Authors: Xuanning Hu, Anchen Li, Qianli Xing, Jinglong Ji, Hao Tuo, Bo Yang
Abstract: Large Language Models (LLMs) possess strong representation and reasoning capabilities, but their application to structure-based drug design (SBDD) is limited by insufficient understanding of protein structures and unpredictable molecular generation. To address these challenges, we propose Exploration-Augmented Latent Inference for LLMs (ELILLM), a framework that reinterprets the LLM generation process as an encoding, latent space exploration, and decoding workflow. ELILLM explicitly explores portions of the design problem beyond the model's current knowledge while using a decoding module to handle familiar regions, generating chemically valid and synthetically reasonable molecules. In our implementation, Bayesian optimization guides the systematic exploration of latent embeddings, and a position-aware surrogate model efficiently predicts binding affinity distributions to inform the search. Knowledge-guided decoding further reduces randomness and effectively imposes chemical validity constraints. We demonstrate ELILLM on the CrossDocked2020 benchmark, showing strong controlled exploration and high binding affinity scores compared with seven baseline methods. These results demonstrate that ELILLM can effectively enhance LLMs capabilities for SBDD.
Authors: Shourya Jain, Paras Chopra
Abstract: Users should not be systemically disadvantaged by the language they use for interacting with LLMs; i.e. users across languages should get responses of similar quality irrespective of language used. In this work, we create a set of real-world open-ended questions based on our analysis of the WildChat dataset and use it to evaluate whether responses vary by language, specifically, whether answer quality depends on the language used to query the model. We also investigate how language and culture are entangled in LLMs such that choice of language changes the cultural information and context used in the response by using LLM-as-a-Judge to identify the cultural context present in responses. To further investigate this, we evaluate LLMs on a translated subset of the CulturalBench benchmark across multiple languages. Our evaluations reveal that LLMs consistently provide lower quality answers to open-ended questions in low resource languages. We find that language significantly impacts the cultural context used by the model. This difference in context impacts the quality of the downstream answer.
Authors: Maciej Kilian, Oleg Mkrtchyan, Luke Zettlemoyer, Akshat Shrivastava, Armen Aghajanyan
Abstract: Mixture-of-Experts layers achieve compute efficiency through weight sparsity: each token activates only a subset of experts. Data sparsity, where each expert processes only a subset of tokens, offers a complementary axis. Expert-choice routing implements data sparsity directly but violates causality in autoregressive models, creating train-inference mismatch. We recover data sparsity within causal token-choice MoE by leveraging zero-compute (null) experts within the routing pool. When a token routes to null experts, those slots consume no compute. The standard load balancing objective trains the model to uniformly use all experts (real and null) therefore creating data sparsity in expectation without the causality violations. We evaluate on vision-language model training, where data heterogeneity is pronounced: vision encoders produce many low-information tokens while text tokens are denser. At matched expected FLOPs, composing weight and data sparsity yields a more compute-efficient frontier than weight sparsity alone, with gains in training loss and downstream performance. The model learns implicit modality-aware allocation, routing vision tokens to null experts more aggressively than text, without explicit modality routing.
Authors: Elon Litman, Gabe Guo
Abstract: We generalize the attention mechanism by viewing it through the lens of Entropic Optimal Transport, revealing that standard attention corresponds to a transport problem regularized by an implicit uniform prior. We introduce Generalized Optimal transport Attention with Trainable priors (GOAT), a new attention mechanism that replaces this naive assumption with a learnable, continuous prior. This prior maintains full compatibility with optimized kernels such as FlashAttention. GOAT also provides an EOT-based explanation of attention sinks and materializes a solution for them, avoiding the representational trade-offs of standard attention. Finally, by absorbing spatial information into the core attention computation, GOAT learns an extrapolatable prior that combines the flexibility of learned positional embeddings with the length generalization of fixed encodings.
Authors: Zhaolong Su, Leheng Zhao, Xiaoying Wu, Ziyue Xu, Jindong Wang
Abstract: Unified multimodal models (UMMs) are emerging as strong foundation models that can do both generation and understanding tasks in a single architecture. However, they are typically trained in centralized settings where all training and downstream datasets are gathered in a central server, limiting the deployment in privacy-sensitive and geographically distributed scenarios. In this paper, we present FedUMM, a general federated learning framework for UMMs under non-IID multimodal data with low communication cost. Built on NVIDIA FLARE, FedUMM instantiates federation for a BLIP3o backbone via parameter-efficient fine-tuning: clients train lightweight LoRA adapters while freezing the foundation models, and the server aggregates only adapter updates. We evaluate on VQA v2 and the GenEval compositional generation benchmarks under Dirichlet-controlled heterogeneity with up to 16 clients. Results show slight degradation as client count and heterogeneity increase, while remaining competitive with centralized training. We further analyze computation--communication trade-offs and demonstrate that adapter-only federation reduces per-round communication by over an order of magnitude compared to full fine-tuning, enabling practical federated UMM training. This work provides empirical experience for future research on privacy-preserving federated unified multimodal models.
Authors: Ashna Nawar Ahmed, Banooqa Banday, Terry Jones, Tanzima Z. Islam
Abstract: High-Performance Computing (HPC) schedulers must balance user performance with facility-wide resource constraints. The task boils down to selecting the optimal number of nodes for a given job. We present a surrogate-assisted multi-objective Bayesian optimization (MOBO) framework to automate this complex decision. Our core hypothesis is that surrogate models informed by attention-based embeddings of job telemetry can capture performance dynamics more effectively than standard regression techniques. We pair this with an intelligent sample acquisition strategy to ensure the approach is data-efficient. On two production HPC datasets, our embedding-informed method consistently identified higher-quality Pareto fronts of runtime-power trade-offs compared to baselines. Furthermore, our intelligent data sampling strategy drastically reduced training costs while improving the stability of the results. To our knowledge, this is the first work to successfully apply embedding-informed surrogates in a MOBO framework to the HPC scheduling problem, jointly optimizing for performance and power on production workloads.
Authors: Adri\'an Rodr\'iguez-Mu\~noz, William Daspit, Adam Klivans, Antonio Torralba, Constantinos Daskalakis, Giannis Daras
Abstract: We propose Ambient Dataloops, an iterative framework for refining datasets that makes it easier for diffusion models to learn the underlying data distribution. Modern datasets contain samples of highly varying quality, and training directly on such heterogeneous data often yields suboptimal models. We propose a dataset-model co-evolution process; at each iteration of our method, the dataset becomes progressively higher quality, and the model improves accordingly. To avoid destructive self-consuming loops, at each generation, we treat the synthetically improved samples as noisy, but at a slightly lower noisy level than the previous iteration, and we use Ambient Diffusion techniques for learning under corruption. Empirically, Ambient Dataloops achieve state-of-the-art performance in unconditional and text-conditional image generation and de novo protein design. We further provide a theoretical justification for the proposed framework that captures the benefits of the data looping procedure.
Authors: Lorian Bannis
Abstract: We introduce Lattice, a hybrid sequential prediction system that conditionally activates learned behavioral structure using binary confidence gating. The system clusters behavior windows into behavioral archetypes and uses binary confidence gating to activate archetype-based scoring only when confidence exceeds a threshold, falling back to baseline predictions when uncertain. We validate Lattice on recommendation systems (MovieLens), scientific time-series (LIGO), and financial markets, using LSTM and transformer backbones. On MovieLens with LSTM, Lattice achieves +31.9% improvement over LSTM baseline in HR@10 (p < 3.29 x 10^-25, 30 seeds), outperforming transformer baselines by 109.4% over SASRec and 218.6% over BERT4Rec. On LIGO and financial data, the system correctly refuses archetype activation when distribution shift occurs - a successful outcome demonstrating confidence gating prevents false activation. On transformer backbones, Lattice provides 0.0% improvement (neutral, no degradation), gracefully deferring when structure is already present. This bidirectional validation - activating when patterns apply, refusing when they don't, and deferring when redundant - supports confidence gating as a promising architectural principle for managing epistemic uncertainty in safety-critical applications.
Authors: Zhenghao He, Guangzhi Xiong, Boyang Wang, Sanchit Sinha, Aidong Zhang
Abstract: Internal activations of diffusion models encode rich semantic information, but interpreting such representations remains challenging. While Sparse Autoencoders (SAEs) have shown promise in disentangling latent representations, existing SAE-based methods for diffusion model understanding rely on unsupervised approaches that fail to align sparse features with human-understandable concepts. This limits their ability to provide reliable semantic control over generated images. We introduce CASL (Concept-Aligned Sparse Latents), a supervised framework that aligns sparse latent dimensions of diffusion models with semantic concepts. CASL first trains an SAE on frozen U-Net activations to obtain disentangled latent representations, and then learns a lightweight linear mapping that associates each concept with a small set of relevant latent dimensions. To validate the semantic meaning of these aligned directions, we propose CASL-Steer, a controlled latent intervention that shifts activations along the learned concept axis. Unlike editing methods, CASL-Steer is used solely as a causal probe to reveal how concept-aligned latents influence generated content. We further introduce the Editing Precision Ratio (EPR), a metric that jointly measures concept specificity and the preservation of unrelated attributes. Experiments show that our method achieves superior editing precision and interpretability compared to existing approaches. To the best of our knowledge, this is the first work to achieve supervised alignment between latent representations and semantic concepts in diffusion models.
Authors: Kareem Amin, Alex Bie, Weiwei Kong, Umar Syed, Sergei Vassilvitskii
Abstract: The prevalence and low cost of LLMs have led to a rise of synthetic content. From review sites to court documents, ``natural'' content has been contaminated by data points that appear similar to natural data, but are in fact LLM-generated. In this work we revisit fundamental learning theory questions in this, now ubiquitous, setting. We model this scenario as a sequence of learning tasks where the input is a mix of natural and synthetic data, and the learning algorithms are oblivious to the origin of any individual example. We study the possibilities and limitations of ERM in this setting. For the problem of estimating the mean of an arbitrary $d$-dimensional distribution, we find that while ERM converges to the true mean, it is outperformed by an algorithm that assigns non-uniform weights to examples from different generations of data. For the PAC learning setting, the disparity is even more stark. We find that ERM does not always converge to the true concept, echoing the model collapse literature. However, we show there are algorithms capable of learning the correct hypothesis for arbitrary VC classes and arbitrary amounts of contamination.
Authors: Fahd Seddik, Abdulrahman Elbedewy, Gaser Sami, Mohamed Abdelmoniem, Yahia Zakaria
Abstract: Training modern deep learning models is increasingly constrained by GPU memory and compute limits. While Randomized Numerical Linear Algebra (RandNLA) offers proven techniques to compress these models, the lack of a unified, production-grade library prevents widely adopting these methods. We present Panther, a PyTorch-compatible library that consolidates established RandNLA algorithms into a single high-performance framework. Panther engineers efficient, drop-in replacements for standard components including sketched linear layers, 2D convolution, multi-head attention, and randomized matrix decompositions (such as pivoted CholeskyQR). By implementing a custom C++/CUDA backend (pawX), Panther provides an optimized implementation that can run on both CPUs and GPUs. We demonstrate the effectiveness of RandNLA techniques and Panther's ease of adoption. By replacing standard PyTorch linear layers with Panther layers (requiring only a few lines of code) we achieve significant memory savings (up to 75%) on BERT while maintaining comparable loss. Source code is available (MIT License) at https://github.com/FahdSeddik/panther, along with demonstration video at https://youtu.be/7M3RQb4KWxs.
URLs: https://github.com/FahdSeddik/panther,, https://youtu.be/7M3RQb4KWxs.
Authors: Md Nabi Newaz Khan, Abdullah Arafat Miah, Yu Bi
Abstract: Graph neural network (GNN) have demonstrated exceptional performance in solving critical problems across diverse domains yet remain susceptible to backdoor attacks. Existing studies on backdoor attack for graph classification are limited to single target attack using subgraph replacement based mechanism where the attacker implants only one trigger into the GNN model. In this paper, we introduce the first multi-targeted backdoor attack for graph classification task, where multiple triggers simultaneously redirect predictions to different target labels. Instead of subgraph replacement, we propose subgraph injection which preserves the structure of the original graphs while poisoning the clean graphs. Extensive experiments demonstrate the efficacy of our approach, where our attack achieves high attack success rates for all target labels with minimal impact on the clean accuracy. Experimental results on five dataset demonstrate the superior performance of our attack framework compared to the conventional subgraph replacement-based attack. Our analysis on four GNN models confirms the generalization capability of our attack which is effective regardless of the GNN model architectures and training parameters settings. We further investigate the impact of the attack design parameters including injection methods, number of connections, trigger sizes, trigger edge density and poisoning ratios. Additionally, our evaluation against state-of-the-art defenses (randomized smoothing and fine-pruning) demonstrates the robustness of our proposed multi-target attacks. This work highlights the GNN vulnerability against multi-targeted backdoor attack in graph classification task. Our source codes will be available at https://github.com/SiSL-URI/Multi-Targeted-Graph-Backdoor-Attack.
URLs: https://github.com/SiSL-URI/Multi-Targeted-Graph-Backdoor-Attack.
Authors: Jakub Antczak, James Montgomery, Ma{\l}gorzata O'Reilly, Zbigniew Palmowski, Richard Turner
Abstract: Emergency Department overcrowding is a critical issue that compromises patient safety and operational efficiency, necessitating accurate demand forecasting for effective resource allocation. This study evaluates and compares three distinct predictive models: Seasonal AutoRegressive Integrated Moving Average with eXogenous regressors (SARIMAX), EXtreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) networks for forecasting daily ED arrivals over a seven-day horizon. Utilizing data from an Australian tertiary referral hospital spanning January 2017 to December 2021, this research distinguishes itself by decomposing demand into eight specific ward categories and stratifying patients by clinical complexity. To address data distortions caused by the COVID-19 pandemic, the study employs the Prophet model to generate synthetic counterfactual values for the anomalous period. Experimental results demonstrate that all three proposed models consistently outperform a seasonal naive baseline. XGBoost demonstrated the highest accuracy for predicting total daily admissions with a Mean Absolute Error of 6.63, while the statistical SARIMAX model proved marginally superior for forecasting major complexity cases with an MAE of 3.77. The study concludes that while these techniques successfully reproduce regular day-to-day patterns, they share a common limitation in underestimating sudden, infrequent surges in patient volume.
Authors: Huayu Li, ZhengXiao He, Siyuan Tian, Jinghao Wen, Ao Li
Abstract: Standard autoregressive decoding in large language models (LLMs) is inherently short-sighted, often failing to find globally optimal reasoning paths due to its token-by-token generation process. While inference-time strategies like foresight sampling attempt to mitigate this by simulating future steps, they typically rely on ad-hoc heuristics for valuing paths and pruning the search space. This paper introduces Martingale Foresight Sampling (MFS), a principled framework that reformulates LLM decoding as a problem of identifying an optimal stochastic process. By modeling the quality of a reasoning path as a stochastic process, we leverage Martingale theory to design a theoretically-grounded algorithm. Our approach replaces heuristic mechanisms with principles from probability theory: step valuation is derived from the Doob Decomposition Theorem to measure a path's predictable advantage, path selection uses Optional Stopping Theory for principled pruning of suboptimal candidates, and an adaptive stopping rule based on the Martingale Convergence Theorem terminates exploration once a path's quality has provably converged. Experiments on six reasoning benchmarks demonstrate that MFS surpasses state-of-the-art methods in accuracy while significantly improving computational efficiency. Code will be released at https://github.com/miraclehetech/EACL2026-Martingale-Foresight-Sampling.
URLs: https://github.com/miraclehetech/EACL2026-Martingale-Foresight-Sampling.
Authors: Jingwei Song, Xinyu Wang, Hanbin Wang, Xiaoxuan Lei, Bill Shi, Shixin Han, Eric Yang, Xiao-Wen Chang, Lynn Ai
Abstract: Speculative Decoding (SD) accelerates autoregressive large language model (LLM) inference by decoupling generation and verification. While recent methods improve draft quality by tightly coupling the drafter with the target model, the verification mechanism itself remains largely unchanged, relying on strict token-level rejection sampling. In practice, modern LLMs frequently operate in low-margin regimes where the target model exhibits weak preference among top candidates. In such cases, rejecting plausible runner-up tokens yields negligible information gain while incurring substantial rollback cost, leading to a fundamental inefficiency in verification. We propose Margin-Aware Speculative Verification, a training-free and domain-agnostic verification strategy that adapts to the target model's local decisiveness. Our method conditions verification on decision stability measured directly from the target logits and relaxes rejection only when strict verification provides minimal benefit. Importantly, the approach modifies only the verification rule and is fully compatible with existing target-coupled speculative decoding frameworks. Extensive experiments across model scales ranging from 8B to 235B demonstrate that our method delivers consistent and significant inference speedups over state-of-the-art baselines while preserving generation quality across diverse benchmarks.
Authors: Rishit Chatterjee, Tahiya Chowdhury
Abstract: Volunteer-led lake monitoring yields irregular, seasonal time series with many gaps arising from ice cover, weather-related access constraints, and occasional human errors, complicating forecasting and early warning of harmful algal blooms. We study Secchi Disk Depth (SDD) forecasting on a 30-lake, data-rich subset drawn from three decades of in situ records collected across Maine lakes. Missingness is handled via Multiple Imputation by Chained Equations (MICE), and we evaluate performance with a normalized Mean Absolute Error (nMAE) metric for cross-lake comparability. Among six candidates, ridge regression provides the best mean test performance. Using ridge regression, we then quantify the minimal sample size, showing that under a backward, recent-history protocol, the model reaches within 5% of full-history accuracy with approximately 176 training samples per lake on average. We also identify a minimal feature set, where a compact four-feature subset matches the thirteen-feature baseline within the same 5% tolerance. Bringing these results together, we introduce a joint feasibility function that identifies the minimal training history and fewest predictors sufficient to achieve the target of staying within 5% of the complete-history, full-feature baseline. In our study, meeting the 5% accuracy target required about 64 recent samples and just one predictor per lake, highlighting the practicality of targeted monitoring. Hence, our joint feasibility strategy unifies recent-history length and feature choice under a fixed accuracy target, yielding a simple, efficient rule for setting sampling effort and measurement priorities for lake researchers.
Authors: Xianghao Zhan, Jingyu Xu, Yuanning Zheng, Zinaida Good, Olivier Gevaert
Abstract: Spatial transcriptomics enables spatial gene expression profiling, motivating computational models that capture spatially conditioned regulatory relationships. We introduce SAGE-FM, a lightweight spatial transcriptomics foundation model based on graph convolutional networks (GCNs) trained with a masked central spot prediction objective. Trained on 416 human Visium samples spanning 15 organs, SAGE-FM learns spatially coherent embeddings that robustly recover masked genes, with 91% of masked genes showing significant correlations (p < 0.05). The embeddings generated by SAGE-FM outperform MOFA and existing spatial transcriptomics methods in unsupervised clustering and preservation of biological heterogeneity. SAGE-FM generalizes to downstream tasks, enabling 81% accuracy in pathologist-defined spot annotation in oropharyngeal squamous cell carcinoma and improving glioblastoma subtype prediction relative to MOFA. In silico perturbation experiments further demonstrate that the model captures directional ligand-receptor and upstream-downstream regulatory effects consistent with ground truth. These results demonstrate that simple, parameter-efficient GCNs can serve as biologically interpretable and spatially aware foundation models for large-scale spatial transcriptomics.
Authors: Megan A. Witherow, Michael L. Evans, Ahmed Temtam, Hamid Okhravi, Khan M. Iftekharuddin
Abstract: Alzheimer's disease (AD), defined as an abnormal buildup of amyloid plaques and tau tangles in the brain can be diagnosed with high accuracy based on protein biomarkers via PET or CSF analysis. However, due to the invasive nature of biomarker collection, most AD diagnoses are made in memory clinics using cognitive tests and evaluation of hippocampal atrophy based on MRI. While clinical assessment and hippocampal volume show high diagnostic accuracy for amnestic or typical AD (tAD), a substantial subgroup of AD patients with atypical presentation (atAD) are routinely misdiagnosed. To improve diagnosis of atAD patients, we propose a machine learning approach to distinguish between atAD and non-AD cognitive impairment using clinical testing battery and MRI data collected as standard-of-care. We develop and evaluate our approach using 1410 subjects across four groups (273 tAD, 184 atAD, 235 non-AD, and 685 cognitively normal) collected from one private data set and two public data sets from the National Alzheimer's Coordinating Center (NACC) and the Alzheimer's Disease Neuroimaging Initiative (ADNI). We perform multiple atAD vs. non-AD classification experiments using clinical features and hippocampal volume as well as a comprehensive set of MRI features from across the brain. The best performance is achieved by incorporating additional important MRI features, which outperforms using hippocampal volume alone. Furthermore, we use the Boruta statistical approach to identify and visualize significant brain regions distinguishing between diagnostic groups. Our ML approach improves the percentage of correctly diagnosed atAD cases (the recall) from 52% to 69% for NACC and from 34% to 77% for ADNI, while achieving high precision. The proposed approach has important implications for improving diagnostic accuracy for non-amnestic atAD in clinical settings using only clinical testing battery and MRI.
