often utilize transition matrices which enable efficient computation but pose restrictions on the model's expressivity, as measured in terms of the ability to emulate finite-state automata (FSA). While unstructured transition matrices are optimal in terms of expressivity, they come at a prohibitively high compute and memory cost even for moderate state sizes. We propose a structured sparse parametrization of transition matrices in SSMs that enables FSA state tracking with optimal state size and depth, while keeping the computational cost of the recurrence comparable to that of diagonal SSMs. Our method, PD-SSM, parametrizes the transition matrix as the product of a column one-hot matrix ($P$) and a complex-valued diagonal matrix ($D$). Consequently, the computational cost of parallel scans scales linearly with the state size. Theoretically, the model is BIBO-stable and can emulate any $N$-state FSA with one layer of dimension $N$ and a linear readout of size $N \times N$, significantly improving on all current structured SSM guarantees. Experimentally, the model significantly outperforms a wide collection of modern SSM variants on various FSA state tracking tasks. On multiclass time-series classification, the performance is comparable to that of neural controlled differential equations, a paradigm explicitly built for time-series analysis. Finally, we integrate PD-SSM into a hybrid Transformer-SSM architecture and demonstrate that the model can effectively track the states of a complex FSA in which transitions are encoded as a set of variable-length English sentences. The code is available at https://github.com/IBM/expressive-sparse-state-space-model
URLs: https://github.com/IBM/expressive-sparse-state-space-model
Authors: Yikai Wang, Xiaocheng Li, Guanting Chen
Abstract: Large language models (LLMs) are increasingly used for decision-making tasks under uncertainty; however, their risk profiles and how they are influenced by prompting and alignment methods remain underexplored. Existing studies have primarily examined personality prompting or multi-agent interactions, leaving open the question of how post-training influences the risk behavior of LLMs. In this work, we propose a new pipeline for eliciting, steering, and modulating LLMs' risk profiles, drawing on tools from behavioral economics and finance. Using utility-theoretic models, we compare pre-trained, instruction-tuned, and RLHF-aligned LLMs, and find that while instruction-tuned models exhibit behaviors consistent with some standard utility formulations, pre-trained and RLHF-aligned models deviate more from any utility models fitted. We further evaluate modulation strategies, including prompt engineering, in-context learning, and post-training, and show that post-training provides the most stable and effective modulation of risk preference. Our findings provide insights into the risk profiles of different classes and stages of LLMs and demonstrate how post-training modulates these profiles, laying the groundwork for future research on behavioral alignment and risk-aware LLM design.
Authors: Fahad Mostafa, Kannon Hossain, Hafiz Khan
Abstract: Early and accurate diagnosis of Alzheimer Disease is critical for effective clinical intervention, particularly in distinguishing it from Mild Cognitive Impairment, a prodromal stage marked by subtle structural changes. In this study, we propose a hybrid deep learning ensemble framework for Alzheimer Disease classification using structural magnetic resonance imaging. Gray and white matter slices are used as inputs to three pretrained convolutional neural networks such as ResNet50, NASNet, and MobileNet, each fine tuned through an end to end process. To further enhance performance, we incorporate a stacked ensemble learning strategy with a meta learner and weighted averaging to optimally combine the base models. Evaluated on the Alzheimer Disease Neuroimaging Initiative dataset, the proposed method achieves state of the art accuracy of 99.21% for Alzheimer Disease vs. Mild Cognitive Impairment and 91.0% for Mild Cognitive Impairment vs. Normal Controls, outperforming conventional transfer learning and baseline ensemble methods. To improve interpretability in image based diagnostics, we integrate Explainable AI techniques by Gradient weighted Class Activation, which generates heatmaps and attribution maps that highlight critical regions in gray and white matter slices, revealing structural biomarkers that influence model decisions. These results highlight the frameworks potential for robust and scalable clinical decision support in neurodegenerative disease diagnostics.