Authors: Himanshu Mishra, Kanwal Mehreen
Abstract: Machine unlearning aims to remove specific knowledge (e.g., copyrighted or private data) from a trained model without full retraining. In practice, models are often quantized (e.g., 4-bit) for deployment, but we find that quantization can catastrophically restore forgotten information [1]. In this paper, we (1) analyze why low-bit quantization undermines unlearning, and (2) propose a quantization-aware unlearning method to mitigate this. We first compute weight-change statistics and bucket overlaps in quantization to show that typical unlearning updates are too small to cross quantization thresholds. Building on this insight, we introduce a logits space hinge loss: for each forget example, we force the output logits of the unlearned model to differ from the original model by at least a margin (half the quantization step). This ensures forgotten examples remain distinguishable even after quantization. We evaluate on language and classification tasks (including a Twitter misinformation dataset) and show our method preserves forgetting under 4-bit quantization, whereas existing methods almost entirely recover the forgotten knowledge.
Authors: Dongchen Huang
Abstract: Deep learning models, particularly Transformers, are often criticized as "black boxes" and lack interpretability. We propose Prism, a white-box attention-based architecture derived from the principles of Maximizing Coding Rate Reduction ($\text{MCR}^2$). By modeling the attention mechanism as a gradient ascent process on a distinct signal-noise manifold, we introduce two physical constraints: an overcomplete dictionary to expand the representational phase space, and an irrational frequency separation ($\pi$-RoPE) to enforce incoherence between signal and noise subspaces. We demonstrate that these geometric inductive biases can be viewed as a physical constraint and they are sufficient to induce unsupervised functional disentanglement alone. Using TinyStories as a controlled testbed for verifying spectral dynamics, we observe that Prism spontaneously specializes its attention heads into spectrally distinct regimes: low-frequency heads capturing long-range causal dependencies (signal) and high-frequency heads handling local syntactic constraints (noise). Our results suggest that interpretability and performance are not a trade-off, but can be unified through principled geometric construction.
Authors: Himanshu Mishra
Abstract: Continual Test-Time Adaptation (CTTA) seeks to update a pretrained model during deployment using only the incoming, unlabeled data stream. Although prior approaches such as Tent, EATA etc. provide meaningful improvements under short evolving shifts, they struggle when the test distribution changes rapidly or over extremely long horizons. This challenge is exemplified by the CCC benchmark, where models operate over streams of 7.5M samples with continually changing corruption types and severities. We propose RDumb++, a principled extension of RDumb that introduces two drift-detection mechanisms i.e entropy-based drift scoring and KL-divergence drift scoring, together with adaptive reset strategies. These mechanisms allow the model to detect when accumulated adaptation becomes harmful and to recover before prediction collapse occurs. Across CCC-medium with three speeds and three seeds (nine runs, each containing one million samples), RDumb++ consistently surpasses RDumb, yielding approx 3% absolute accuracy gains while maintaining stable adaptation throughout the entire stream. Ablation experiments on drift thresholds and reset strengths further show that drift-aware resetting is essential for preventing collapse and achieving reliable long-horizon CTTA.
Authors: Charles B. Delahunt, Courosh Mehanian, Daniel E. Shea, Matthew P. Horning
Abstract: A key task in ML is to optimize models at various stages, e.g. by choosing hyperparameters or picking a stopping point. A traditional ML approach is to use validation loss, i.e. to apply the training loss function on a validation set to guide these optimizations. However, ML for healthcare has a distinct goal from traditional ML: Models must perform well relative to specific clinical requirements, vs. relative to the loss function used for training. These clinical requirements can be captured more precisely by tailored metrics. Since many optimization tasks do not require the driving metric to be differentiable, they allow a wider range of options, including the use of metrics tailored to be clinically-relevant. In this paper we describe two controlled experiments which show how the use of clinically-tailored metrics provide superior model optimization compared to validation loss, in the sense of better performance on the clinical task. The use of clinically-relevant metrics for optimization entails some extra effort, to define the metrics and to code them into the pipeline. But it can yield models that better meet the central goal of ML for healthcare: strong performance in the clinic.
Authors: Jingren Hou, Hong Wang, Pengyu Xu, Chang Gao, Huafeng Liu, Liping Jing
Abstract: Real-world scientific applications frequently encounter incomplete observational data due to sensor limitations, geographic constraints, or measurement costs. Although neural operators significantly advanced PDE solving in terms of computational efficiency and accuracy, their underlying assumption of fully-observed spatial inputs severely restricts applicability in real-world applications. We introduce the first systematic framework for learning neural operators from partial observation. We identify and formalize two fundamental obstacles: (i) the supervision gap in unobserved regions that prevents effective learning of physical correlations, and (ii) the dynamic spatial mismatch between incomplete inputs and complete solution fields. Specifically, our proposed Latent Autoregressive Neural Operator~(\ours) introduces two novel components designed explicitly to address the core difficulties of partial observations: (i) a mask-to-predict training strategy that creates artificial supervision by strategically masking observed regions, and (ii) a Physics-Aware Latent Propagator that reconstructs solutions through boundary-first autoregressive generation in latent space. Additionally, we develop POBench-PDE, a dedicated and comprehensive benchmark designed specifically for evaluating neural operators under partial observation conditions across three PDE-governed tasks. \ours achieves state-of-the-art performance with 18--69$\%$ relative L2 error reduction across all benchmarks under patch-wise missingness with less than 50$\%$ missing rate, including real-world climate prediction. Our approach effectively addresses practical scenarios involving up to 75$\%$ missing rate, to some extent bridging the existing gap between idealized research settings and the complexities of real-world scientific computing.
Authors: Phuc Nguyen, Benjamin Zelditch, Joyce Chen, Rohit Patra, Changshuai Wei
Abstract: We present BanditLP, a scalable multi-stakeholder contextual bandit framework that unifies neural Thompson Sampling for learning objective-specific outcomes with a large-scale linear program for constrained action selection at serving time. The methodology is application-agnostic, compatible with arbitrary neural architectures, and deployable at web scale, with an LP solver capable of handling billions of variables. Experiments on public benchmarks and synthetic data show consistent gains over strong baselines. We apply this approach in LinkedIn's email marketing system and demonstrate business win, illustrating the value of integrated exploration and constrained optimization in production.
Authors: Xuan Liao, Zhenkang Peng, Ying Rong
Abstract: Managing perishable products with limited lifetimes is a fundamental challenge in inventory management, as poor ordering decisions can quickly lead to stockouts or excessive waste. We study a perishable inventory system with random lead times in which both the demand process and the lead time distribution are unknown. We consider a practical setting where orders are placed using limited historical data together with observed covariates and current system states. To improve learning efficiency under limited data, we adopt a marginal cost accounting scheme that assigns each order a single lifetime cost and yields a unified loss function for end-to-end learning. This enables training a deep learning-based policy that maps observed covariates and system states directly to order quantities. We develop two end-to-end variants: a purely black-box approach that outputs order quantities directly (E2E-BB), and a structure-guided approach that embeds the projected inventory level (PIL) policy, capturing inventory effects through explicit computation rather than additional learning (E2E-PIL). We further show that the objective induced by E2E-PIL is homogeneous of degree one, enabling a boosting technique from operational data analytics (ODA) that yields an enhanced policy (E2E-BPIL). Experiments on synthetic and real data establish a robust performance ordering: E2E-BB is dominated by E2E-PIL, which is further improved by E2E-BPIL. Using an excess-risk decomposition, we show that embedding heuristic policy structure reduces effective model complexity and improves learning efficiency with only a modest loss of flexibility. More broadly, our results suggest that deep learning-based decision tools are more effective and robust when guided by human knowledge, highlighting the value of integrating advanced analytics with inventory theory.
Authors: Liusha Yang, Siqi Zhao, Shuqi Chai
Abstract: This paper introduces a neural network-based nonlinear shrinkage estimator of covariance matrices for the purpose of minimum variance portfolio optimization. It is a hybrid approach that integrates statistical estimation with machine learning. Starting from the Ledoit-Wolf (LW) shrinkage estimator, we decompose the LW covariance matrix into its eigenvalues and eigenvectors, and apply a lightweight transformer-based neural network to learn a nonlinear eigenvalue shrinkage function. Trained with portfolio risk as the loss function, the resulting precision matrix (the inverse covariance matrix) estimator directly targets portfolio risk minimization. By conditioning on the sample-to-dimension ratio, the approach remains scalable across different sample sizes and asset universes. Empirical results on stock daily returns from Standard & Poor's 500 Index (S&P500) demonstrate that the proposed method consistently achieves lower out-of-sample realized risk than benchmark approaches. This highlights the promise of integrating structural statistical models with data-driven learning.
Authors: Mingyuan Fan, Weiguang Han, Daixin Wang, Cen Chen, Zhiqiang Zhang, Jun Zhou
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is a central paradigm for turning large language models (LLMs) into reliable problem solvers, especially in logic-heavy domains. Despite its empirical success, it remains unclear whether RLVR elicits novel capabilities or merely sharpens the distribution over existing knowledge. We study this by formalizing over-sharpening, a phenomenon where the policy collapses onto limited modes, suppressing valid alternatives. At a high level, we discover finite-batch updates intrinsically bias learning toward sampled modes, triggering a collapse that propagates globally via semantic coupling. To mitigate this, we propose inverse-success advantage calibration to prioritize difficult queries and distribution-level calibration to diversify sampling via a memory network. Empirical evaluations validate that our strategies can effectively improve generalization.
Authors: Zitian Li, Wang Chi Cheung
Abstract: 1-identification is a fundamental multi-armed bandit formulation on pure exploration. An agent aims to determine whether there exists a qualified arm whose mean reward is not less than a known threshold $\mu_0$, or to output \textsf{None} if it believes such an arm does not exist. The agent needs to guarantee its output is correct with probability at least $1-\delta$, while making expected total pulling times $\mathbb{E}\tau$ as small as possible. We work on 1-identification with two main contributions. (1) We utilize an optimization formulation to derive a new lower bound of $\mathbb{E}\tau$, when there is at least one qualified arm. (2) We design a new algorithm, deriving tight upper bounds whose gap to lower bounds are up to a polynomial of logarithm factor across all problem instance. Our result complements the analysis of $\mathbb{E}\tau$ when there are multiple qualified arms, which is an open problem left by history literature.
Authors: Zhiwei Zhang, Fei Zhao, Rui Wang, Zezhong Wang, Bin Liang, Jiakang Wang, Yao Hu, Shaosheng Cao, Kam-Fai Wong
Abstract: Large language models (LLMs) can call tools effectively, yet they remain brittle in multi-turn execution: following a tool call error, smaller models often degenerate into repetitive invalid re-invocations, failing to interpret error feedback and self-correct. This brittleness hinders reliable real-world deployment, where the execution errors are inherently inevitable during tool interaction procedures. We identify a key limitation of current approaches: standard reinforcement learning (RL) treats errors as sparse negative rewards, providing no guidance on how to recover, while pre-collected synthetic error-correction datasets suffer from distribution mismatch with the model's on-policy error modes. To bridge this gap, we propose Fission-GRPO, a framework that converts execution errors into corrective supervision within the RL training loop. Our core mechanism fissions each failed trajectory into a new training instance by augmenting it with diagnostic feedback from a finetuned Error Simulator, then resampling recovery rollouts on-policy. This enables the model to learn from the precise errors it makes during exploration, rather than from static, pre-collected error cases. On the BFCL v4 Multi-Turn, Fission-GRPO improves the error recovery rate of Qwen3-8B by 5.7% absolute, crucially, yielding a 4% overall accuracy gain (42.75% to 46.75%) over GRPO and outperforming specialized tool-use agents.
Authors: Natasha Trinkle, Huong Ha, Jeffrey Chan
Abstract: Bayesian optimisation is a sample efficient method for finding a global optimum of expensive black-box objective functions. Historic datasets from related problems can be exploited to help improve performance of Bayesian optimisation by adapting transfer learning methods to various components of the Bayesian optimisation pipeline. In this study we perform an empirical analysis of various ensemble-based transfer learning Bayesian optimisation methods and pipeline components. We expand on previous work in the literature by contributing some specific pipeline components, and three new real-time transfer learning Bayesian optimisation benchmarks. In particular we propose to use a weighting strategy for ensemble surrogate model predictions based on regularised regression with weights constrained to be positive, and a related component for handling the case when transfer learning is not improving Bayesian optimisation performance. We find that in general, two components that help improve transfer learning Bayesian optimisation performance are warm start initialisation and constraining weights used with ensemble surrogate model to be positive.
Authors: Yinxi Tian, Changwu Huang, Ke Tang, Xin Yao
Abstract: Knowledge distillation (KD) transfers knowledge from large teacher models to compact student models, enabling efficient deployment on resource constrained devices. While diverse KD methods, including response based, feature based, and relation based approaches, capture different aspects of teacher knowledge, integrating multiple methods or knowledge sources is promising but often hampered by complex implementation, inflexible combinations, and catastrophic forgetting, which limits practical effectiveness. This work proposes SMSKD (Sequential Multi Stage Knowledge Distillation), a flexible framework that sequentially integrates heterogeneous KD methods. At each stage, the student is trained with a specific distillation method, while a frozen reference model from the previous stage anchors learned knowledge to mitigate forgetting. In addition, we introduce an adaptive weighting mechanism based on the teacher true class probability (TCP) that dynamically adjusts the reference loss per sample to balance knowledge retention and integration. By design, SMSKD supports arbitrary method combinations and stage counts with negligible computational overhead. Extensive experiments show that SMSKD consistently improves student accuracy across diverse teacher student architectures and method combinations, outperforming existing baselines. Ablation studies confirm that stage wise distillation and reference model supervision are primary contributors to performance gains, with TCP based adaptive weighting providing complementary benefits. Overall, SMSKD is a practical and resource efficient solution for integrating heterogeneous KD methods.
Authors: Jingjing Bai, Yoshinobu Kawahara
Abstract: Transformer-based models, despite their promise for long-term time series forecasting (LTSF), suffer from an inherent low-pass filtering effect that limits their effectiveness. This issue arises due to undifferentiated propagation of frequency components across layers, causing a progressive attenuation of high-frequency information crucial for capturing fine-grained temporal variations. To address this limitation, we propose Dualformer, a principled dual-domain framework that rethinks frequency modeling from a layer-wise perspective. Dualformer introduces three key components: (1) a dual-branch architecture that concurrently models complementary temporal patterns in both time and frequency domains; (2) a hierarchical frequency sampling module that allocates distinct frequency bands to different layers, preserving high-frequency details in lower layers while modeling low-frequency trends in deeper layers; and (3) a periodicity-aware weighting mechanism that dynamically balances contributions from the dual branches based on the harmonic energy ratio of inputs, supported theoretically by a derived lower bound. This design enables structured frequency modeling and adaptive integration of time-frequency features, effectively preserving high-frequency information and enhancing generalization. Extensive experiments conducted on eight widely used benchmarks demonstrate Dualformer's robustness and superior performance, particularly on heterogeneous or weakly periodic data. Our code is publicly available at https://github.com/Akira-221/Dualformer.
Authors: Xinyu Wang, Sicheng Lyu, Yu Gu, Jerry Huang, Peng Lu, Yufei Cui, Xiao-Wen Chang
Abstract: Model editing updates a pre-trained LLM with new facts or rules without re-training, while preserving unrelated behavior. In real deployment, edits arrive as long streams, and existing editors often face a plasticity-stability dilemma: locate-then-edit "hard writes" can accumulate interference over time, while null-space-style "hard preservation" preserves only what is explicitly constrained, so past edits can be overwritten and unconstrained behaviors may deviate, degrading general capabilities in the many-edits regime. We propose RLSEdit, a recursive least-squares editor for long sequential editing. RLSEdit formulates editing as an online quadratic optimization with soft constraints, minimizing a cumulative key-value fitting objective with two regularizers that control for both deviation from the pre-trained weights and from a designated anchor mapping. The resulting update admits an efficient online recursion via the Woodbury identity, with per-edit cost independent of history length and scaling only with the current edit size. We further provide deviation bounds and an asymptotic characterization of the adherence-preservation trade-off in the many-edits regime. Experiments on multiple model families demonstrate stable scaling to 10K edits, outperforming strong baselines in both edit success and holistic stability -- crucially retaining early edits, and preserving general capabilities on GLUE and held-out reasoning/code benchmarks.
Authors: Ryoma Sato
Abstract: We propose Zero-Error Horizon (ZEH) for trustworthy LLMs, which represents the maximum range that a model can solve without any errors. While ZEH itself is simple, we demonstrate that evaluating the ZEH of state-of-the-art LLMs yields abundant insights. For example, by evaluating the ZEH of GPT-5.2, we found that GPT-5.2 cannot even compute the parity of a short string like 11000, and GPT-5.2 cannot determine whether the parentheses in ((((()))))) are balanced. This is surprising given the excellent capabilities of GPT-5.2. The fact that LLMs make mistakes on such simple problems serves as an important lesson when applying LLMs to safety-critical domains. By applying ZEH to Qwen2.5 and conducting detailed analysis, we found that while ZEH correlates with accuracy, the detailed behaviors differ, and ZEH provides clues about the emergence of algorithmic capabilities. Finally, while computing ZEH incurs significant computational cost, we discuss how to mitigate this cost by achieving up to one order of magnitude speedup using tree structures and online softmax.
Authors: Xiuling Wang, Xin Huang, Haibo Hu, Jianliang Xu
Abstract: Graph Neural Networks (GNNs) unlock new ways of learning from graph-structured data, proving highly effective in capturing complex relationships and patterns. Federated GNNs (FGNNs) have emerged as a prominent distributed learning paradigm for training GNNs over decentralized data. However, FGNNs face two significant challenges: high communication overhead from multiple rounds of parameter exchanges and non-IID data characteristics across clients. To address these issues, we introduce CeFGC, a novel FGNN paradigm that facilitates efficient GNN training over non-IID data by limiting communication between the server and clients to three rounds only. The core idea of CeFGC is to leverage generative diffusion models to minimize direct client-server communication. Each client trains a generative diffusion model that captures its local graph distribution and shares this model with the server, which then redistributes it back to all clients. Using these generative models, clients generate synthetic graphs combined with their local graphs to train local GNN models. Finally, clients upload their model weights to the server for aggregation into a global GNN model. We theoretically analyze the I/O complexity of communication volume to show that CeFGC reduces to a constant of three communication rounds only. Extensive experiments on several real graph datasets demonstrate the effectiveness and efficiency of CeFGC against state-of-the-art competitors, reflecting our superior performance on non-IID graphs by aligning local and global model objectives and enriching the training set with diverse graphs.
Authors: Yang Yu, Peiyu Zang, Chi Hsu Tsai, Haiming Wu, Yixin Shen, Jialing Zhang, Haoyu Wang, Zhiyou Xiao, Jingze Shi, Yuyu Luo, Wentao Zhang, Chunlei Men, Guang Liu, Yonghua Lin
Abstract: The performance of modern AI systems is fundamentally constrained by the quality of their underlying kernels, which translate high-level algorithmic semantics into low-level hardware operations. Achieving near-optimal kernels requires expert-level understanding of hardware architectures and programming models, making kernel engineering a critical but notoriously time-consuming and non-scalable process. Recent advances in large language models (LLMs) and LLM-based agents have opened new possibilities for automating kernel generation and optimization. LLMs are well-suited to compress expert-level kernel knowledge that is difficult to formalize, while agentic systems further enable scalable optimization by casting kernel development as an iterative, feedback-driven loop. Rapid progress has been made in this area. However, the field remains fragmented, lacking a systematic perspective for LLM-driven kernel generation. This survey addresses this gap by providing a structured overview of existing approaches, spanning LLM-based approaches and agentic optimization workflows, and systematically compiling the datasets and benchmarks that underpin learning and evaluation in this domain. Moreover, key open challenges and future research directions are further outlined, aiming to establish a comprehensive reference for the next generation of automated kernel optimization. To keep track of this field, we maintain an open-source GitHub repository at https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.
URLs: https://github.com/flagos-ai/awesome-LLM-driven-kernel-generation.
Authors: Dong Xu, Jiantao Wu, Qihua Pan, Sisi Yuan, Zexuan Zhu, Junkai Ji
Abstract: Drug-drug interaction (DDI) prediction is central to drug discovery and clinical development, particularly in the context of increasingly prevalent polypharmacy. Although existing computational methods achieve strong performance on standard benchmarks, they often fail to generalize to realistic deployment scenarios, where most candidate drug pairs involve previously unseen drugs and validated interactions are scarce. We demonstrate that proximity in the embedding spaces of prevailing molecule-centric DDI models does not reliably correspond to interaction labels, and that simply scaling up model capacity therefore fails to improve generalization. To address these limitations, we propose GenRel-DDI, a generalizable relation learning framework that reformulates DDI prediction as a relation-centric learning problem, in which interaction representations are learned independently of drug identities. This relation-level abstraction enables the capture of transferable interaction patterns that generalize to unseen drugs and novel drug pairs. Extensive experiments across multiple benchmark demonstrate that GenRel-DDI consistently and significantly outperforms state-of-the-art methods, with particularly large gains on strict entity-disjoint evaluations, highlighting the effectiveness and practical utility of relation learning for robust DDI prediction. The code is available at https://github.com/SZU-ADDG/GenRel-DDI.
Authors: Yuanyuan Qi, Xiaohao Yang, Jueqing Lu, Guoxiang Guo, Joanne Enticott, Gang Liu, Lan Du
Abstract: With the rapid advancement and strong generalization capabilities of large language models (LLMs), they have been increasingly incorporated into the active learning pipelines as annotators to reduce annotation costs. However, considering the annotation quality, labels generated by LLMs often fall short of real-world applicability. To address this, we propose a novel active learning framework, Mixture of LLMs in the Loop Active Learning, replacing human annotators with labels generated through a Mixture-of-LLMs-based annotation model, aimed at enhancing LLM-based annotation robustness by aggregating the strengths of multiple LLMs. To further mitigate the impact of the noisy labels, we introduce annotation discrepancy and negative learning to identify the unreliable annotations and enhance learning effectiveness. Extensive experiments demonstrate that our framework achieves performance comparable to human annotation and consistently outperforms single-LLM baselines and other LLM-ensemble-based approaches. Moreover, our framework is built on lightweight LLMs, enabling it to operate fully on local machines in real-world applications.
Authors: Fengheng Chu, Jiahao Chen, Yuhong Wang, Jun Wang, Zhihui Fu, Shouling Ji, Songze Li
Abstract: While Large Language Models (LLMs) are aligned to mitigate risks, their safety guardrails remain fragile against jailbreak attacks. This reveals limited understanding of components governing safety. Existing methods rely on local, greedy attribution that assumes independent component contributions. However, they overlook the cooperative interactions between different components in LLMs, such as attention heads, which jointly contribute to safety mechanisms. We propose \textbf{G}lobal \textbf{O}ptimization for \textbf{S}afety \textbf{V}ector Extraction (GOSV), a framework that identifies safety-critical attention heads through global optimization over all heads simultaneously. We employ two complementary activation repatching strategies: Harmful Patching and Zero Ablation. These strategies identify two spatially distinct sets of safety vectors with consistently low overlap, termed Malicious Injection Vectors and Safety Suppression Vectors, demonstrating that aligned LLMs maintain separate functional pathways for safety purposes. Through systematic analyses, we find that complete safety breakdown occurs when approximately 30\% of total heads are repatched across all models. Building on these insights, we develop a novel inference-time white-box jailbreak method that exploits the identified safety vectors through activation repatching. Our attack substantially outperforms existing white-box attacks across all test models, providing strong evidence for the effectiveness of the proposed GOSV framework on LLM safety interpretability.
Authors: Lina Felsner, Henriette Bast, Tina Dorosti, Florian Schaff, Franz Pfeiffer, Daniela Pfeiffer, Julia Schnabel
Abstract: X-ray dark-field radiography provides complementary diagnostic information to conventional attenuation imaging by visualizing microstructural tissue changes through small-angle scattering. However, the limited availability of such data poses challenges for developing robust deep learning models. In this work, we present the first framework for generating dark-field images directly from standard attenuation chest X-rays using an Uncertainty-Guided Progressive Generative Adversarial Network. The model incorporates both aleatoric and epistemic uncertainty to improve interpretability and reliability. Experiments demonstrate high structural fidelity of the generated images, with consistent improvement of quantitative metrics across stages. Furthermore, out-of-distribution evaluation confirms that the proposed model generalizes well. Our results indicate that uncertainty-guided generative modeling enables realistic dark-field image synthesis and provides a reliable foundation for future clinical applications.
Authors: Jidong Jin
Abstract: Inference in large-scale AI models is typically performed on dense parameter matrices, leading to inference cost and system complexity that scale unsustainably with model size. This limitation does not arise from insufficient model capacity, but from treating post-training inference systems as monolithic operators while ignoring internal structures formed during learning. We show that gradient update events in large models are highly localized and selective, leaving many parameter dependencies statistically indistinguishable from their initialization distribution after training. As a result, post-training inference systems are structurally non-uniform and inherently decomposable. Based on this observation, we introduce a post-training statistical criterion and a structural annealing procedure that removes unsupported dependencies and reveals stable, independent substructures. This work establishes a post-training, model-agnostic structural view of inference systems and enables structured, parallel inference without modifying model functionality or interfaces.
Authors: Cristina P\^era, T\^ania Carvalho, Maxime Cordy, Lu\'is Antunes
Abstract: Membership Inference Attacks (MIAs) are currently a dominant approach for evaluating privacy in machine learning applications. Despite their significance in identifying records belonging to the training dataset, several concerns remain unexplored, particularly with regard to tabular data. In this paper, first, we provide an extensive review and analysis of MIAs considering two main learning paradigms: centralized and federated learning. We extend and refine the taxonomy for both. Second, we demonstrate the efficacy of MIAs in tabular data using several attack strategies, also including defenses. Furthermore, in a federated learning scenario, we consider the threat posed by an outsider adversary, which is often neglected. Third, we demonstrate the high vulnerability of single-outs (records with a unique signature) to MIAs. Lastly, we explore how MIAs transfer across model architectures. Our results point towards a general poor performance of these attacks in tabular data which contrasts with previous state-of-the-art. Notably, even attacks with limited attack performance can still successfully expose a large portion of single-outs. Moreover, our findings suggest that using different surrogate models makes MIAs more effective.
Authors: Simon W. Penninga, Ruud J. G. van Sloun
Abstract: In this paper we propose the Iterative Amortized Hierarchical Variational Autoencoder (IA-HVAE), which expands on amortized inference with a hybrid scheme containing an initial amortized guess and iterative refinement with decoder gradients. We achieve this by creating a linearly separable decoder in a transform domain (e.g. Fourier space), enabling real-time applications with very high model depths. The architectural change leads to a 35x speed-up for iterative inference with respect to the traditional HVAE. We show that our hybrid approach outperforms fully amortized and fully iterative equivalents in accuracy and speed respectively. Moreover, the IAHVAE shows improved reconstruction quality over a vanilla HVAE in inverse problems such as deblurring and denoising.
Authors: Binbin Lin, Lei Zou, Hao Tian, Heng Cai, Yifan Yang, Bing Zhou
Abstract: Healthcare visitation patterns are influenced by a complex interplay of hospital attributes, population socioeconomics, and spatial factors. However, existing research often adopts a fragmented approach, examining these determinants in isolation. This study addresses this gap by integrating hospital capacities, occupancy rates, reputation, and popularity with population SES and spatial mobility patterns to predict visitation flows and analyze influencing factors. Utilizing four years of SafeGraph mobility data and user experience data from Google Maps Reviews, five flow prediction models, Naive Regression, Gradient Boosting, Multilayer Perceptrons (MLPs), Deep Gravity, and Heterogeneous Graph Neural Networks (HGNN),were trained and applied to simulate visitation flows in Houston, Texas, U.S. The Shapley additive explanation (SHAP) analysis and the Partial Dependence Plot (PDP) method were employed to examine the combined impacts of different factors on visitation patterns. The findings reveal that Deep Gravity outperformed other models. Hospital capacities, ICU occupancy rates, ratings, and popularity significantly influence visitation patterns, with their effects varying across different travel distances. Short-distance visits are primarily driven by convenience, whereas long-distance visits are influenced by hospital ratings. White-majority areas exhibited lower sensitivity to hospital ratings for short-distance visits, while Asian populations and those with higher education levels prioritized hospital rating in their visitation decisions. SES further influence these patterns, as areas with higher proportions of Hispanic, Black, under-18, and over-65 populations tend to have more frequent hospital visits, potentially reflecting greater healthcare needs or limited access to alternative medical services.
Authors: Naram Mhaisen, George Iosifidis
Abstract: We introduce $k$-lazyGD, an online learning algorithm that bridges the gap between greedy Online Gradient Descent (OGD, for $k=1$) and lazy GD/dual-averaging (for $k=T$), creating a spectrum between reactive and stable updates. We analyze this spectrum in Smoothed Online Convex Optimization (SOCO), where the learner incurs both hitting and movement costs. Our main contribution is establishing that laziness is possible without sacrificing hitting performance: we prove that $k$-lazyGD achieves the optimal dynamic regret $\mathcal{O}(\sqrt{(P_T+1)T})$ for any laziness slack $k$ up to $\Theta(\sqrt{T/P_T})$, where $P_T$ is the comparator path length. This result formally connects the allowable laziness to the comparator's shifts, showing that $k$-lazyGD can retain the inherently small movements of lazy methods without compromising tracking ability. We base our analysis on the Follow the Regularized Leader (FTRL) framework, and derive a matching lower bound. Since the slack depends on $P_T$, an ensemble of learners with various slacks is used, yielding a method that is provably stable when it can be, and agile when it must be.
Authors: Moritz Wedemeyer, Eike Cramer, Alexander Mitsos, Manuel Dahmen
Abstract: With the increasing flexibilization of processes, determining robust scheduling decisions has become an important goal. Traditionally, the flexibility index has been used to identify safe operating schedules by approximating the admissible uncertainty region using simple admissible uncertainty sets, such as hypercubes. Presently, available contextual information, such as forecasts, has not been considered to define the admissible uncertainty set when determining the flexibility index. We propose the conditional flexibility index (CFI), which extends the traditional flexibility index in two ways: by learning the parametrized admissible uncertainty set from historical data and by using contextual information to make the admissible uncertainty set conditional. This is achieved using a normalizing flow that learns a bijective mapping from a Gaussian base distribution to the data distribution. The admissible latent uncertainty set is constructed as a hypersphere in the latent space and mapped to the data space. By incorporating contextual information, the CFI provides a more informative estimate of flexibility by defining admissible uncertainty sets in regions that are more likely to be relevant under given conditions. Using an illustrative example, we show that no general statement can be made about data-driven admissible uncertainty sets outperforming simple sets, or conditional sets outperforming unconditional ones. However, both data-driven and conditional admissible uncertainty sets ensure that only regions of the uncertain parameter space containing realizations are considered. We apply the CFI to a security-constrained unit commitment example and demonstrate that the CFI can improve scheduling quality by incorporating temporal information.
Authors: Ricardo N. Ferreira, Cl\'audia Soares, Jo\~ao Xavier
Abstract: We study Constrained Online Convex Optimization (COCO), where a learner chooses actions iteratively, observes both unanticipated convex loss and convex constraint, and accumulates loss while incurring penalties for constraint violations. We introduce CLASP (Convex Losses And Squared Penalties), an algorithm that minimizes cumulative loss together with squared constraint violations. Our analysis departs from prior work by fully leveraging the firm non-expansiveness of convex projectors, a proof strategy not previously applied in this setting. For convex losses, CLASP achieves regret $O\left(T^{\max\{\beta,1-\beta\}}\right)$ and cumulative squared penalty $O\left(T^{1-\beta}\right)$ for any $\beta \in (0,1)$. Most importantly, for strongly convex problems, CLASP provides the first logarithmic guarantees on both regret and cumulative squared penalty. In the strongly convex case, the regret is upper bounded by $O( \log T )$ and the cumulative squared penalty is also upper bounded by $O( \log T )$.
Authors: Annemarie Jutte, Uraz Odyurt
Abstract: Industrial Cyber-Physical Systems (CPS) are sensitive infrastructure from both safety and economics perspectives, making their reliability critically important. Machine Learning (ML), specifically deep learning, is increasingly integrated in industrial CPS, but the inherent complexity of ML models results in non-transparent operation. Rigorous evaluation is needed to prevent models from exhibiting unexpected behaviour on future, unseen data. Explainable AI (XAI) can be used to uncover model reasoning, allowing a more extensive analysis of behaviour. We apply XAI to to improve predictive performance of ML models intended for industrial CPS. We analyse the effects of components from time-series data decomposition on model predictions using SHAP values. Through this method, we observe evidence on the lack of sufficient contextual information during model training. By increasing the window size of data instances, informed by the XAI findings, we are able to improve model performance.
Authors: Matthew Shorvon, Frederik Mallmann-Trenn, David S. Watson
Abstract: Computing the conditional mode of a distribution, better known as the $\mathit{maximum\ a\ posteriori}$ (MAP) assignment, is a fundamental task in probabilistic inference. However, MAP estimation is generally intractable, and remains hard even under many common structural constraints and approximation schemes. We introduce $\mathit{probably\ approximately\ correct}$ (PAC) algorithms for MAP inference that provide provably optimal solutions under variable and fixed computational budgets. We characterize tractability conditions for PAC-MAP using information theoretic measures that can be estimated from finite samples. Our PAC-MAP solvers are efficiently implemented using probabilistic circuits with appropriate architectures. The randomization strategies we develop can be used either as standalone MAP inference techniques or to improve on popular heuristics, fortifying their solutions with rigorous guarantees. Experiments confirm the benefits of our method in a range of benchmarks.
Authors: Adithya Sineesh, Akshita Kamsali
Abstract: Deep learning classifiers for Raman spectroscopy are increasingly reported to outperform classical chemometric approaches. However their evaluations are often conducted in isolation or compared against traditional machine learning methods or trivially adapted vision-based architectures that were not originally proposed for Raman spectroscopy. As a result, direct comparisons between existing deep learning models developed specifically for Raman spectral analysis on shared open-source datasets remain scarce. To the best of our knowledge, this study presents one of the first systematic benchmarks comparing three or more published Raman-specific deep learning classifiers across multiple open-source Raman datasets. We evaluate five representative deep learning architectures under a unified training and hyperparameter tuning protocol across three open-source Raman datasets selected to support standard evaluation, fine-tuning, and explicit distribution-shift testing. We report classification accuracies and macro-averaged F1 scores to provide a fair and reproducible comparison of deep learning models for Raman spectra based classification.
Authors: Yuta Nakahara, Shota Saito, Kohei Horinouchi, Koshi Shimada, Naoki Ichijo, Manabu Kobayashi, Toshiyasu Matsushima
Abstract: We propose a variable splitting binary tree (VSBT) model based on Bayesian context tree (BCT) models for time series segmentation. Unlike previous applications of BCT models, the tree structure in our model represents interval partitioning on the time domain. Moreover, interval partitioning is represented by recursive logistic regression models. By adjusting logistic regression coefficients, our model can represent split positions at arbitrary locations within each interval. This enables more compact tree representations. For simultaneous estimation of both split positions and tree depth, we develop an effective inference algorithm that combines local variational approximation for logistic regression with the context tree weighting (CTW) algorithm. We present numerical examples on synthetic data demonstrating the effectiveness of our model and algorithm.
Authors: Rustem Takhanov
Abstract: The manifold hypothesis suggests that the generalization performance of machine learning methods improves significantly when the intrinsic dimension of the input distribution's support is low. In the context of KRR, we investigate two alternative notions of intrinsic dimension. The first, denoted $d_\rho$, is the upper Minkowski dimension defined with respect to the canonical metric induced by a kernel function $K$ on a domain $\Omega$. The second, denoted $d_K$, is the effective dimension, derived from the decay rate of Kolmogorov $n$-widths associated with $K$ on $\Omega$. Given a probability measure $\mu$ on $\Omega$, we analyze the relationship between these $n$-widths and eigenvalues of the integral operator $\phi \to \int_\Omega K(\cdot,x)\phi(x)d\mu(x)$. We show that, for a fixed domain $\Omega$, the Kolmogorov $n$-widths characterize the worst-case eigenvalue decay across all probability measures $\mu$ supported on $\Omega$. These eigenvalues are central to understanding the generalization behavior of constrained KRR, enabling us to derive an excess error bound of order $O(n^{-\frac{2+d_K}{2+2d_K} + \epsilon})$ for any $\epsilon > 0$, when the training set size $n$ is large. We also propose an algorithm that estimates upper bounds on the $n$-widths using only a finite sample from $\mu$. For distributions close to uniform, we prove that $\epsilon$-accurate upper bounds on all $n$-widths can be computed with high probability using at most $O\left(\epsilon^{-d_\rho}\log\frac{1}{\epsilon}\right)$ samples, with fewer required for small $n$. Finally, we compute the effective dimension $d_K$ for various fractal sets and present additional numerical experiments. Our results show that, for kernels such as the Laplace kernel, the effective dimension $d_K$ can be significantly smaller than the Minkowski dimension $d_\rho$, even though $d_K = d_\rho$ provably holds on regular domains.
Authors: Muhammad Ilham Rizqyawan, Peter Macfarlane, Stathis Hadjidemetriou, Fani Deligianni
Abstract: Obtaining labelled ECG data for developing supervised models is challenging. Contrastive learning (CL) has emerged as a promising pretraining approach that enables effective transfer learning with limited labelled data. However, existing CL frameworks either focus solely on global context or fail to exploit ECG-specific characteristics. Furthermore, these methods rely on hard contrastive targets, which may not adequately capture the continuous nature of feature similarity in ECG signals. In this paper, we propose Beat-SSL, a contrastive learning framework that performs dual-context learning through both rhythm-level and heartbeat-level contrasting with soft targets. We evaluated our pretrained model on two downstream tasks: 1) multilabel classification for global rhythm assessment, and 2) ECG segmentation to assess its capacity to learn representations across both contexts. We conducted an ablation study and compared the best configuration with three other methods, including one ECG foundation model. Despite the foundation model's broader pretraining, Beat-SSL reached 93% of its performance in multilabel classification task and surpassed all other methods in the segmentation task by 4%.
Authors: Mert Yuksekgonul, Daniel Koceja, Xinhao Li, Federico Bianchi, Jed McCaleb, Xiaolong Wang, Jan Kautz, Yejin Choi, James Zou, Carlos Guestrin, Yu Sun
Abstract: How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erd\H{o}s' minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to $2\times$ faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.
Authors: Song Xia, Meiwen Ding, Chenqi Kong, Wenhan Yang, Xudong Jiang
Abstract: Multimodal large language models (MLLMs) exhibit strong capabilities across diverse applications, yet remain vulnerable to adversarial perturbations that distort their feature representations and induce erroneous predictions. To address this vulnerability, we propose the Feature-space Smoothing (FS) and theoretically prove that FS offers certified robustness on the feature representations of MLLMs. Specifically, FS transforms any feature encoder into a smoothed variant that is guaranteed to maintain a certified lower bound on the feature cosine similarity between clean and adversarial representations under $\ell_2$-bounded attacks. Moreover, we indicate that the value of this Feature Cosine Similarity Bound (FCSB) derived from FS can be improved by enlarging the defined Gaussian robustness score on the vanilla encoder. Building upon this, we introduce the Purifier and Smoothness Mapper (PSM), a plug-and-play module that improves the Gaussian robustness score of MLLMs and thus enhances their certified robustness under FS, without requiring any retraining on MLLMs. We demonstrate that the FS with PSM not only provides a strong theoretical robustness guarantee but also exhibits superior empirical performance compared to adversarial training. Extensive experiments across diverse MLLMs and downstream tasks indicate the effectiveness of the FS-PSM, reducing the Attack Success Rate (ASR) of various white-box attacks from nearly 90\% to about 1\%.
Authors: Patrick Altmeyer, Aleksander Buszydlik, Arie van Deursen, Cynthia C. S. Liem
Abstract: We propose a novel training regime termed counterfactual training that leverages counterfactual explanations to increase the explanatory capacity of models. Counterfactual explanations have emerged as a popular post-hoc explanation method for opaque machine learning models: they inform how factual inputs would need to change in order for a model to produce some desired output. To be useful in real-world decision-making systems, counterfactuals should be plausible with respect to the underlying data and actionable with respect to the feature mutability constraints. Much existing research has therefore focused on developing post-hoc methods to generate counterfactuals that meet these desiderata. In this work, we instead hold models directly accountable for the desired end goal: counterfactual training employs counterfactuals during the training phase to minimize the divergence between learned representations and plausible, actionable explanations. We demonstrate empirically and theoretically that our proposed method facilitates training models that deliver inherently desirable counterfactual explanations and additionally exhibit improved adversarial robustness.
Authors: Sandeep Neela
Abstract: Market integrity monitoring is difficult because suspicious price/volume behavior can arise from many benign mechanisms, while modern detection systems often rely on opaque models that are hard to audit and communicate. We present AIMM-X, an explainable monitoring pipeline that combines market microstructure-style signals derived from OHLCV time series with multi-source public attention signals (e.g., news and online discussion proxies) to surface time windows that merit analyst review. The system detects candidate anomalous windows using transparent thresholding and aggregation, then assigns an interpretable integrity score decomposed into a small set of additive components, allowing practitioners to trace why a window was flagged and which factors drove the score. We provide an end-to-end, reproducible implementation that downloads data, constructs attention features, builds unified panels, detects windows, computes component signals, and generates summary figures/tables. Our goal is not to label manipulation, but to provide a practical, auditable screening tool that supports downstream investigation by compliance teams, exchanges, or researchers.