Authors: Renee Ge, Qianli Liao, Tomaso Poggio
Abstract: Transformers have demonstrated remarkable performance in natural language processing and related domains, as they largely focus on sequential, autoregressive next-token prediction tasks. Yet, they struggle in logical reasoning, not necessarily because of a fundamental limitation of these models, but possibly due to the lack of exploration of more creative uses, such as latent space and recurrent reasoning. An emerging exploration in this direction is the Hierarchical Reasoning Model (Wang et. al., 2025), which introduces a novel type of recurrent reasoning in the latent space of transformers, achieving remarkable performance on a wide range of 2D reasoning tasks. Despite the promising results, this line of models is still at an early stage and calls for in-depth investigation. In this work, we review this class of models, examine key design choices, test alternative variants and clarify common misconceptions.
Authors: Eric Gelphman, Deepanshu Verma, Nicole Tianjiao Yang, Stanley Osher, Samy Wu Fung
Abstract: Neural network approaches that parameterize value functions have succeeded in approximating high-dimensional optimal feedback controllers when the Hamiltonian admits explicit formulas. However, many practical problems, such as the space shuttle reentry problem and bicycle dynamics, among others, may involve implicit Hamiltonians that do not admit explicit formulas, limiting the applicability of existing methods. Rather than directly parameterizing controls, which does not leverage the Hamiltonian's underlying structure, we propose an end-to-end implicit deep learning approach that directly parameterizes the value function to learn optimal control laws. Our method enforces physical principles by ensuring trained networks adhere to the control laws by exploiting the fundamental relationship between the optimal control and the value function's gradient; this is a direct consequence of the connection between Pontryagin's Maximum Principle and dynamic programming. Using Jacobian-Free Backpropagation (JFB), we achieve efficient training despite temporal coupling in trajectory optimization. We show that JFB produces descent directions for the optimal control objective and experimentally demonstrate that our approach effectively learns high-dimensional feedback controllers across multiple scenarios involving implicit Hamiltonians, which existing methods cannot address.
Authors: Sagnik Anupam, Davis Brown, Shuo Li, Eric Wong, Hamed Hassani, Osbert Bastani
Abstract: LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.
Authors: Sina Alemohammad, Zhangyang Wang, Richard G. Baraniuk
Abstract: Scaling generative AI models is bottlenecked by the scarcity of high-quality training data. The ease of synthesizing from a generative model suggests using (unverified) synthetic data to augment a limited corpus of real data for the purpose of fine-tuning in the hope of improving performance. Unfortunately, however, the resulting positive feedback loop leads to model autophagy disorder (MAD, aka model collapse) that results in a rapid degradation in sample quality and/or diversity. In this paper, we introduce Neon (for Negative Extrapolation frOm self-traiNing), a new learning method that turns the degradation from self-training into a powerful signal for self-improvement. Given a base model, Neon first fine-tunes it on its own self-synthesized data but then, counterintuitively, reverses its gradient updates to extrapolate away from the degraded weights. We prove that Neon works because typical inference samplers that favor high-probability regions create a predictable anti-alignment between the synthetic and real data population gradients, which negative extrapolation corrects to better align the model with the true data distribution. Neon is remarkably easy to implement via a simple post-hoc merge that requires no new real data, works effectively with as few as 1k synthetic samples, and typically uses less than 1% additional training compute. We demonstrate Neon's universality across a range of architectures (diffusion, flow matching, autoregressive, and inductive moment matching models) and datasets (ImageNet, CIFAR-10, and FFHQ). In particular, on ImageNet 256x256, Neon elevates the xAR-L model to a new state-of-the-art FID of 1.02 with only 0.36% additional training compute. Code is available at https://github.com/VITA-Group/Neon
Authors: Yaxin Hou, Bo Han, Yuheng Jia, Hui Liu, Junhui Hou