Authors: Rishit Chugh
Abstract: The deployment of large language models (LLMs) has raised security concerns due to their susceptibility to producing harmful or policy-violating outputs when exposed to adversarial prompts. While alignment and guardrails mitigate common misuse, they remain vulnerable to automated jailbreaking methods such as GCG, PEZ, and GBDA, which generate adversarial suffixes via training and gradient-based search. Although effective, these methods particularly GCG are computationally expensive, limiting their practicality for organisations with constrained resources. This paper introduces a resource-efficient adversarial prompting approach that eliminates the need for retraining by matching new prompts to a database of pre-trained adversarial prompts. A dataset of 1,000 prompts was classified into seven harm-related categories, and GCG, PEZ, and GBDA were evaluated on a Llama 3 8B model to identify the most effective attack method per category. Results reveal a correlation between prompt type and algorithm effectiveness. By retrieving semantically similar successful adversarial prompts, the proposed method achieves competitive attack success rates with significantly reduced computational cost. This work provides a practical framework for scalable red-teaming and security evaluation of aligned LLMs, including in settings where model internals are inaccessible.
Authors: Fabiana Taglietti, Andrea Pulici, Maxwell Roxburgh, Gabriele Seguini, Ian Vidamour, Stephan Menzel, Edoardo Franco, Michele Laus, Eleni Vasilaki, Michele Perego, Thomas J. Hayward, Marco Fanciulli, Jack C. Gartside
Abstract: Physical neural networks typically train linear synaptic weights while treating device nonlinearities as fixed. We show the opposite - by training the synaptic nonlinearity itself, as in Kolmogorov-Arnold Network (KAN) architectures, we yield markedly higher task performance per physical resource and improved performance-parameter scaling than conventional linear weight-based networks, demonstrating ability of KAN topologies to exploit reconfigurable nonlinear physical dynamics. We experimentally realise physical KANs in silicon-on-insulator devices we term 'Synaptic Nonlinear Elements' (SYNEs), operating at room temperature, 0.1-1 microampere currents, and 2 MHz speeds with no observed degradation over 10^13 measurements and months-long timescales. We demonstrate nonlinear function regression, classification, and prediction of Li-Ion battery dynamics from noisy real-world multi-sensor data. Physical KANs outperform equivalently-parameterised software multilayer perceptron networks across all tasks, with up to two orders of magnitude fewer parameters, and two orders of magnitude fewer devices than linear weight based physical networks. These results establish learned physical nonlinearity as a hardware-native computational primitive for compact and efficient learning systems, and SYNE devices as effective substrates for heterogenous nonlinear computing.
Authors: Chuanqing Wang, Zhenmin Zhao, Shanshan Du, Chaoqun Fei, Songmao Zhang, Ruqian Lu
Abstract: The rash development of knowledge graph research has brought big driving force to its application in many areas, including the medicine and healthcare domain. However, we have found that the application of some major information processing techniques on knowledge graph still lags behind. This defect includes the failure to make sufficient use of advanced logic reasoning, advanced artificial intelligence techniques, special-purpose programming languages, modern probabilistic and statistic theories et al. on knowledge graphs development and application. In particular, the multiple knowledge graphs cooperation and competition techniques have not got enough attention from researchers. This paper develops a systematic theory, technique and application of the concept 'knowledge graph network' and its application in medical and healthcare domain. Our research covers its definition, development, reasoning, computing and application under different conditions such as unsharp, uncertain, multi-modal, vectorized, distributed, federated. Almost in each case we provide (real data) examples and experiment results. Finally, a conclusion of innovation is provided.
Authors: Asim H. Gazi, Yongyi Guo, Daiqi Gao, Ziping Xu, Kelly W. Zhang, Susan A. Murphy
Abstract: Reinforcement learning (RL) has achieved remarkable success in real-world decision-making across diverse domains, including gaming, robotics, online advertising, public health, and natural language processing. Despite these advances, a substantial gap remains between RL research and its deployment in many practical settings. Two recurring challenges often underlie this gap. First, many settings offer limited opportunity for the agent to interact extensively with the target environment due to practical constraints. Second, many target environments often undergo substantial changes, requiring redesign and redeployment of RL systems (e.g., advancements in science and technology that change the landscape of healthcare delivery). Addressing these challenges and bridging the gap between basic research and application requires theory and methodology that directly inform the design, implementation, and continual improvement of RL systems in real-world settings. In this paper, we frame the application of RL in practice as a three-component process: (i) online learning and optimization during deployment, (ii) post- or between-deployment offline analyses, and (iii) repeated cycles of deployment and redeployment to continually improve the RL system. We provide a narrative review of recent advances in statistical RL that address these components, including methods for maximizing data utility for between-deployment inference, enhancing sample efficiency for online learning within-deployment, and designing sequences of deployments for continual improvement. We also outline future research directions in statistical RL that are use-inspired -- aiming for impactful application of RL in practice.
Authors: Eichi Uehara
Abstract: Estimating Heterogeneous Treatment Effects (HTE) in industrial applications such as AdTech and healthcare presents a dual challenge: extreme class imbalance and heavy-tailed outcome distributions. While the X-Learner framework effectively addresses imbalance through cross-imputation, we demonstrate that it is fundamentally vulnerable to "Outlier Smearing" when reliant on Mean Squared Error (MSE) minimization. In this failure mode, the bias from a few extreme observations ("whales") in the minority group is propagated to the entire majority group during the imputation step, corrupting the estimated treatment effect structure. To resolve this, we propose the Robust X-Learner (RX-Learner). This framework integrates a redescending {\gamma}-divergence objective -- structurally equivalent to the Welsch loss under Gaussian assumptions -- into the gradient boosting machinery. We further stabilize the non-convex optimization using a Proxy Hessian strategy grounded in Majorization-Minimization (MM) principles. Empirical evaluation on a semi-synthetic Criteo Uplift dataset demonstrates that the RX-Learner reduces the Precision in Estimation of Heterogeneous Effect (PEHE) metric by 98.6% compared to the standard X-Learner, effectively decoupling the stable "Core" population from the volatile "Periphery".
Authors: Hoshitaro Ohnishi, Hideo Mukai
Abstract: Quantum error correction is indispensable to achieving reliable quantum computation. When quantum information is encoded redundantly, a larger Hilbert space is constructed using multiple physical qubits, and the computation is performed within a designated subspace. When applying deep learning to the decoding of quantum error-correcting codes, a key challenge arises from the non-uniqueness between the syndrome measurements provided to the decoder and the corresponding error patterns that constitute the ground-truth labels. Building upon prior work that addressed this issue for the toric code by re-optimizing the decoder with respect to the symmetry inherent in the parity-check structure, we generalize this approach to arbitrary stabilizer codes. In our experiments, we employed multilayer perceptrons to approximate continuous functions that complement the syndrome measurements of the Color code and the Golay code. Using these models, we performed decoder re-optimization for each code. For the Color code, we achieved an improvement of approximately 0.8% in decoding accuracy at a physical error rate of 5%, while for the Golay code the accuracy increased by about 0.1%. Furthermore, from the evaluation of the geometric and algebraic structures in the continuous function approximation for each code, we showed that the design of generalized continuous functions is advantageous for learning the geometric structure inherent in the code. Our results also indicate that approximations that faithfully reproduce the code structure can have a significant impact on the effectiveness of reoptimization. This study demonstrates that the re-optimization technique previously shown to be effective for the Toric code can be generalized to address the challenge of label degeneracy that arises when applying deep learning to the decoding of stabilizer codes.
Authors: Jason Bohne, Ieva Petrulionyte, Michael Arbel, Julien Mairal, Pawe{\l} Polak
Abstract: Functional bilevel optimization (FBO) provides a powerful framework for hierarchical learning in function spaces, yet current methods are limited to static offline settings and perform suboptimally in online, non-stationary scenarios. We propose SmoothFBO, the first algorithm for non-stationary FBO with both theoretical guarantees and practical scalability. SmoothFBO introduces a time-smoothed stochastic hypergradient estimator that reduces variance through a window parameter, enabling stable outer-loop updates with sublinear regret. Importantly, the classical parametric bilevel case is a special reduction of our framework, making SmoothFBO a natural extension to online, non-stationary settings. Empirically, SmoothFBO consistently outperforms existing FBO methods in non-stationary hyperparameter optimization and model-based reinforcement learning, demonstrating its practical effectiveness. Together, these results establish SmoothFBO as a general, theoretically grounded, and practically viable foundation for bilevel optimization in online, non-stationary scenarios.
Authors: Francesca Pia Panaccione, Carlo Sgaravatti, Pietro Pinoli
Abstract: Biomedical research increasingly relies on integrating diverse data modalities, including gene expression profiles, medical images, and clinical metadata. While medical images and clinical metadata are routinely collected in clinical practice, gene expression data presents unique challenges for widespread research use, mainly due to stringent privacy regulations and costly laboratory experiments. To address these limitations, we present GeMM-GAN, a novel Generative Adversarial Network conditioned on histopathology tissue slides and clinical metadata, designed to synthesize realistic gene expression profiles. GeMM-GAN combines a Transformer Encoder for image patches with a final Cross Attention mechanism between patches and text tokens, producing a conditioning vector to guide a generative model in generating biologically coherent gene expression profiles. We evaluate our approach on the TCGA dataset and demonstrate that our framework outperforms standard generative models and generates more realistic and functionally meaningful gene expression profiles, improving by more than 11\% the accuracy on downstream disease type prediction compared to current state-of-the-art generative models. Code will be available at: https://github.com/francescapia/GeMM-GAN
Authors: Pablo Messina, Andr\'es Villa, Juan Le\'on Alc\'azar, Karen S\'anchez, Carlos Hinojosa, Denis Parra, \'Alvaro Soto, Bernard Ghanem
Abstract: Medical vision-language models can automate the generation of radiology reports but struggle with accurate visual grounding and factual consistency. Existing models often misalign textual findings with visual evidence, leading to unreliable or weakly grounded predictions. We present CURE, an error-aware curriculum learning framework that improves grounding and report quality without any additional data. CURE fine-tunes a multimodal instructional model on phrase grounding, grounded report generation, and anatomy-grounded report generation using public datasets. The method dynamically adjusts sampling based on model performance, emphasizing harder samples to improve spatial and textual alignment. CURE improves grounding accuracy by +0.37 IoU, boosts report quality by +0.188 CXRFEScore, and reduces hallucinations by 18.6%. CURE is a data-efficient framework that enhances both grounding accuracy and report reliability. Code is available at https://github.com/PabloMessina/CURE and model weights at https://huggingface.co/pamessina/medgemma-4b-it-cure
URLs: https://github.com/PabloMessina/CURE, https://huggingface.co/pamessina/medgemma-4b-it-cure
Authors: Alex Goessmann, Janina Sch\"utte, Maximilian Fr\"ohlich, Martin Eigel
Abstract: The unification of neural and symbolic approaches to artificial intelligence remains a central open challenge. In this work, we introduce a tensor network formalism, which captures sparsity principles originating in the different approaches in tensor decompositions. In particular, we describe a basis encoding scheme for functions and model neural decompositions as tensor decompositions. The proposed formalism can be applied to represent logical formulas and probability distributions as structured tensor decompositions. This unified treatment identifies tensor network contractions as a fundamental inference class and formulates efficiently scaling reasoning algorithms, originating from probability theory and propositional logic, as contraction message passing schemes. The framework enables the definition and training of hybrid logical and probabilistic models, which we call Hybrid Logic Network. The theoretical concepts are accompanied by the python library tnreason, which enables the implementation and practical use of the proposed architectures.
Authors: Saptarshi Roy, Alessandro Rinaldo, Purnamrita Sarkar
Abstract: In recent years, Rectified flow (RF) has gained considerable popularity largely due to its generation efficiency and state-of-the-art performance. In this paper, we investigate the degree to which RF automatically adapts to the intrinsic low dimensionality of the support of the target distribution to accelerate sampling. We show that, using a carefully designed choice of the time-discretization scheme and with sufficiently accurate drift estimates, the RF sampler enjoys an iteration complexity of order $O(k/\varepsilon)$ (up to log factors), where $\varepsilon$ is the precision in total variation distance and $k$ is the intrinsic dimension of the target distribution. In addition, we show that the denoising diffusion probabilistic model (DDPM) procedure is equivalent to a stochastic version of RF by establishing a novel connection between these processes and stochastic localization. Building on this connection, we further design a stochastic RF sampler that also adapts to the low-dimensionality of the target distribution under milder requirements on the accuracy of the drift estimates, and also with a specific time schedule. We illustrate with simulations on the synthetic data and text-to-image data experiments the improved performance of the proposed samplers implementing the newly designed time-discretization schedules.
Authors: Jason Chuan-Chih Chou, Abhinav Kumar, Shivank Garg
Abstract: Drawing inspiration from recent findings including surprisingly decent performance of transformers without positional encoding (NoPE) in the domain of language models and how registers (additional throwaway tokens not tied to input) may improve the performance of large vision transformers (ViTs), we invent and test a variant of ViT called fractal ViT that breaks permutation invariance among the tokens by applying an attention mask between the regular tokens and ``summary tokens'' similar to registers, in isolation or in combination with various positional encodings. These models do not improve upon ViT with registers, highlighting the fact that these findings may be scale, domain, or application-specific.
Authors: William Huang, Siyou Pei, Leyi Zou, Eric J. Gonzalez, Ishan Chatterjee, Yang Zhang
Abstract: The proliferation of XR devices has made egocentric hand pose estimation a vital task, yet this perspective is inherently challenged by frequent finger occlusions. To address this, we propose a novel approach that leverages the rich information in dorsal hand skin deformation, unlocked by recent advances in dense visual featurizers. We introduce a dual-stream delta encoder that learns pose by contrasting features from a dynamic hand with a baseline relaxed position. Our evaluation demonstrates that, using only cropped dorsal images, our method reduces the Mean Per Joint Angle Error (MPJAE) by 18% in self-occluded scenarios (fingers >=50% occluded) compared to state-of-the-art techniques that depend on the whole hand's geometry and large model backbones. Consequently, our method not only enhances the reliability of downstream tasks like index finger pinch and tap estimation in occluded scenarios but also unlocks new interaction paradigms, such as detecting isometric force for a surface "click" without visible movement while minimizing model size.
Authors: Wenxin Zhou, Ritesh Mehta, Anthony Miyaguchi
Abstract: We develop a two-stage retrieval system that combines multiple complementary retrieval methods with a learned reranker and LLM-based reranking, to address the TREC Tip-of-the-Tongue (ToT) task. In the first stage, we employ hybrid retrieval that merges LLM-based retrieval, sparse (BM25), and dense (BGE-M3) retrieval methods. We also introduce topic-aware multi-index dense retrieval that partitions the Wikipedia corpus into 24 topical domains. In the second stage, we evaluate both a trained LambdaMART reranker and LLM-based reranking. To support model training, we generate 5000 synthetic ToT queries using LLMs. Our best system achieves recall of 0.66 and NDCG@1000 of 0.41 on the test set by combining hybrid retrieval with Gemini-2.5-flash reranking, demonstrating the effectiveness of fusion retrieval.
Authors: Ali Khreis, Ro'Yah Radaideh, Quinn McGill
Abstract: Early detection of malignant skin lesions is critical for improving patient outcomes in aggressive, metastatic skin cancers. This study evaluates a comprehensive system for preliminary skin lesion assessment that combines the clinically established ABCD rule of dermoscopy (analyzing Asymmetry, Borders, Color, and Dermoscopic Structures) with machine learning classification. Using a 1,000-image subset of the HAM10000 dataset, the system implements an automated, rule-based pipeline to compute a Total Dermoscopy Score (TDS) for each lesion. This handcrafted approach is compared against various machine learning solutions, including traditional classifiers (Logistic Regression, Random Forest, and SVM) and deep learning models. While the rule-based system provides high clinical interpretability, results indicate a performance bottleneck when reducing complex morphology to five numerical features. Experimental findings show that transfer learning with EfficientNet-B0 failed significantly due to domain shift between natural and medical images. In contrast, a custom three-layer Convolutional Neural Network (CNN) trained from scratch achieved 78.5% accuracy and 86.5% recall on median-filtered images, representing a 19-point accuracy improvement over traditional methods. The results demonstrate that direct pixel-level learning captures diagnostic patterns beyond handcrafted features and that purpose-built lightweight architectures can outperform large pretrained models for small, domain-specific medical datasets.
Authors: Naoya Onizawa, Takahiro Hanyu
Abstract: This article critically investigates the limitations of the simulated annealing algorithm using probabilistic bits (pSA) in solving large-scale combinatorial optimization problems. The study begins with an in-depth analysis of the pSA process, focusing on the issues resulting from unexpected oscillations among p-bits. These oscillations hinder the energy reduction of the Ising model and thus obstruct the successful execution of pSA in complex tasks. Through detailed simulations, we unravel the root cause of this energy stagnation, identifying the feedback mechanism inherent to the pSA operation as the primary contributor to these disruptive oscillations. To address this challenge, we propose two novel algorithms, time average pSA (TApSA) and stalled pSA (SpSA). These algorithms are designed based on partial deactivation of p-bits and are thoroughly tested using Python simulations on maximum cut benchmarks that are typical combinatorial optimization problems. On the 16 benchmarks from 800 to 5,000 nodes, the proposed methods improve the normalized cut value from 0.8% to 98.4% on average in comparison with the conventional pSA.
Authors: Cyril Shih-Huan Hsu, Xi Li, Lanfranco Zanzi, Zhiheng Yang, Chrysa Papagianni, Xavier Costa P\'erez
Abstract: Recent advancements in mobile and wireless networks are unlocking the full potential of robotic autonomy, enabling robots to take advantage of ultra-low latency, high data throughput, and ubiquitous connectivity. However, for robots to navigate and operate seamlessly, efficiently and reliably, they must have an accurate understanding of both their surrounding environment and the quality of radio signals. Achieving this in highly dynamic and ever-changing environments remains a challenging and largely unsolved problem. In this paper, we introduce MapViT, a two-stage Vision Transformer (ViT)-based framework inspired by the success of pre-train and fine-tune paradigm for Large Language Models (LLMs). MapViT is designed to predict both environmental changes and expected radio signal quality. We evaluate the framework using a set of representative Machine Learning (ML) models, analyzing their respective strengths and limitations across different scenarios. Experimental results demonstrate that the proposed two-stage pipeline enables real-time prediction, with the ViT-based implementation achieving a strong balance between accuracy and computational efficiency. This makes MapViT a promising solution for energy- and resource-constrained platforms such as mobile robots. Moreover, the geometry foundation model derived from the self-supervised pre-training stage improves data efficiency and transferability, enabling effective downstream predictions even with limited labeled data. Overall, this work lays the foundation for next-generation digital twin ecosystems, and it paves the way for a new class of ML foundation models driving multi-modal intelligence in future 6G-enabled systems.
Authors: Yangyang Zhong, Yanmei Gu, Zhengqing Zang, Xiaomeng Li, Yuqi Ding, Xibei Jia, Yuting Shen, Zhenzhong Lan, Liwang Zhu, Weiping Liu, Junlin Zhou, Haisheng Liu, Zhong Xin Yu, Pengxin Luo, Donglian Qi, Yunfeng Yan, Junbo Zhao
Abstract: Masked Diffusion Language Models (MDLMs) promise parallel token generation and arbitrary-order decoding, yet it remains unclear to what extent current models truly realize these capabilities. We characterize MDLM behavior along two dimensions -- parallelism strength and generation order -- using Average Finalization Parallelism (AFP) and Kendall's tau. We evaluate eight mainstream MDLMs (up to 100B parameters) on 58 benchmarks spanning knowledge, reasoning, and programming. The results show that MDLMs still lag behind comparably sized autoregressive models, mainly because parallel probabilistic modeling weakens inter-token dependencies. Meanwhile, MDLMs exhibit adaptive decoding behavior: their parallelism and generation order vary significantly with the task domain, the stage of reasoning, and whether the output is correct. On tasks that require "backward information" (e.g., Sudoku), MDLMs adopt a solution order that tends to fill easier Sudoku blanks first, highlighting their advantages. Finally, we provide theoretical motivation and design insights supporting a Generate-then-Edit paradigm, which mitigates dependency loss while retaining the efficiency of parallel decoding.
Authors: Xinjie Zhou, Zhihui Yang, Lechao Cheng, Sai Wu, Gang Chen
Abstract: Large language models (LLMs) exhibit powerful capabilities but risk memorizing sensitive personally identifiable information (PII) from their training data, posing significant privacy concerns. While machine unlearning techniques aim to remove such data, they predominantly depend on access to the training data. This requirement is often impractical, as training data in real-world deployments is commonly proprietary or inaccessible. To address this limitation, we propose Data-Free Selective Unlearning (DFSU), a novel privacy-preserving framework that removes sensitive PII from an LLM without requiring its training data. Our approach first synthesizes pseudo-PII through language model inversion, then constructs token-level privacy masks for these synthetic samples, and finally performs token-level selective unlearning via a contrastive mask loss within a low-rank adaptation (LoRA) subspace. Extensive experiments on the AI4Privacy PII-Masking dataset using Pythia models demonstrate that our method effectively removes target PII while maintaining model utility.
Authors: Larry Bowden, Qi Chu, Bernard Cena, Kentaro Ohno, Bob Parney, Deepak Sharma, Mitsuharu Takeori
Abstract: Detecting machine failures promptly is of utmost importance in industry for maintaining efficiency and minimizing downtime. This paper introduces a failure detection algorithm based on quantum computing and a statistical change-point detection approach. Our method leverages the potential of projected quantum feature maps to enhance the precision of anomaly detection in machine monitoring systems. We empirically validate our approach on benchmark multi-dimensional time series datasets as well as on a real-world dataset comprising IoT sensor readings from operational machines, ensuring the practical relevance of our study. The algorithm was executed on IBM's 133-qubit Heron quantum processor, demonstrating the feasibility of integrating quantum computing into industrial maintenance procedures. The presented results underscore the effectiveness of our quantum-based failure detection system, showcasing its capability to accurately identify anomalies in noisy time series data. This work not only highlights the potential of quantum computing in industrial diagnostics but also paves the way for more sophisticated quantum algorithms in the realm of predictive maintenance.
Authors: Kristen Moore, Diksha Goel, Cody James Christopher, Zhen Wang, Minjune Kim, Ahmed Ibrahim, Ahmad Mohsin, Seyit Camtepe
Abstract: Realistic network traffic simulation is critical for evaluating intrusion detection systems, stress-testing network protocols, and constructing high-fidelity environments for cybersecurity training. While attack traffic can often be layered into training environments using red-teaming or replay methods, generating authentic benign background traffic remains a core challenge -- particularly in simulating the complex temporal and communication dynamics of real-world networks. This paper introduces TempoNet, a novel generative model that combines multi-task learning with multi-mark temporal point processes to jointly model inter-arrival times and all packet- and flow-header fields. TempoNet captures fine-grained timing patterns and higher-order correlations such as host-pair behavior and seasonal trends, addressing key limitations of GAN-, LLM-, and Bayesian-based methods that fail to reproduce structured temporal variation. TempoNet produces temporally consistent, high-fidelity traces, validated on real-world datasets. Furthermore, we show that intrusion detection models trained on TempoNet-generated background traffic perform comparably to those trained on real data, validating its utility for real-world security applications.
Authors: Hengfan Zhang, Yueqian Lin, Hai Helen Li, Yiran Chen
Abstract: Deploying Audio-Language Models (Audio-LLMs) on edge infrastructure exposes a persistent tension between perception depth and computational efficiency. Lightweight local models tend to produce passive perception - generic summaries that miss the subtle evidence required for multi-step audio reasoning - while indiscriminate cloud offloading incurs unacceptable latency, bandwidth cost, and privacy risk. We propose CoFi-Agent (Tool-Augmented Coarse-to-Fine Agent), a hybrid architecture targeting edge servers and gateways. It performs fast local perception and triggers conditional forensic refinement only when uncertainty is detected. CoFi-Agent runs an initial single-pass on a local 7B Audio-LLM, then a cloud controller gates difficult cases and issues lightweight plans for on-device tools such as temporal re-listening and local ASR. On the MMAR benchmark, CoFi-Agent improves accuracy from 27.20% to 53.60%, while achieving a better accuracy-efficiency trade-off than an always-on investigation pipeline. Overall, CoFi-Agent bridges the perception gap via tool-enabled, conditional edge-cloud collaboration under practical system constraints.
Authors: Mengyu Yao, Ziqi Zhang, Ning Luo, Shaofei Li, Yifeng Cai, Xiangqun Chen, Yao Guo, Ding Li
Abstract: Retrieval-augmented generation (RAG) systems integrate document retrieval with large language models and have been widely adopted. However, in privacy-related scenarios, RAG introduces a new privacy risk: adversaries can issue carefully crafted queries to exfiltrate sensitive content from the underlying corpus gradually. Although recent studies have demonstrated multi-turn extraction attacks, they rely on heuristics and fail to perform long-term extraction planning. To address these limitations, we formulate the RAG extraction attack as an adaptive stochastic coverage problem (ASCP). In ASCP, each query is treated as a probabilistic action that aims to maximize conditional marginal gain (CMG), enabling principled long-term planning under uncertainty. However, integrating ASCP with practical RAG attack faces three key challenges: unobservable CMG, intractability in the action space, and feasibility constraints. To overcome these challenges, we maintain a global attacker-side state to guide the attack. Building on this idea, we introduce RAGCRAWLER, which builds a knowledge graph to represent revealed information, uses this global state to estimate CMG, and plans queries in semantic space that target unretrieved regions. In comprehensive experiments across diverse RAG architectures and datasets, our proposed method, RAGCRAWLER, consistently outperforms all baselines. It achieves up to 84.4% corpus coverage within a fixed query budget and deliver an average improvement of 20.7% over the top-performing baseline. It also maintains high semantic fidelity and strong content reconstruction accuracy with low attack cost. Crucially, RAGCRAWLER proves its robustness by maintaining effectiveness against advanced RAG systems employing query rewriting and multi-query retrieval strategies. Our work reveals significant security gaps and highlights the pressing need for stronger safeguards for RAG.
Authors: Zhixuan Liang, Xingyu Zeng, Rui Zhao, Ping Luo
Abstract: Active learning (AL) strategies aim to train high-performance models with minimal labeling efforts, only selecting the most informative instances for annotation. Current approaches to evaluating data informativeness predominantly focus on the data's distribution or intrinsic information content and do not directly correlate with downstream task performance, such as mean average precision (mAP) in object detection. Thus, we propose Performance-guided (i.e. mAP-guided) Reinforced Active Learning for Object Detection (MGRAL), a novel approach that leverages the concept of expected model output changes as informativeness. To address the combinatorial explosion challenge of batch sample selection and the non-differentiable correlation between model performance and selected batches, MGRAL skillfully employs a reinforcement learning-based sampling agent that optimizes selection using policy gradient with mAP improvement as reward. Moreover, to reduce the computational overhead of mAP estimation with unlabeled samples, MGRAL utilizes an unsupervised way with fast look-up tables, ensuring feasible deployment. We evaluate MGRAL's active learning performance on detection tasks over PASCAL VOC and COCO benchmarks. Our approach demonstrates the highest AL curve with convincing visualizations, establishing a new paradigm in reinforcement learning-driven active object detection.
Authors: Binu V P, Deepthy K Bhaskar, Minimol B
Abstract: As digital threats continue to grow, organizations must find ways to enhance security while protecting user privacy. This paper explores how artificial intelligence (AI) plays a crucial role in achieving this balance. AI technologies can improve security by detecting threats, monitoring systems, and automating responses. However, using AI also raises privacy concerns that need careful consideration.We examine real-world examples from the healthcare sector to illustrate how organizations can implement AI solutions that strengthen security without compromising patient privacy. Additionally, we discuss the importance of creating transparent AI systems and adhering to privacy regulations.Ultimately, this paper provides insights and recommendations for integrating AI into healthcare security practices, helping organizations navigate the challenges of modern management while keeping patient data safe.
Authors: Asim Biswal, Chuan Lei, Xiao Qin, Aodong Li, Balakrishnan Narayanaswamy, Tim Kraska
Abstract: Recent advances in LLM-based Text-to-SQL have achieved remarkable gains on public benchmarks such as BIRD and Spider. Yet, these systems struggle to scale in realistic enterprise settings with large, complex schemas, diverse SQL dialects, and expensive multi-step reasoning. Emerging agentic approaches show potential for adaptive reasoning but often suffer from inefficiency and instability-repeating interactions with databases, producing inconsistent outputs, and occasionally failing to generate valid answers. To address these challenges, we introduce Agent Semantic Memory (AgentSM), an agentic framework for Text-to-SQL that builds and leverages interpretable semantic memory. Instead of relying on raw scratchpads or vector retrieval, AgentSM captures prior execution traces-or synthesizes curated ones-as structured programs that directly guide future reasoning. This design enables systematic reuse of reasoning paths, which allows agents to scale to larger schemas, more complex questions, and longer trajectories efficiently and reliably. Compared to state-of-the-art systems, AgentSM achieves higher efficiency by reducing average token usage and trajectory length by 25% and 35%, respectively, on the Spider 2.0 benchmark. It also improves execution accuracy, reaching a state-of-the-art accuracy of 44.8% on the Spider 2.0 Lite benchmark.
Authors: Jiahao Zhang, Zifan He, Nicholas Fraser, Michaela Blott, Yizhou Sun, Jason Cong
Abstract: We present FlexLLM, a composable High-Level Synthesis (HLS) library for rapid development of domain-specific LLM accelerators. FlexLLM exposes key architectural degrees of freedom for stage-customized inference, enabling hybrid designs that tailor temporal reuse and spatial dataflow differently for prefill and decode, and provides a comprehensive quantization suite to support accurate low-bit deployment. Using FlexLLM, we build a complete inference system for the Llama-3.2 1B model in under two months with only 1K lines of code. The system includes: (1) a stage-customized accelerator with hardware-efficient quantization (12.68 WikiText-2 PPL) surpassing SpinQuant baseline, and (2) a Hierarchical Memory Transformer (HMT) plug-in for efficient long-context processing. On the AMD U280 FPGA at 16nm, the accelerator achieves 1.29$\times$ end-to-end speedup, 1.64$\times$ higher decode throughput, and 3.14$\times$ better energy efficiency than an NVIDIA A100 GPU (7nm) running BF16 inference; projected results on the V80 FPGA at 7nm reach 4.71$\times$, 6.55$\times$, and 4.13$\times$, respectively. In long-context scenarios, integrating the HMT plug-in reduces prefill latency by 23.23$\times$ and extends the context window by 64$\times$, delivering 1.10$\times$/4.86$\times$ lower end-to-end latency and 5.21$\times$/6.27$\times$ higher energy efficiency on the U280/V80 compared to the A100 baseline. FlexLLM thus bridges algorithmic innovation in LLM inference and high-performance accelerators with minimal manual effort.
Authors: Ajvad Haneef K, Karan Kuwar Singh, Madhu Kumar S D
Abstract: High-dimensional malware datasets often exhibit feature redundancy, instability, and scalability limitations, which hinder the effectiveness and interpretability of machine learning-based malware detection systems. Although feature selection is commonly employed to mitigate these issues, many existing approaches lack robustness when applied to large-scale and heterogeneous malware data. To address this gap, this paper proposes CAFE-GB (Chunk-wise Aggregated Feature Estimation using Gradient Boosting), a scalable feature selection framework designed to produce stable and globally consistent feature rankings for high-dimensional malware detection. CAFE-GB partitions training data into overlapping chunks, estimates local feature importance using gradient boosting models, and aggregates these estimates to derive a robust global ranking. Feature budget selection is performed separately through a systematic k-selection and stability analysis to balance detection performance and robustness. The proposed framework is evaluated on two large-scale malware datasets: BODMAS and CIC-AndMal2020, representing large and diverse malware feature spaces. Experimental results show that classifiers trained on CAFE-GB -selected features achieve performance parity with full-feature baselines across multiple metrics, including Accuracy, F1-score, MCC, ROC-AUC, and PR-AUC, while reducing feature dimensionality by more than 95\%. Paired Wilcoxon signed-rank tests confirm that this reduction does not introduce statistically significant performance degradation. Additional analyses demonstrate low inter-feature redundancy and improved interpretability through SHAP-based explanations. Runtime and memory profiling further indicate reduced downstream classification overhead. Overall, CAFE-GB provides a stable, interpretable, and scalable feature selection strategy for large-scale malware detection.
Authors: Qi Zeng, Weide Liu, Bo Li, Ryne Didier, P. Ellen Grant, Davood Karimi
Abstract: This paper presents FeTal-SAM, a novel adaptation of the Segment Anything Model (SAM) tailored for fetal brain MRI segmentation. Traditional deep learning methods often require large annotated datasets for a fixed set of labels, making them inflexible when clinical or research needs change. By integrating atlas-based prompts and foundation-model principles, FeTal-SAM addresses two key limitations in fetal brain MRI segmentation: (1) the need to retrain models for varying label definitions, and (2) the lack of insight into whether segmentations are driven by genuine image contrast or by learned spatial priors. We leverage multi-atlas registration to generate spatially aligned label templates that serve as dense prompts, alongside a bounding-box prompt, for SAM's segmentation decoder. This strategy enables binary segmentation on a per-structure basis, which is subsequently fused to reconstruct the full 3D segmentation volumes. Evaluations on two datasets, the dHCP dataset and an in-house dataset demonstrate FeTal-SAM's robust performance across gestational ages. Notably, it achieves Dice scores comparable to state-of-the-art baselines which were trained for each dataset and label definition for well-contrasted structures like cortical plate and cerebellum, while maintaining the flexibility to segment any user-specified anatomy. Although slightly lower accuracy is observed for subtle, low-contrast structures (e.g., hippocampus, amygdala), our results highlight FeTal-SAM's potential to serve as a general-purpose segmentation model without exhaustive retraining. This method thus constitutes a promising step toward clinically adaptable fetal brain MRI analysis tools.
Authors: Clare Chemery, Hendrik Edelhoff, Ludwig Bothmann
Abstract: We introduce a lightweight experimentation pipeline designed to lower the barrier for applying machine learning (ML) methods for classifying images in ecological research. We enable ecologists to experiment with ML models independently, thus they can move beyond off-the-shelf models and generate insights tailored to local datasets and specific classification tasks and target variables. Our tool combines a simple command-line interface for preprocessing, training, and evaluation with a graphical interface for annotation, error analysis, and model comparison. This design enables ecologists to build and iterate on compact, task-specific classifiers without requiring advanced ML expertise. As a proof of concept, we apply the pipeline to classify red deer (Cervus elaphus) by age and sex from 3392 camera trap images collected in the Veldenstein Forest, Germany. Using 4352 cropped images containing individual deer labeled by experts, we trained and evaluated multiple backbone architectures with a wide variety of parameters and data augmentation strategies. Our best-performing models achieved 90.77% accuracy for age classification and 96.15% for sex classification. These results demonstrate that reliable demographic classification is feasible even with limited data to answer narrow, well-defined ecological problems. More broadly, the framework provides ecologists with an accessible tool for developing ML models tailored to specific research questions, paving the way for broader adoption of ML in wildlife monitoring and demographic analysis.
Authors: Jaya Chaturvedi, Saniya Deshpande, Chenkai Ma, Robert Cobb, Angus Roberts, Robert Stewart, Daniel Stahl, Diana Shamsutdinova
Abstract: Introduction: Clinical text classification using natural language processing (NLP) models requires adequate training data to achieve optimal performance. For that, 200-500 documents are typically annotated. The number is constrained by time and costs and lacks justification of the sample size requirements and their relationship to text vocabulary properties. Methods: Using the publicly available MIMIC-III dataset containing hospital discharge notes with ICD-9 diagnoses as labels, we employed pre-trained BERT embeddings followed by Random Forest classifiers to identify 10 randomly selected diagnoses, varying training corpus sizes from 100 to 10,000 documents, and analyzed vocabulary properties by identifying strong and noisy predictive words through Lasso logistic regression on bag-of-words embeddings. Results: Learning curves varied significantly across the 10 classification tasks despite identical preprocessing and algorithms, with 600 documents sufficient to achieve 95% of the performance attainable with 10,000 documents for all tasks. Vocabulary analysis revealed that more strong predictors and fewer noisy predictors were associated with steeper learning curves, where every 100 additional noisy words decreased accuracy by approximately 0.02 while 100 additional strong predictors increased maximum accuracy by approximately 0.04.
Authors: Jingsong Xia, Siqi Wang
Abstract: Background: Coronary angiography (CAG) is a cornerstone imaging modality for assessing coronary artery disease and guiding interventional treatment decisions. However, in real-world clinical settings, angiographic images are often characterized by complex lesion morphology, severe class imbalance, label uncertainty, and limited computational resources, posing substantial challenges to conventional deep learning approaches in terms of robustness and generalization.Methods: The proposed framework is built upon a pretrained convolutional neural network to construct a lightweight hybrid neural representation. A selective neural plasticity training strategy is introduced to enable efficient parameter adaptation. Furthermore, a brain-inspired attention-modulated loss function, combining Focal Loss with label smoothing, is employed to enhance sensitivity to hard samples and uncertain annotations. Class-imbalance-aware sampling and cosine annealing with warm restarts are adopted to mimic rhythmic regulation and attention allocation mechanisms observed in biological neural systems.Results: Experimental results demonstrate that the proposed lightweight brain-inspired model achieves strong and stable performance in binary coronary angiography classification, yielding competitive accuracy, recall, F1-score, and AUC metrics while maintaining high computational efficiency.Conclusion: This study validates the effectiveness of brain-inspired learning mechanisms in lightweight medical image analysis and provides a biologically plausible and deployable solution for intelligent clinical decision support under limited computational resources.
Authors: Jaekwon Im, Natalia Polouliakh, Taketo Akama
Abstract: Dance-to-music generation aims to generate music that is aligned with dance movements. Existing approaches typically rely on body motion features extracted from a single human dancer and limited dance-to-music datasets, which restrict their performance and applicability to real-world scenarios involving multiple dancers and non-human dancers. In this paper, we propose PF-D2M, a universal diffusion-based dance-to-music generation model that incorporates visual features extracted from dance videos. PF-D2M is trained with a progressive training strategy that effectively addresses data scarcity and generalization challenges. Both objective and subjective evaluations show that PF-D2M achieves state-of-the-art performance in dance-music alignment and music quality.
Authors: Chen Xu
Abstract: We propose a novel first-order method for non-convex optimization of the form $\max_{\bm{w}\in\mathbb{R}^d}\mathbb{E}_{\bm{x}\sim\mathcal{D}}[f_{\bm{w}}(\bm{x})]$, termed Progressive Power Homotopy (Prog-PowerHP). The method applies stochastic gradient ascent to a surrogate objective obtained by first performing a power transformation and then Gaussian smoothing, $F_{N,\sigma}(\bm{\mu}):=\mathbb{E}_{\bm{w}\sim\mathcal{N}(\bm{\mu},\sigma^2I_d),\bm{x}\sim\mathcal{D}}[e^{Nf_w(\bm{x})}]$, while progressively increasing the power parameter $N$ and decreasing the smoothing scale $\sigma$ along the optimization trajectory. We prove that, under mild regularity conditions, Prog-PowerHP converges to a small neighborhood of the global optimum with an iteration complexity scaling nearly as $O(d^2\varepsilon^{-2})$. Empirically, Prog-PowerHP demonstrates clear advantages in phase retrieval when the samples-to-dimension ratio approaches the information-theoretic limit, and in training two-layer neural networks in under-parameterized regimes. These results suggest that Prog-PowerHP is particularly effective for navigating cluttered non-convex landscapes where standard first-order methods struggle.
Authors: Brainard Philemon Jagati, Jitendra Tembhurne, Harsh Goud, Rudra Pratap Singh, Chandrashekhar Meshram
Abstract: Deep neural network models degrade significantly in the long-tailed data distribution, with the overall training data dominated by a small set of classes in the head, and the tail classes obtaining less training examples. Addressing the imbalance in the classes, attention in the related literature was given mainly to the adjustments carried out in the decision space in terms of either corrections performed at the logit level in order to compensate class-prior bias, with the least attention to the optimization process resulting from the adjustments introduced through the differences in the confidences among the samples. In the current study, we present the design of a class and confidence-aware re-weighting scheme for long-tailed learning. This scheme is purely based upon the loss level and has a complementary nature to the existing methods performing the adjustment of the logits. In the practical implementation stage of the proposed scheme, we use an {\Omega}(p_t, f_c) function. This function enables the modulation of the contribution towards the training task based upon the confidence value of the prediction, as well as the relative frequency of the corresponding class. Our observations in the experiments are corroborated by significant experimental results performed on the CIFAR-100-LT, ImageNet-LT, and iNaturalist2018 datasets under various values of imbalance factors that clearly authenticate the theoretical discussions above.