Abstract: Current long-tailed semi-supervised learning methods assume that labeled data exhibit a long-tailed distribution, and unlabeled data adhere to a typical predefined distribution (i.e., long-tailed, uniform, or inverse long-tailed). However, the distribution of the unlabeled data is generally unknown and may follow an arbitrary distribution. To tackle this challenge, we propose a Controllable Pseudo-label Generation (CPG) framework, expanding the labeled dataset with the progressively identified reliable pseudo-labels from the unlabeled dataset and training the model on the updated labeled dataset with a known distribution, making it unaffected by the unlabeled data distribution. Specifically, CPG operates through a controllable self-reinforcing optimization cycle: (i) at each training step, our dynamic controllable filtering mechanism selectively incorporates reliable pseudo-labels from the unlabeled dataset into the labeled dataset, ensuring that the updated labeled dataset follows a known distribution; (ii) we then construct a Bayes-optimal classifier using logit adjustment based on the updated labeled data distribution; (iii) this improved classifier subsequently helps identify more reliable pseudo-labels in the next training step. We further theoretically prove that this optimization cycle can significantly reduce the generalization error under some conditions. Additionally, we propose a class-aware adaptive augmentation module to further improve the representation of minority classes, and an auxiliary branch to maximize data utilization by leveraging all labeled and unlabeled samples. Comprehensive evaluations on various commonly used benchmark datasets show that CPG achieves consistent improvements, surpassing state-of-the-art methods by up to $\textbf{15.97\%}$ in accuracy. The code is available at https://github.com/yaxinhou/CPG.
Authors: Dustin Wright, Sarah Masud, Jared Moore, Srishti Yadav, Maria Antoniak, Chan Young Park, Isabelle Augenstein
Abstract: Large language models (LLMs) tend to generate lexically, semantically, and stylistically homogenous texts. This poses a risk of knowledge collapse, where homogenous LLMs mediate a shrinking in the range of accessible information over time. Existing works on homogenization are limited by a focus on closed-ended multiple-choice setups or fuzzy semantic features, and do not look at trends across time and cultural contexts. To overcome this, we present a new methodology to measure epistemic diversity, i.e., variation in real-world claims in LLM outputs, which we use to perform a broad empirical study of LLM knowledge collapse. We test 27 LLMs, 155 topics covering 12 countries, and 200 prompt variations sourced from real user chats. For the topics in our study, we show that while newer models tend to generate more diverse claims, nearly all models are less epistemically diverse than a basic web search. We find that model size has a negative impact on epistemic diversity, while retrieval-augmented generation (RAG) has a positive impact, though the improvement from RAG varies by the cultural context. Finally, compared to a traditional knowledge source (Wikipedia), we find that country-specific claims reflect the English language more than the local one, highlighting a gap in epistemic representation
Authors: Khalid Mehtab Khan, Anagha Kulkarni
Abstract: Identifying cultural capital (CC) themes in student reflections can offer valuable insights that help foster equitable learning environments in classrooms. However, themes such as aspirational goals or family support are often woven into narratives, rather than appearing as direct keywords. This makes them difficult to detect for standard NLP models that process sentences in isolation. The core challenge stems from a lack of awareness, as standard models are pre-trained on general corpora, leaving them blind to the domain-specific language and narrative context inherent to the data. To address this, we introduce AWARE, a framework that systematically attempts to improve a transformer model's awareness for this nuanced task. AWARE has three core components: 1) Domain Awareness, adapting the model's vocabulary to the linguistic style of student reflections; 2) Context Awareness, generating sentence embeddings that are aware of the full essay context; and 3) Class Overlap Awareness, employing a multi-label strategy to recognize the coexistence of themes in a single sentence. Our results show that by making the model explicitly aware of the properties of the input, AWARE outperforms a strong baseline by 2.1 percentage points in Macro-F1 and shows considerable improvements across all themes. This work provides a robust and generalizable methodology for any text classification task in which meaning depends on the context of the narrative.