Authors: Xiangyu Wang, Zhixin Lv, Yongjiao Sun, Anrui Han, Ye Yuan, Hangxu Ji
Abstract: Text-Based Person Search (TBPS) holds unique value in real-world surveillance bridging visual perception and language understanding, yet current paradigms utilizing pre-training models often fail to transfer effectively to complex open-world scenarios. The reliance on "Passive Observation" leads to multifaceted spurious correlations and spatial semantic misalignment, causing a lack of robustness against distribution shifts. To fundamentally resolve these defects, this paper proposes ICON (Invariant Counterfactual Optimization with Neuro-symbolic priors), a framework integrating causal and topological priors. First, we introduce Rule-Guided Spatial Intervention to strictly penalize sensitivity to bounding box noise, forcibly severing location shortcuts to achieve geometric invariance. Second, Counterfactual Context Disentanglement is implemented via semantic-driven background transplantation, compelling the model to ignore background interference for environmental independence. Then, we employ Saliency-Driven Semantic Regularization with adaptive masking to resolve local saliency bias and guarantee holistic completeness. Finally, Neuro-Symbolic Topological Alignment utilizes neuro-symbolic priors to constrain feature matching, ensuring activated regions are topologically consistent with human structural logic. Experimental results demonstrate that ICON not only maintains leading performance on standard benchmarks but also exhibits exceptional robustness against occlusion, background interference, and localization noise. This approach effectively advances the field by shifting from fitting statistical co-occurrences to learning causal invariance.
Authors: Liang Wang, Kanzhong Yao, Yang Liu, Weikai Qin, Jun Wu, Zhe Sun, Qiuguo Zhu
Abstract: Parkour tasks for quadrupeds have emerged as a promising benchmark for agile locomotion. While human athletes can effectively perceive environmental characteristics to select appropriate footholds for obstacle traversal, endowing legged robots with similar perceptual reasoning remains a significant challenge. Existing methods often rely on hierarchical controllers that follow pre-computed footholds, thereby constraining the robot's real-time adaptability and the exploratory potential of reinforcement learning. To overcome these challenges, we present PUMA, an end-to-end learning framework that integrates visual perception and foothold priors into a single-stage training process. This approach leverages terrain features to estimate egocentric polar foothold priors, composed of relative distance and heading, guiding the robot in active posture adaptation for parkour tasks. Extensive experiments conducted in simulation and real-world environments across various discrete complex terrains, demonstrate PUMA's exceptional agility and robustness in challenging scenarios.
Authors: Yifan Zhu, Yekai Pan, Chen Ding
Abstract: High-performance attention kernels are essential for Large Language Models. This paper presents analysis of CuTile-based Flash Attention memory behavior and a technique to improve its cache performance. In particular, our analysis on the NVIDIA GB10 (Grace Blackwell) identifies the main cause of L2 cache miss. Leveraging this insight, we introduce a new programming technique called Sawtooth Wavefront Reordering that reduces L2 misses. We validate it in both CUDA and CuTile, observing 50\% or greater reduction in L2 misses and up to 60\% increase in throughput on GB10.
Authors: Omar Al-Ghattas
Abstract: In constrained stochastic optimization, one naturally expects that imposing a stricter feasible set does not increase the statistical risk of an estimator defined by projection onto that set. In this paper, we show that this intuition can fail even in canonical settings. We study the Gaussian sequence model, a deliberately austere test best, where for a compact, convex set $\Theta \subset \mathbb{R}^d$ one observes \[ Y = \theta^\star + \sigma Z, \qquad Z \sim N(0, I_d), \] and seeks to estimate an unknown parameter $\theta^\star \in \Theta$. The natural estimator is the least squares estimator (LSE), which coincides with the Euclidean projection of $Y$ onto $\Theta$. We construct an explicit example exhibiting \emph{risk reversal}: for sufficiently large noise, there exist nested compact convex sets $\Theta_S \subset \Theta_L$ and a parameter $\theta^\star \in \Theta_S$ such that the LSE constrained to $\Theta_S$ has strictly larger risk than the LSE constrained to $\Theta_L$. We further show that this phenomenon can persist at the level of worst-case risk, with the supremum risk over the smaller constraint set exceeding that over the larger one. We clarify this behavior by contrasting noise regimes. In the vanishing-noise limit, the risk admits a first-order expansion governed by the statistical dimension of the tangent cone at $\theta^\star$, and tighter constraints uniformly reduce risk. In contrast, in the diverging-noise regime, the risk is determined by global geometric interactions between the constraint set and random noise directions. Here, the embedding of $\Theta_S$ within $\Theta_L$ can reverse the risk ordering. These results reveal a previously unrecognized failure mode of projection-based estimators: in sufficiently noisy settings, tightening a constraint can paradoxically degrade statistical performance.
Authors: Jingfu Peng, Yuhong Yang
Abstract: Deep neural networks (DNNs) typically involve a large number of parameters and are trained to achieve zero or near-zero training error. Despite such interpolation, they often exhibit strong generalization performance on unseen data, a phenomenon that has motivated extensive theoretical investigations. Comforting results show that interpolation indeed may not affect the minimax rate of convergence under the squared error loss. In the mean time, DNNs are well known to be highly vulnerable to adversarial perturbations in future inputs. A natural question then arises: Can interpolation also escape from suboptimal performance under a future $X$-attack? In this paper, we investigate the adversarial robustness of interpolating estimators in a framework of nonparametric regression. A finding is that interpolating estimators must be suboptimal even under a subtle future $X$-attack, and achieving perfect fitting can substantially damage their robustness. An interesting phenomenon in the high interpolation regime, which we term the curse of simple size, is also revealed and discussed. Numerical experiments support our theoretical findings.
Authors: Saar Cohen
Abstract: Clustering is a fundamental problem, aiming to partition a set of elements, like agents or data points, into clusters such that elements in the same cluster are closer to each other than to those in other clusters. In this paper, we present a new framework for studying online non-centroid clustering with delays, where elements, that arrive one at a time as points in a finite metric space, should be assigned to clusters, but assignments need not be immediate. Specifically, upon arrival, each point's location is revealed, and an online algorithm has to irrevocably assign it to an existing cluster or create a new one containing, at this moment, only this point. However, we allow decisions to be postponed at a delay cost, instead of following the more common assumption of immediate decisions upon arrival. This poses a critical challenge: the goal is to minimize both the total distance costs between points in each cluster and the overall delay costs incurred by postponing assignments. In the classic worst-case arrival model, where points arrive in an arbitrary order, no algorithm has a competitive ratio better than sublogarithmic in the number of points. To overcome this strong impossibility, we focus on a stochastic arrival model, where points' locations are drawn independently across time from an unknown and fixed probability distribution over the finite metric space. We offer hope for beyond worst-case adversaries: we devise an algorithm that is constant competitive in the sense that, as the number of points grows, the ratio between the expected overall costs of the output clustering and an optimal offline clustering is bounded by a constant.
Authors: Zack Dewis, Yimin Zhu, Zhengsen Xu, Mabel Heffring, Saeid Taleghanidoozdoozan, Quinn Ledingham, Lincoln Linlin Xu
Abstract: Although Mamba models greatly improve Hyperspectral Image (HSI) classification, they have critical challenges in terms defining efficient and adaptive token sequences for improve performance. This paper therefore presents CSSMamba (Clustering-guided Spatial-Spectral Mamba) framework to better address the challenges, with the following contributions. First, to achieve efficient and adaptive token sequences for improved Mamba performance, we integrate the clustering mechanism into a spatial Mamba architecture, leading to a cluster-guided spatial Mamba module (CSpaMamba) that reduces the Mamba sequence length and improves Mamba feature learning capability. Second, to improve the learning of both spatial and spectral information, we integrate the CSpaMamba module with a spectral mamba module (SpeMamba), leading to a complete clustering-guided spatial-spectral Mamba framework. Third, to further improve feature learning capability, we introduce an Attention-Driven Token Selection mechanism to optimize Mamba token sequencing. Last, to seamlessly integrate clustering into the Mamba model in a coherent manner, we design a Learnable Clustering Module that learns the cluster memberships in an adaptive manner. Experiments on the Pavia University, Indian Pines, and Liao-Ning 01 datasets demonstrate that CSSMamba achieves higher accuracy and better boundary preservation compared to state-of-the-art CNN, Transformer, and Mamba-based methods.
Authors: Zhengchi Ma, Anru R. Zhang
Abstract: Imbalanced classification, where one class is observed far less frequently than the other, often causes standard training procedures to prioritize the majority class and perform poorly on rare but important cases. A classic and widely used remedy is to augment the minority class with synthetic examples, but two basic questions remain under-resolved: when does synthetic augmentation actually help, and how many synthetic samples should be generated? We develop a unified statistical framework for synthetic augmentation in imbalanced learning, studying models trained on imbalanced data augmented with synthetic minority samples and evaluated under the balanced population risk. Our theory shows that synthetic data is not always beneficial. In a ``local symmetry" regime, imbalance is not the dominant source of error near the balanced optimum, so adding synthetic samples cannot improve learning rates and can even degrade performance by amplifying generator mismatch. When augmentation can help (a ``local asymmetry" regime), the optimal synthetic size depends on generator accuracy and on whether the generator's residual mismatch is directionally aligned with the intrinsic majority-minority shift. This structure can make the best synthetic size deviate from naive full balancing, sometimes by a small refinement and sometimes substantially when generator bias is systematic. Practically, we recommend Validation-Tuned Synthetic Size (VTSS): select the synthetic size by minimizing balanced validation loss over a range centered near the fully balanced baseline, while allowing meaningful departures when the data indicate them. Simulations and a real sepsis prediction study support the theory and illustrate when synthetic augmentation helps, when it cannot, and how to tune its quantity effectively.
Authors: Zainab Alhathloul, Irfan Ahmad
Abstract: The Arabic language has undergone notable transformations over time, including the emergence of new vocabulary, the obsolescence of others, and shifts in word usage. This evolution is evident in the distinction between the classical and modern Arabic eras. Although historians and linguists have partitioned Arabic literature into multiple eras, relatively little research has explored the automatic classification of Arabic texts by time period, particularly beyond the domain of poetry. This paper addresses this gap by employing neural networks and deep learning techniques to automatically classify Arabic texts into distinct eras and periods. The proposed models are evaluated using two datasets derived from two publicly available corpora, covering texts from the pre-Islamic to the modern era. The study examines class setups ranging from binary to 15-class classification and considers both predefined historical eras and custom periodizations. Results range from F1-scores of 0.83 and 0.79 on the binary-era classification task using the OpenITI and APCD datasets, respectively, to 0.20 on the 15-era classification task using OpenITI and 0.18 on the 12-era classification task using APCD.
Authors: Paolo Baldan, Sebastian Gurke, Barbara K\"onig, Florian Wittbold
Abstract: The problem of determining the (least) fixpoint of (higher-dimensional) functions over the non-negative reals frequently occurs when dealing with systems endowed with a quantitative semantics. We focus on the situation in which the functions of interest are not known precisely but can only be approximated. As a first contribution we generalize an iteration scheme called dampened Mann iteration, recently introduced in the literature. The improved scheme relaxes previous constraints on parameter sequences, allowing learning rates to converge to zero or not converge at all. While seemingly minor, this flexibility is essential to enable the implementation of chaotic iterations, where only a subset of components is updated in each step, allowing to tackle higher-dimensional problems. Additionally, by allowing learning rates to converge to zero, we can relax conditions on the convergence speed of function approximations, making the method more adaptable to various scenarios. We also show that dampened Mann iteration applies immediately to compute the expected payoff in various probabilistic models, including simple stochastic games, not covered by previous work.
Authors: Prakash Dhungana, Sayed Ahmad Salehi
Abstract: Keyword Spotting (KWS) systems with small footprint models deployed on edge devices face significant accuracy and robustness challenges due to domain shifts caused by varying noise and recording conditions. To address this, we propose a comprehensive framework for continual learning designed to adapt to new domains while maintaining computational efficiency. The proposed pipeline integrates a dual-input Convolutional Neural Network, utilizing both Mel Frequency Cepstral Coefficients (MFCC) and Mel-spectrogram features, supported by a multi-stage denoising process, involving discrete wavelet transform and spectral subtraction techniques, plus model and prototype update blocks. Unlike prior methods that restrict updates to specific layers, our approach updates the complete quantized model, made possible due to compact model architecture. A subset of input samples are selected during runtime using class prototypes and confidence-driven filtering, which are then pseudo-labeled and combined with rehearsal buffer for incremental model retraining. Experimental results on noisy test dataset demonstrate the framework's effectiveness, achieving 99.63\% accuracy on clean data and maintaining robust performance (exceeding 94\% accuracy) across diverse noisy environments, even at -10 dB Signal-to-Noise Ratio. The proposed framework work confirms that integrating efficient denoising with prototype-based continual learning enables KWS models to operate autonomously and robustly in resource-constrained, dynamic environments.
Authors: Yiyao Yang
Abstract: Uncertainty estimation in machine learning has traditionally focused on the prediction stage, aiming to quantify confidence in model outputs while treating learned representations as deterministic and reliable by default. In this work, we challenge this implicit assumption and argue that reliability should be regarded as a first-class property of learned representations themselves. We propose a principled framework for reliable representation learning that explicitly models representation-level uncertainty and leverages structural constraints as inductive biases to regularize the space of feasible representations. Our approach introduces uncertainty-aware regularization directly in the representation space, encouraging representations that are not only predictive but also stable, well-calibrated, and robust to noise and structural perturbations. Structural constraints, such as sparsity, relational structure, or feature-group dependencies, are incorporated to define meaningful geometry and reduce spurious variability in learned representations, without assuming fully correct or noise-free structure. Importantly, the proposed framework is independent of specific model architectures and can be integrated with a wide range of representation learning methods.
Authors: El Mehdi Er Raqabi, Kevin Dalmeijer, Pascal Van Hentenryck
Abstract: This paper investigates the multi-compartment vehicle routing problem with multiple time windows (MCVRPMTW), an extension of the classical vehicle routing problem with time windows that considers vehicles equipped with multiple compartments and customers requiring service across several delivery time windows. The problem incorporates three key compartment-related features: (i) compartment flexibility in the number of compartments, (ii) item-to-compartment compatibility, and (iii) item-to-item compatibility. The problem also accommodates practical operational requirements such as driver breaks. To solve the MCVRPMTW, we develop an exact branch-and-price (B&P) algorithm in which the pricing problem is solved using a labeling algorithm. Several acceleration strategies are introduced to limit symmetry during label extensions, improve the stability of dual solutions in column generation, and enhance the branching process. To handle large-scale instances, we propose a rolling-space B&P algorithm that integrates clustering techniques into the solution framework. Extensive computational experiments on instances inspired by a real-world industrial application demonstrate the effectiveness of the proposed approach and provide useful managerial insights for practical implementation.
Authors: Ofir Nabati, Guy Tennenholtz, Shie Mannor
Abstract: We present a representation-driven framework for reinforcement learning. By representing policies as estimates of their expected values, we leverage techniques from contextual bandits to guide exploration and exploitation. Particularly, embedding a policy network into a linear feature space allows us to reframe the exploration-exploitation problem as a representation-exploitation problem, where good policy representations enable optimal exploration. We demonstrate the effectiveness of this framework through its application to evolutionary and policy gradient-based approaches, leading to significantly improved performance compared to traditional methods. Our framework provides a new perspective on reinforcement learning, highlighting the importance of policy representation in determining optimal exploration-exploitation strategies.
Authors: Chakib Fettal, Lazhar Labiod, Mohamed Nadif
Abstract: The proliferation of high-dimensional data from sources such as social media, sensor networks, and online platforms has created new challenges for clustering algorithms. Multi-view clustering, which integrates complementary information from multiple data perspectives, has emerged as a powerful solution. However, existing methods often struggle with scalability and efficiency, particularly on large attributed networks. In this work, we address these limitations by leveraging explicit kernel feature maps and a non-iterative optimization strategy, enabling efficient and accurate clustering on datasets with millions of points.
Authors: Steven Kolawole, Lucio Dery, Jean-Fran\c{c}ois Kagy, Virginia Smith, Graham Neubig, Ameet Talwalkar
Abstract: Structured pruning is a promising approach to create smaller, faster large language models. However, existing methods typically rely on computing the gradient via backward passes, which can inflate memory requirements and compute costs. In this work we introduce Bonsai, a gradient-free structured pruning method that eliminates the need for backpropagation, significantly reducing memory requirements and compute costs while achieving state-of-the-art pruning performance. Bonsai uses forward-pass-only perturbative pruning to enable efficient compression of large models on a broader range of hardware configurations. Unlike existing structured pruning approaches, Bonsai not only achieves better compression with fewer resources but also produces models that are twice as fast as those generated by semi-structured pruning. As a concrete demonstration, we use Bonsai to prune 7B and 8B models to 50% sparsity on a single A6000 GPU -- a task challenging for backprop-based methods in memory-constrained settings, as they require 2-3x the memory. Our results show that removing backprop as a requirement not only enables pruning larger models on constrained hardware but can also lead to state-of-the-art efficiency and performance.
Authors: Hugo Melchers, Joost Prins, Michael Abdelmalik
Abstract: This work introduces a paradigm for constructing parametric neural operators that are derived from finite-dimensional representations of Green's operators for linear partial differential equations (PDEs). We refer to such neural operators as Neural Green's Operators (NGOs). Our construction of NGOs preserves the linear action of Green's operators on the inhomogeneity fields, while approximating the nonlinear dependence of the Green's function on the coefficients of the PDE using neural networks. This construction reduces the complexity of the problem from learning the entire solution operator and its dependence on all parameters to only learning the Green's function and its dependence on the PDE coefficients. Furthermore, we show that our explicit representation of Green's functions enables the embedding of desirable mathematical attributes in our NGO architectures, such as symmetry, spectral, and conservation properties. Through numerical benchmarks on canonical PDEs, we demonstrate that NGOs achieve comparable or superior accuracy to Deep Operator Networks, Variationally Mimetic Operator Networks, and Fourier Neural Operators with similar parameter counts, while generalizing significantly better when tested on out-of-distribution data. For parametric time-dependent PDEs, we show that NGOs that are trained on a single time step can produce pointwise-accurate dynamics in an auto-regressive manner over arbitrarily large numbers of time steps. For parametric nonlinear PDEs, we demonstrate that NGOs trained exclusively on solutions of corresponding linear problems can be embedded within iterative solvers to yield accurate solutions, provided a suitable initial guess is available. Finally, we show that we can leverage the explicit representation of Green's functions returned by NGOs to construct effective matrix preconditioners that accelerate iterative solvers for PDEs.
Authors: Zhirui Chen, Vincent Y. F. Tan
Abstract: We consider the problem of offline reinforcement learning from human feedback (RLHF) with pairwise comparisons proposed by Zhu et al. (2023), where the implicit reward is a linear function of an unknown parameter. Given an offline dataset, our objective consists in ascertaining the optimal action for each state, with the ultimate goal of minimizing the {\em simple regret}. We propose an algorithm, \underline{RL} with \underline{L}ocally \underline{O}ptimal \underline{W}eights or {\sc RL-LOW}, which yields an exponential form of simple regret of $\exp ( - \Omega(n/H) )$ where $n$ is the number of data samples and $H$ denotes an instance-dependent hardness quantity that depends explicitly on the suboptimality gap of each action. Furthermore, we derive a first-of-its-kind instance-dependent lower bound in offline RLHF with pairwise comparisons. Interestingly, we observe that the lower and upper bounds on the simple regret match order-wise in the exponent, demonstrating order-wise optimality of our {\sc RL-LOW}. In view of privacy considerations in practical applications, we also extend {\sc RL-LOW} to the setting of $(\varepsilon,\delta)$-differential privacy and show, somewhat surprisingly, that the hardness parameter $H$ is unchanged in the asymptotic regime as $n$ tends to infinity; this underscores the inherent efficiency of {\sc RL-LOW} in terms of preserving the privacy of the observed rewards. Given our focus on establishing instance-dependent bounds of exponential convergence, our research fills the research gap in existing studies that concentrate on establishing worst-case regrets of {\em inverse polynomial convergence} (e.g., $\widetilde{O}(\frac{1}{\sqrt{n}})$) for offline RLHF with pairwise comparisons.
Authors: Da Li, Junping Yin, Jin Xu, Xinxin Li, Juan Zhang
Abstract: Extracting simple mathematical expression from an observational dataset to describe complex natural phenomena is one of the core objectives of artificial intelligence (AI). This field is known as symbolic regression (SR). Traditional SR models are based on genetic programming (GP) or reinforcement learning (RL), facing well-known challenges, such as low efficiency and overfitting. Recent studies have integrated SR with large language models (LLMs), enabling fast zero-shot inference by learning mappings from millions of dataset-expression pairs. However, since the input and output are inherently different modalities, such models often struggle to converge effectively. In this paper, we introduce ViSymRe, a vision-guided multimodal SR model that incorporates the third resource, expression graph, to bridge the modality gap. Different from traditional multimodal models, ViSymRe is trained to extract vision, termed virtual vision, from datasets, without relying on the global availability of expression graphs, which addresses the essential challenge of visual SR, i.e., expression graphs are not available during inference. Evaluation results on multiple mainstream benchmarks show that ViSymRe achieves more competitive performance than the state-of-the-art dataset-only baselines. The expressions predicted by ViSymRe not only fit the dataset well but are also simple and structurally accurate, goals that SR models strive to achieve.
Authors: Eric Hirsch, Christian Friedrich
Abstract: Accurate tool wear prediction is essential for maintaining productivity and minimizing costs in machining. However, the complex nature of the tool wear process poses significant challenges to achieving reliable predictions. This study explores data-driven methods, in particular deep learning, for tool wear prediction. Traditional data-driven approaches often focus on a single process, relying on multi-sensor setups and extensive data generation, which limits generalization to new settings. Moreover, multi-sensor integration is often impractical in industrial environments. To address these limitations, this research investigates the transferability of predictive models using minimal training data, validated across two processes. Furthermore, it uses a simple setup with a single acceleration sensor to establish a low-cost data generation approach that facilitates the generalization of models to other processes via transfer learning. The study evaluates several machine learning models, including transformer-inspired convolutional neural networks (CNN), long short-term memory networks (LSTM), support vector machines (SVM), and decision trees, trained on different input formats such as feature vectors and short-time Fourier transform (STFT). The performance of the models is evaluated on two machines and on different amounts of training data, including scenarios with significantly reduced datasets, providing insight into their effectiveness under constrained data conditions. The results demonstrate the potential of specific models and configurations for effective tool wear prediction, contributing to the development of more adaptable and efficient predictive maintenance strategies in machining. Notably, the ConvNeXt model has an exceptional performance, achieving 99.1\% accuracy in identifying tool wear using data from only four milling tools operated until they are worn.
Authors: Pablo Barcel\'o, Alexander Kozachinskiy, Miguel Romero Orth, Bernardo Subercaseaux, Jos\'e Verschae
Abstract: Despite the wide use of $k$-Nearest Neighbors as classification models, their explainability properties remain poorly understood from a theoretical perspective. While nearest neighbors classifiers offer interpretability from a ``data perspective'', in which the classification of an input vector $\bar{x}$ is explained by identifying the vectors $\bar{v}_1, \ldots, \bar{v}_k$ in the training set that determine the classification of $\bar{x}$, we argue that such explanations can be impractical in high-dimensional applications, where each vector has hundreds or thousands of features and it is not clear what their relative importance is. Hence, we focus on understanding nearest neighbor classifications through a ``feature perspective'', in which the goal is to identify how the values of the features in $\bar{x}$ affect its classification. Concretely, we study abductive explanations such as ``minimum sufficient reasons'', which correspond to sets of features in $\bar{x}$ that are enough to guarantee its classification, and counterfactual explanations based on the minimum distance feature changes one would have to perform in $\bar{x}$ to change its classification. We present a detailed landscape of positive and negative complexity results for counterfactual and abductive explanations, distinguishing between discrete and continuous feature spaces, and considering the impact of the choice of distance function involved. Finally, we show that despite some negative complexity results, Integer Quadratic Programming and SAT solving allow for computing explanations in practice.
Authors: Phil Ostheimer, Mayank Nagda, Andriy Balinskyy, Jean Radig, Carl Herrmann, Stephan Mandt, Marius Kloft, Sophie Fellenz
Abstract: Sparse data is fundamental to scientific simulations in biology and physics, from single-cell gene expression to particle calorimetry, where exact zeros encode physical absence rather than weak signal. However, existing diffusion models lack the physical rigor to faithfully represent this sparsity. This work introduces Sparse Data Diffusion (SDD), a generative method that explicitly models exact zeros via Sparsity Bits, unifying efficient ML generation with physically grounded sparsity handling. Empirical validation in particle physics and single-cell biology demonstrates that SDD achieves higher fidelity than baseline methods in capturing sparse patterns critical for scientific analysis, advancing scalable and physically faithful simulation.
Authors: Quentin Nater, Mourad Khayati
Abstract: With the prevalence of sensor failures, imputation, the process of estimating missing values, has emerged as the cornerstone of time series data pre-processing. While numerous imputation algorithms have been developed to repair these data gaps, existing time series libraries provide limited imputation support. Furthermore, they often lack the ability to simulate realistic time series missingness patterns and fail to account for the impact of the imputed data on subsequent downstream analysis. This paper introduces ImputeGAP, a comprehensive library for time series imputation that supports a diverse range of imputation methods and modular missing data simulation, catering to datasets with varying characteristics. The library includes extensive customization options, such as automated hyperparameter tuning, benchmarking, explainability, downstream evaluation, and compatibility with popular time series frameworks.
Authors: Vugar Ismailov
Abstract: We study feedforward neural networks with inputs from a topological space (TFNNs). We prove a universal approximation theorem for shallow TFNNs, which demonstrates their capacity to approximate any continuous function defined on this topological space. As an application, we obtain an approximative version of Kolmogorov's superposition theorem for compact metric spaces.
Authors: Wensheng Li, Yichao Tian, Hao Wang, Ruifeng Zhou, Hanting Guan, Chao Zhang, Dacheng Tao
Abstract: Curriculum learning (CL) mimics human learning, in which easy samples are learned first, followed by harder samples, and has become an effective method for training deep networks. However, many existing automatic CL methods maintain a preference for easy samples during the entire training process regardless of the constantly evolving training state. This is just like a human curriculum that fails to provide individualized instruction, which can delay learning progress. To address this issue, we propose an adaptively point-weighting (APW) curriculum learning method that assigns a weight to each training sample based on its training loss. The weighting strategy of APW follows the easy-to-hard training paradigm, guided by the current training state of the network. We present a theoretical analysis of APW, including training effectiveness, training stability, and generalization performance. Experimental results validate these theoretical findings and demonstrate the superiority of the proposed APW method.
Authors: Alejandro Velez-Arce, Jesus Caraballo, Marinka Zitnik
Abstract: Existing biomedical benchmarks do not provide end-to-end infrastructure for training, evaluation, and inference of models that integrate multimodal biological data and a broad range of machine learning tasks in therapeutics. We present PyTDC, an open-source machine-learning platform providing streamlined training, evaluation, and inference software for multimodal biological AI models. PyTDC unifies distributed, heterogeneous, continuously updated data sources and model weights and standardizes benchmarking and inference endpoints. This paper discusses the components of PyTDC's architecture and, to our knowledge, the first-of-its-kind case study on the introduced single-cell drug-target nomination ML task. We find state-of-the-art methods in graph representation learning and domain-specific methods from graph theory perform poorly on this task. Though we find a context-aware geometric deep learning method that outperforms the evaluated SoTA and domain-specific baseline methods, the model is unable to generalize to unseen cell types or incorporate additional modalities, highlighting PyTDC's capacity to facilitate an exciting avenue of research developing multimodal, context-aware, foundation models for open problems in biomedical AI.
Authors: Shutong Ding, Ke Hu, Shan Zhong, Haoyang Luo, Weinan Zhang, Jingya Wang, Jun Wang, Ye Shi
Abstract: Recent advances in reinforcement learning (RL) have demonstrated the powerful exploration capabilities and multimodality of generative diffusion-based policies. While substantial progress has been made in offline RL and off-policy RL settings, integrating diffusion policies into on-policy frameworks like PPO remains underexplored. This gap is particularly significant given the widespread use of large-scale parallel GPU-accelerated simulators, such as IsaacLab, which are optimized for on-policy RL algorithms and enable rapid training of complex robotic tasks. A key challenge lies in computing state-action log-likelihoods under diffusion policies, which is straightforward for Gaussian policies but intractable for flow-based models due to irreversible forward-reverse processes and discretization errors (e.g., Euler-Maruyama approximations). To bridge this gap, we propose GenPO, a generative policy optimization framework that leverages exact diffusion inversion to construct invertible action mappings. GenPO introduces a novel doubled dummy action mechanism that enables invertibility via alternating updates, resolving log-likelihood computation barriers. Furthermore, we also use the action log-likelihood for unbiased entropy and KL divergence estimation, enabling KL-adaptive learning rates and entropy regularization in on-policy updates. Extensive experiments on eight IsaacLab benchmarks, including legged locomotion (Ant, Humanoid, Anymal-D, Unitree H1, Go2), dexterous manipulation (Shadow Hand), aerial control (Quadcopter), and robotic arm tasks (Franka), demonstrate GenPO's superiority over existing RL baselines. Notably, GenPO is the first method to successfully integrate diffusion policies into on-policy RL, unlocking their potential for large-scale parallelized training and real-world robotic deployment.
Authors: Han Liu, Keyan Ding, Peilin Chen, Yinwei Wei, Liqiang Nie, Dapeng Wu, Shiqi Wang
Abstract: Accurate prediction of protein-ligand binding affinity is critical for drug discovery. While recent deep learning approaches have demonstrated promising results, they often rely solely on structural features of proteins and ligands, overlooking their valuable biochemical knowledge associated with binding affinity. To address this limitation, we propose KEPLA, a novel deep learning framework that explicitly integrates prior knowledge from Gene Ontology and ligand properties to enhance prediction performance. KEPLA takes protein sequences and ligand molecular graphs as input and optimizes two complementary objectives: (1) aligning global representations with knowledge graph relations to capture domain-specific biochemical insights, and (2) leveraging cross attention between local representations to construct fine-grained joint embeddings for prediction. Experiments on two benchmark datasets across both in-domain and cross-domain scenarios demonstrate that KEPLA consistently outperforms state-of-the-art baselines. Furthermore, interpretability analyses based on knowledge graph relations and cross attention maps provide valuable insights into the underlying predictive mechanisms.
Authors: Mohammad Hashemi, Hossein Amiri, Andreas Zufle
Abstract: Learning effective representations of urban environments requires capturing spatial structure beyond fixed administrative boundaries. Existing geospatial representation learning approaches typically aggregate Points of Interest(POI) into pre-defined administrative regions such as census units or ZIP code areas, assigning a single embedding to each region. However, POIs often form semantically meaningful groups that extend across, within, or beyond these boundaries, defining places that better reflect human activity and urban function. To address this limitation, we propose PlaceRep, a training-free geospatial representation learning method that constructs place-level representations by clustering spatially and semantically related POIs. PlaceRep summarizes large-scale POI graphs from U.S. Foursquare data to produce general-purpose urban region embeddings while automatically identifying places across multiple spatial scales. By eliminating model pre-training, PlaceRep provides a scalable and efficient solution for multi-granular geospatial analysis. Experiments using the tasks of population density estimation and housing price prediction as downstream tasks show that PlaceRep outperforms most state-of-the-art graph-based geospatial representation learning methods and achieves up to a 100x speedup in generating region-level representations on large-scale POI graphs. The implementation of PlaceRep is available at https://github.com/mohammadhashemii/PlaceRep.
Authors: Mario Tuci, Lennart Bastian, Benjamin Dupuis, Nassir Navab, Tolga Birdal, Umut \c{S}im\c{s}ekli
Abstract: Providing generalization guarantees for stochastic optimization algorithms remains a key challenge in learning theory. Recently, numerous works demonstrated the impact of the geometric properties of optimization trajectories on generalization performance. These works propose worst-case generalization bounds in terms of various notions of intrinsic dimension and/or topological complexity, which were found to empirically correlate with the generalization error. However, most of these approaches involve intractable mutual information terms, which limit a full understanding of the bounds. In contrast, some authors built on algorithmic stability to obtain worst-case bounds involving geometric quantities of a combinatorial nature, which are impractical to compute. In this paper, we address these limitations by combining empirically relevant complexity measures with a framework that avoids intractable quantities. To this end, we introduce the concept of \emph{random set stability}, tailored for the data-dependent random sets produced by stochastic optimization algorithms. Within this framework, we show that the worst-case generalization error can be bounded in terms of (i) the random set stability parameter and (ii) empirically relevant, data- and algorithm-dependent complexity measures of the random set. Moreover, our framework improves existing topological generalization bounds by recovering previous complexity notions without relying on mutual information terms. Through a series of experiments in practically relevant settings, we validate our theory by evaluating the tightness of our bounds and the interplay between topological complexity and stability.
Authors: Haowei Lin, Haotian Ye, Wenzheng Feng, Quzhe Huang, Yujun Li, Hubert Lim, Zhengrui Li, Xiangyu Wang, Jianzhu Ma, Yitao Liang, James Zou
Abstract: Discovering scaling laws for predicting model performance at scale is a fundamental and open-ended challenge, mostly reliant on slow, case specific human experimentation. To investigate the potential for LLMs to automate this process, we collect over 5,000 experiments from existing literature and curate eight diverse scaling law discovery tasks. While existing agents struggle to produce accurate law formulas, this paper introduces SLDAgent, an evolution-based agent that co-optimize the scaling law model and the parameters, enabling it to autonomously explore complex relationships between variables. For the first time, we demonstrates that SLDAgent can automatically discover laws that exhibit consistently more accurate extrapolation than their established, human-derived counterparts across all tasks. Through comprehensive analysis, we elucidate why these discovered laws are superior and verify their practical utility in both pretraining and finetuning applications. This work establishes a new paradigm for agentic scientific discovery, showing that AI systems can understand their own scaling behavior, and can contribute novel and practical knowledge back to the research community.
Authors: Ye Su, Hezhe Qiao, Wei Huang, Lin Chen
Abstract: Semi-supervised regression (SSR), which aims to predict continuous scores of samples while reducing reliance on a large amount of labeled data, has recently received considerable attention across various applications, including computer vision, natural language processing, and audio and medical analysis. Existing SSR methods typically train models on scarce labeled data by introducing constraint-based regularization or ordinal ranking to reduce overfitting. However, these approaches fail to fully exploit the abundance of unlabeled samples. While consistency-driven pseudo-labeling methods attempt to incorporate unlabeled data, they are highly sensitive to pseudo-label quality and noisy predictions. To address these challenges, we introduce a Dual-stream Knowledge Distillation framework (DKD), which is specially designed for the SSR task to distill knowledge from both continuous-valued knowledge and distribution information, better preserving regression magnitude information and improving sample efficiency. Specifically, in DKD, the teacher is optimized solely with ground-truth labels for label distribution estimation, while the student learns from a mixture of real labels and teacher-generated pseudo targets on unlabeled data. The distillation design ensures the effective supervision transfer, allowing the student to leverage pseudo labels more robustly. Then, we introduce an advanced Decoupled Distribution Alignment (DDA) to align the target class and non-target class between teacher and student on the distribution, enhancing the student's capacity to mitigate noise in pseudo-label supervision and learn a more well-calibrated regression predictor.
Authors: Zihao Wang, Yunjie Li, Lingmin Zan, Zheng Gong, Mengtao Zhu
Abstract: The Extended Long Short-Term Memory (xLSTM) network has demonstrated strong capability in modeling complex long-term dependencies in time series data. Despite its success, the deterministic architecture of xLSTM limits its representational capacity and forecasting performance, especially on challenging real-world time series datasets characterized by inherent uncertainty, stochasticity, and complex hierarchical latent dynamics. In this work, we propose StoxLSTM, a stochastic xLSTM within a designed state space modeling framework, which integrates latent stochastic variables directly into the recurrent units to effectively model deep latent temporal dynamics and uncertainty. The designed state space model follows an efficient non-autoregressive generative approach, achieving strong predictive performance without complex modifications to the original xLSTM architecture. Extensive experiments on publicly available benchmark datasets demonstrate that StoxLSTM consistently outperforms state-of-the-art baselines, achieving superior performance and generalization.
Authors: Zhanting Zhou, KaHou Tam, Yiding Feng, Ziqiang Zheng, Zeyu Ma, Yang Yang
Abstract: Federated Graph Learning (FGL) enables a central server to coordinate model training across distributed clients without local graph data being shared. However, FGL significantly suffers from cross-silo domain shifts, where each "silo" (domain) contains a limited number of clients with distinct graph topologies. These heterogeneities induce divergent optimization trajectories, ultimately leading to global model divergence. In this work, we reveal a severe architectural pathology termed Structural Orthogonality: the topology-dependent message passing mechanism forces gradients from different domains to target disjoint coordinates in the parameter space. Through a controlled comparison between backbones, we statistically prove that GNN updates are near-perpendicular across domains (with projection ratios $\to$ 0). Consequently, naive averaging leads to Consensus Collapse, a phenomenon where sparse, informative structural signals from individual domains are diluted by the near-zero updates of others. This forces the global model into a "sub-optimal" state that fails to represent domain-specific structural patterns, resulting in poor generalization. To address this, we propose FedIA, a lightweight server-side framework designed to reconcile update conflicts without auxiliary communication. FedIA operates in two stages: (1) Global Importance Masking (GIM) identifies a shared parameter subspace to filter out domain-specific structural noise and prevent signal dilution; (2) Confidence-Aware Momentum Weighting (CAM) dynamically re-weights client contributions based on gradient reliability to amplify stable optimization signals.
Authors: Niek Tax, Lorenzo Perini, Fridolin Linder, Daniel Haimovich, Dima Karamshuk, Nastaran Okati, Milan Vojnovic, Pavlos Athanasios Apostolopoulos
Abstract: We propose MCGrad, a novel and scalable multicalibration algorithm. Multicalibration - calibration in subgroups of the data - is an important property for the performance of machine learning-based systems. Existing multicalibration methods have thus far received limited traction in industry. We argue that this is because existing methods (1) require such subgroups to be manually specified, which ML practitioners often struggle with, (2) are not scalable, or (3) may harm other notions of model performance such as log loss and Area Under the Precision-Recall Curve (PRAUC). MCGrad does not require explicit specification of protected groups, is scalable, and often improves other ML evaluation metrics instead of harming them. MCGrad has been in production at Meta, and is now part of hundreds of production models. We present results from these deployments as well as results on public datasets. We provide an open source implementation of MCGrad at https://github.com/facebookincubator/MCGrad.
Authors: Yoontae Hwang, Stefan Zohren
Abstract: Modern deep learning for asset allocation typically separates forecasting from optimization. We argue this creates a fundamental mismatch where minimizing prediction errors fails to yield robust portfolios. We propose the Signature Informed Transformer to address this by unifying feature extraction and decision making into a single policy. Our model employs path signatures to encode complex path dependencies and introduces a specialized attention mechanism that targets geometric asset relationships. By directly minimizing the Conditional Value at Risk we ensure the training objective aligns with financial goals. We prove that our attention module rigorously amplifies signature derived signals. Experiments across diverse equity universes show our approach significantly outperforms both traditional strategies and advanced forecasting baselines. The code is available at: https://anonymous.4open.science/r/Signature-Informed-Transformer-For-Asset-Allocation-DB88
URLs: https://anonymous.4open.science/r/Signature-Informed-Transformer-For-Asset-Allocation-DB88
Authors: Yorgos Felekis, Theodoros Damoulas, Paris Giampouras
Abstract: Causal Abstraction (CA) theory provides a principled framework for relating causal models that describe the same system at different levels of granularity while ensuring interventional consistency between them. Recent methods for learning CAs, however, assume fixed and well-specified exogenous distributions, leaving them vulnerable to environmental shifts and model misspecification. In this work, we address these limitations by introducing the first class of distributionally robust CAs and their associated learning algorithms. The latter cast robust causal abstraction learning as a constrained min-max optimization problem with Wasserstein ambiguity sets. We provide theoretical guarantees for both empirical and Gaussian environments, enabling principled selection of ambiguity set radii and establish quantitative guarantees on worst-case abstraction error. Furthermore, we present empirical evidence across different problems and CA learning methods, demonstrating our framework's robustness not only to environmental shifts but also to structural and intervention mapping misspecification.
Authors: Shaocong Ma, Heng Huang
Abstract: In financial applications, reinforcement learning (RL) agents are commonly trained on historical data, where their actions do not influence prices. However, during deployment, these agents trade in live markets where their own transactions can shift asset prices, a phenomenon known as market impact. This mismatch between training and deployment environments can significantly degrade performance. Traditional robust RL approaches address this model misspecification by optimizing the worst-case performance over a set of uncertainties, but typically rely on symmetric structures that fail to capture the directional nature of market impact. To address this issue, we develop a novel class of elliptic uncertainty sets. We establish both implicit and explicit closed-form solutions for the worst-case uncertainty under these sets, enabling efficient and tractable robust policy evaluation. Experiments on single-asset and multi-asset trading tasks demonstrate that our method achieves superior Sharpe ratio and remains robust under increasing trade volumes, offering a more faithful and scalable approach to RL in financial markets.
Authors: Ai Jian, Jingqing Ruan, Xing Ma, Dailin Li, Weipeng Zhang, Ke Zeng, Xunliang Cai
Abstract: Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences. Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations. To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM). Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a Task-Adaptive Rubric system that dynamically generates instance-specific criteria for precise evaluation. Extensive experiments demonstrate that PATRM achieves a 8.7% average improvement on RewardBench and RMBench across Qwen3-8B/14B models. Crucially, it boosts downstream RLHF performance by an average relative improvement of 13.6% across IFEval and InFoBench, validating its effectiveness for policy alignment. Our code is available at https://github.com/JaneEyre0530/PaTaRM.
Authors: Georgiy Shakirov, Albert Arakelov
Abstract: A common practice in heterogeneous graph neural networks (HGNNs) is to condition parameters on node/edge types, assuming types reflect semantic roles. However, this can cause overreliance on surface-level labels and impede cross-type knowledge transfer. We explore integrating Mixture-of-Experts (MoE) into HGNNs--a direction underexplored despite MoE's success in homogeneous settings. Crucially, we question the need for type-specific experts. We propose Homogeneous Expert Routing (HER), an MoE layer for Heterogeneous Graph Transformers (HGT) that stochastically masks type embeddings during routing to encourage type-agnostic specialization. Evaluated on IMDB, ACM, and DBLP for link prediction, HER consistently outperforms standard HGT and a type-separated MoE baseline. Analysis on IMDB shows HER experts specialize by semantic patterns (e.g., movie genres) rather than node types, confirming routing is driven by latent semantics. Our work demonstrates that regularizing type dependence in expert routing yields more generalizable, efficient, and interpretable representations--a new design principle for heterogeneous graph learning.
Authors: Yushen Liu, Yanfu Zhang, Xugui Zhou
Abstract: Automated insulin delivery for Type 1 Diabetes must balance glucose control and safety under uncertain meals and physiological variability. While reinforcement learning (RL) enables adaptive personalization, existing approaches struggle to simultaneously guarantee safety, leaving a gap in achieving both personalized and risk-aware glucose control, such as overdosing before meals or stacking corrections. To bridge this gap, we propose TSODE, a safety-aware controller that integrates Thompson Sampling RL with a Neural Ordinary Differential Equation (NeuralODE) forecaster to address this challenge. Specifically, the NeuralODE predicts short-term glucose trajectories conditioned on proposed insulin doses, while a conformal calibration layer quantifies predictive uncertainty to reject or scale risky actions. In the FDA-approved UVa/Padova simulator (adult cohort), TSODE achieved 87.9% time-in-range with less than 10% time below 70 mg/dL, outperforming relevant baselines. These results demonstrate that integrating adaptive RL with calibrated NeuralODE forecasting enables interpretable, safe, and robust glucose regulation.
Authors: Qinghao Hu, Shang Yang, Junxian Guo, Xiaozhe Yao, Yujun Lin, Yuxian Gu, Han Cai, Chuang Gan, Ana Klimovic, Song Han
Abstract: The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.
Authors: Yanxuan Yu, Michael S. Hughes, Julien Lee, Jiacheng Zhou, Andrew F. Laine
Abstract: We study classification under extreme class imbalance where recall and calibration are both critical, for example in medical diagnosis scenarios. We propose AF-SMOTE, a mathematically motivated augmentation framework that first synthesizes minority points and then filters them by an adversarial discriminator and a boundary utility model. We prove that, under mild assumptions on the decision boundary smoothness and class-conditional densities, our filtering step monotonically improves a surrogate of F_beta (for beta >= 1) while not inflating Brier score. On MIMIC-IV proxy label prediction and canonical fraud detection benchmarks, AF-SMOTE attains higher recall and average precision than strong oversampling baselines (SMOTE, ADASYN, Borderline-SMOTE, SVM-SMOTE), and yields the best calibration. We further validate these gains across multiple additional datasets beyond MIMIC-IV. Our successful application of AF-SMOTE to a healthcare dataset using a proxy label demonstrates in a disease-agnostic way its practical value in clinical situations, where missing true positive cases in rare diseases can have severe consequences.
Authors: Duncan Stothers, Ben Stothers, Emily Schaeffer, Kishore Mulpuri
Abstract: We study an ultrasound-first, radiation-preserving policy for developmental dysplasia of the hip (DDH) that requests a radiograph only when needed. We (i) pretrain modality-specific encoders (ResNet-18) with SimSiam on a large unlabelled registry (37186 ultrasound; 19546 radiographs), (ii) freeze the backbones and fit small, measurement-faithful heads on DDH-relevant landmarks and measurements, (iii) calibrate a one-sided conformal deferral rule on ultrasound predictions that provides finite sample marginal coverage guarantees under exchangeability, using a held-out calibration set. Ultrasound heads predict Graf alpha, beta, and femoral head coverage; X-ray heads predict acetabular index (AI), center-edge (CE) angle and IHDI grade. On our held out labeled evaluation set, ultrasound measurement error is modest (e.g., alpha MAE ~= 9.7 degrees, coverage MAE ~= 14.0%), while radiographic probes achieve AI and CE MAEs of ~= 7.6 degrees and ~= 8.9 degrees, respectively. The calibrated US-only policy is explored across rule families (alpha-only; alpha OR coverage; alpha AND coverage), conformal miscoverage levels, and per-utility trade-offs using decision-curve analysis. Conservative settings yield high coverage with near-zero US-only rates; permissive settings (e.g., alpha OR coverage at larger deltas) achieve non-zero US-only throughput with expected coverage tradeoffs. The result is a simple, reproducible pipeline that turns limited labels into interpretable measurements and tunable selective imaging curves suitable for clinical handoff and future external validation.
Authors: Songlin Zhao, Michael Pitts, Zhuwei Qin
Abstract: Large language models (LLMs) are increasingly adapted into domain-specific variants for applications in law, healthcare, and finance. Their scale, however, limits deployment in resource-constrained settings, and existing compression approaches often either degrade after domain adaptation or require substantial additional computation. We introduce EfficientXpert, a lightweight framework for domain pruning that integrates ForeSight Mask, a propagation-aware criterion for selecting weights to prune without backpropagation, and Partial Brain Surgeon, an efficient closed-form update for low-rank adapters under a fixed sparsity pattern. With fine-tuning cost comparable to standard LoRA, EfficientXpert converts a general pretrained model into a sparse, domain-adapted expert in a single pruning step. Across health and legal benchmarks, EfficientXpert reaches up to 98 percent of dense performance at 40 percent sparsity, improving over prior pruning baselines while matching LoRA training time and staying within 1 percent of LoRA peak GPU memory in our experiments.
Authors: Zhiguo Zhang, Xiaoliang Ma, Daniel Schlesinger
Abstract: Accurate and interpretable air pollution forecasting is crucial for public health, but most models face a trade-off between performance and interpretability. This study proposes a physics-guided, interpretable-by-design spatiotemporal learning framework. The model decomposes the spatiotemporal behavior of air pollutant concentrations into two transparent, additive modules. The first is a physics-guided transport kernel with directed weights conditioned on wind and geography (advection). The second is an explainable attention mechanism that learns local responses and attributes future concentrations to specific historical lags and exogenous drivers. Evaluated on a comprehensive dataset from the Stockholm region, our model consistently outperforms state-of-the-art baselines across multiple forecasting horizons. Our model's integration of high predictive performance and spatiotemporal interpretability provides a more reliable foundation for operational air-quality management in real-world applications.
Authors: Libo Wang
Abstract: To address a fundamental limitation in cognitive systems, namely the absence of a time-updatable mediating thought space between semantics and continuous control, this work constructs and trains a vision-language-action model termed Sigma, deployed on a single RTX 4090. The model is built upon the open-source pi0.5_base backbone, with the svla_so101_pickplace dataset preprocessed into a structured training corpus. An independently designed VLA architecture is introduced to integrate deep semantic understanding with associative reasoning, enabling telepathic-style alignment between perception and action. Training proceeds through iterative optimization of data preprocessing, LoRA-based fine-tuning, and inference-stage adapter design. Evaluation is conducted using offline closed-loop replay, comparing Sigma against the untuned pi0.5_base under identical data conditions. Experimental results indicate a consistent reduction in control MSE across vector-, fragment-, and trajectory-level scales, while preserving the stability of the telepathy norm and semantic-text alignment quality. These findings demonstrate that mind-responsive alignment control can be quantitatively achieved through semantic and associative architectural integration without retraining the base model, providing a reproducible pathway for semantic alignment and intention-driven behavior.
Authors: Huzaifa Arif
Abstract: Recent work \cite{arifgroup} introduced Federated Proximal Gradient \textbf{(\texttt{FedProxGrad})} for solving non-convex composite optimization problems in group fair federated learning. However, the original analysis established convergence only to a \textit{noise-dominated neighborhood of stationarity}, with explicit dependence on a variance-induced noise floor. In this work, we provide an improved asymptotic convergence analysis for a generalized \texttt{FedProxGrad}-type analytical framework with inexact local proximal solutions and explicit fairness regularization. We call this extended analytical framework \textbf{DS \texttt{FedProxGrad}} (Decay Step Size \texttt{FedProxGrad}). Under a Robbins-Monro step-size schedule \cite{robbins1951stochastic} and a mild decay condition on local inexactness, we prove that $\liminf_{r\to\infty} \mathbb{E}[\|\nabla F(\mathbf{x}^r)\|^2] = 0$, i.e., the algorithm is asymptotically stationary and the convergence rate does not depend on a variance-induced noise floor.
Authors: Nicolas Tacheny
Abstract: Agentic systems built on large language models operate through recursive feedback loops, where each output becomes the next input. Yet the geometric behavior of these agentic loops (whether they converge, diverge, or exhibit more complex dynamics) remains poorly understood. This paper introduces a geometric framework for analyzing agentic trajectories in semantic embedding space, treating iterative transformations as discrete dynamical systems. We distinguish the artifact space, where linguistic transformations occur, from the embedding space, where geometric measurements are performed. Because cosine similarity is biased by embedding anisotropy, we introduce an isotonic calibration that eliminates systematic bias and aligns similarities with human semantic judgments while preserving high local stability. This enables rigorous measurement of trajectories, clusters and attractors. Through controlled experiments on singular agentic loops, we identify two fundamental regimes. A contractive rewriting loop converges toward a stable attractor with decreasing dispersion, while an exploratory summarize and negate loop produces unbounded divergence with no cluster formation. These regimes display qualitatively distinct geometric signatures of contraction and expansion. Our results show that prompt design directly governs the dynamical regime of an agentic loop, enabling systematic control of convergence, divergence and trajectory structure in iterative LLM transformations.
Authors: Yuseon Choi, Sangjin Kim, Jungjun Oh, Byeongcheol Kim, Hoi-Jun Yoo
Abstract: Low-bit quantization is a promising technique for efficient transformer inference by reducing computational and memory overhead. However, aggressive bitwidth reduction remains challenging due to activation outliers, leading to accuracy degradation. Existing methods, such as outlier-handling and group quantization, achieve high accuracy but incur substantial energy consumption. To address this, we propose SeVeDo, an energy-efficient SVD-based heterogeneous accelerator that structurally separates outlier-sensitive components into a high-precision low-rank path, while the remaining computations are executed in a low-bit residual datapath with group quantization. To further enhance efficiency, Hierarchical Group Quantization (HGQ) combines coarse-grained floating-point scaling with fine-grained shifting, effectively reducing dequantization cost. Also, SVD-guided mixed precision (SVD-MP) statically allocates higher bitwidths to precision-sensitive components identified through low-rank decomposition, thereby minimizing floating-point operation cost. Experimental results show that SeVeDo achieves a peak energy efficiency of 13.8TOPS/W, surpassing conventional designs, with 12.7TOPS/W on ViT-Base and 13.4TOPS/W on Llama2-7B benchmarks.
Authors: Xia Chen
Abstract: Conventional deep learning prioritizes unconstrained optimization, yet biological systems operate under strict metabolic constraints. We propose that these physical constraints shape dynamics to function not as limitations, but as a temporal inductive bias that breeds generalization. Through a phase-space analysis of signal propagation, we reveal a fundamental asymmetry: expansive dynamics amplify noise, whereas proper dissipative dynamics compress phase space that aligns with the network's spectral bias, compelling the abstraction of invariant features. This condition can be imposed externally via input encoding, or intrinsically through the network's own temporal dynamics. Both pathways require architectures capable of temporal integration and proper constraints to decode induced invariants, whereas static architectures fail to capitalize on temporal structure. Through comprehensive evaluations across supervised classification, unsupervised reconstruction, and zero-shot reinforcement learning, we demonstrate that a critical "transition" regime maximizes generalization capability. These findings establish dynamical constraints as a distinct class of inductive bias, suggesting that robust AI development requires not only scaling and removing limitations, but computationally mastering the temporal characteristics that naturally promote generalization.
Authors: George Ma, Zhongyuan Liang, Irene Y. Chen, Somayeh Sojoudi
Abstract: We investigate whether sparse autoencoders identify genuine reasoning features in large language models. We first present a stylized theoretical analysis showing that sparsity-regularized decoding favors stable low-dimensional correlates over high-dimensional within-reasoning variation, biasing learned features toward token-level cues. Motivated by this analysis, we introduce a falsification-based evaluation framework that combines causal token injection with LLM-guided counterexample generation to distinguish genuine reasoning features from superficial linguistic correlates. Across 22 configurations spanning multiple model families, layers and datasets, we find that contrastively selected reasoning features are highly sensitive to token interventions, with 45%-90% activating when only a few associated tokens are injected into non-reasoning text. For the remaining features, LLM-guided falsification reliably constructs non-reasoning inputs that instantiate the feature's token-level cues and trigger activation, and meaning-preserving paraphrases of top-activating reasoning traces that suppress it. Steering the highest-ranked features yields no improvements on benchmarks. Overall, our results suggest that when low-dimensional token-level patterns are coupled with high-dimensional reasoning processes, the sparsity bias of SAEs systematically favors low-dimensional linguistic patterns that consistently co-occur with reasoning. Code is available at https://github.com/GeorgeMLP/reasoning-probing.
Authors: Qiang Zhang, Boli Chen, Fanrui Zhang, Ruixue Ding, Shihang Wang, Qiuchen Wang, Yinfeng Huang, Haonan Zhang, Rongxiang Zhu, Pengyong Wang, Ailin Ren, Xin Li, Pengjun Xie, Jiawei Liu, Ning Guo, Jingren Zhou, Zheng-Jun Zha
Abstract: Reinforcement learning has substantially improved the performance of LLM agents on tasks with verifiable outcomes, but it still struggles on open-ended agent tasks with vast solution spaces (e.g., complex travel planning). Due to the absence of objective ground-truth for these tasks, current RL algorithms largely rely on reward models that assign scalar scores to individual responses. We contend that such pointwise scoring suffers from an inherent discrimination collapse: the reward model struggles to distinguish subtle advantages among different trajectories, resulting in scores within a group being compressed into a narrow range. Consequently, the effective reward signal becomes dominated by noise from the reward model, leading to optimization stagnation. To address this, we propose ArenaRL, a reinforcement learning paradigm that shifts from pointwise scalar scoring to intra-group relative ranking. ArenaRL introduces a process-aware pairwise evaluation mechanism, employing multi-level rubrics to assign fine-grained relative scores to trajectories. Additionally, we construct an intra-group adversarial arena and devise a tournament-based ranking scheme to obtain stable advantage signals. Empirical results confirm that the built seeded single-elimination scheme achieves nearly equivalent advantage estimation accuracy to full pairwise comparisons with O(N^2) complexity, while operating with only O(N) complexity, striking an optimal balance between efficiency and precision. Furthermore, to address the lack of full-cycle benchmarks for open-ended agents, we build Open-Travel and Open-DeepResearch, two high-quality benchmarks featuring a comprehensive pipeline covering SFT, RL training, and multi-dimensional evaluation. Extensive experiments show that ArenaRL substantially outperforms standard RL baselines, enabling LLM agents to generate more robust solutions for complex real-world tasks.
Authors: Jianqi Zhang, Jingyao Wang, Wenwen Qiang, Fanjiang Xu, Changwen Zheng
Abstract: The World Wide Web needs reliable predictive capabilities to respond to changes in user behavior and usage patterns. Time series forecasting (TSF) is a key means to achieve this goal. In recent years, the large language models (LLMs) for TSF (LLM4TSF) have achieved good performance. However, there is a significant difference between pretraining corpora and time series data, making it hard to guarantee forecasting quality when directly applying LLMs to TSF; fine-tuning LLMs can mitigate this issue, but often incurs substantial computational overhead. Thus, LLM4TSF faces a dual challenge of prediction performance and compute overhead. To address this, we aim to explore a method for improving the forecasting performance of LLM4TSF while freezing all LLM parameters to reduce computational overhead. Inspired by in-context learning (ICL), we propose LVICL. LVICL uses our vector-injected ICL to inject example information into a frozen LLM, eliciting its in-context learning ability and thereby enhancing its performance on the example-related task (i.e., TSF). Specifically, we first use the LLM together with a learnable context vector adapter to extract a context vector from multiple examples adaptively. This vector contains compressed, example-related information. Subsequently, during the forward pass, we inject this vector into every layer of the LLM to improve forecasting performance. Compared with conventional ICL that adds examples into the prompt, our vector-injected ICL does not increase prompt length; moreover, adaptively deriving a context vector from examples suppresses components harmful to forecasting, thereby improving model performance. Extensive experiments demonstrate the effectiveness of our approach.
Authors: Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, Jiajun Chai, Guojun Yin, Wei Lin, Shuai Ma, Fuzhen Zhuang, Deqing Wang, Yaodong Yang, Jianxin Li, Yikun Ban
Abstract: Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.
Authors: Sucheta Ghosh, Zahra Monfared, Felix Dietrich
Abstract: We introduce a two-stage multitask learning framework for analyzing Electroencephalography (EEG) signals that integrates denoising, dynamical modeling, and representation learning. In the first stage, a denoising autoencoder is trained to suppress artifacts and stabilize temporal dynamics, providing robust signal representations. In the second stage, a multitask architecture processes these denoised signals to achieve three objectives: motor imagery classification, chaotic versus non-chaotic regime discrimination using Lyapunov exponent-based labels, and self-supervised contrastive representation learning with NT-Xent loss. A convolutional backbone combined with a Transformer encoder captures spatial-temporal structure, while the dynamical task encourages sensitivity to nonlinear brain dynamics. This staged design mitigates interference between reconstruction and discriminative goals, improves stability across datasets, and supports reproducible training by clearly separating noise reduction from higher-level feature learning. Empirical studies show that our framework not only enhances robustness and generalization but also surpasses strong baselines and recent state-of-the-art methods in EEG decoding, highlighting the effectiveness of combining denoising, dynamical features, and self-supervised learning.
Authors: Andrew Kiruluta
Abstract: We introduce Spectral Generative Flow Models (SGFMs), a physics-inspired alternative to transformer-based large language models. Instead of representing text or video as sequences of discrete tokens processed by attention, SGFMs treat generation as the evolution of a continuous field governed by constrained stochastic dynamics in a multiscale wavelet basis. This formulation replaces global attention with local operators, spectral projections, and Navier--Stokes-like transport, yielding a generative mechanism grounded in continuity, geometry, and physical structure. Our framework provides three key innovations: (i) a field-theoretic ontology in which text and video are unified as trajectories of a stochastic partial differential equation; (ii) a wavelet-domain representation that induces sparsity, scale separation, and computational efficiency; and (iii) a constrained stochastic flow that enforces stability, coherence, and uncertainty propagation. Together, these components define a generative architecture that departs fundamentally from autoregressive modeling and diffusion-based approaches. SGFMs offer a principled path toward long-range coherence, multimodal generality, and physically structured inductive bias in next-generation generative models.
Authors: Nilin Abrahamsen
Abstract: This note introduces Projected Microbatch Accumulation (PROMA), a proximal policy method that modifies gradient accumulation across microbatches rather than relying on likelihood ratios relative to a reference policy. During accumulation, PROMA projects the partially accumulated gradient to be orthogonal to the sequence-wise gradients of the current microbatch. This projection is applied layer-wise during the backward pass, enabling efficient implementation. Empirically, PROMA achieves proximal updates without entropy collapse while providing tighter local KL control than GRPO.
Authors: Sofiane Tanji, Samuel Vaiter, Yassine Laguel
Abstract: Despite their promise, fair machine learning methods often yield Pareto-inefficient models, in which the performance of certain groups can be improved without degrading that of others. This issue arises frequently in traditional in-processing approaches such as fairness-through-regularization. In contrast, existing Pareto-efficient approaches are biased towards a certain perspective on fairness and fail to adapt to the broad range of fairness metrics studied in the literature. In this paper, we present BADR, a simple framework to recover the optimal Pareto-efficient model for any fairness metric. Our framework recovers its models through a Bilevel Adaptive Rescalarisation procedure. The lower level is a weighted empirical risk minimization task where the weights are a convex combination of the groups, while the upper level optimizes the chosen fairness objective. We equip our framework with two novel large-scale, single-loop algorithms, BADR-GD and BADR-SGD, and establish their convergence guarantees. We release badr, an open-source Python toolbox implementing our framework for a variety of learning tasks and fairness metrics. Finally, we conduct extensive numerical experiments demonstrating the advantages of BADR over existing Pareto-efficient approaches to fairness.
Authors: Sayeed Shafayet Chowdhury, Snehasis Mukhopadhyay, Shiaofen Fang, Vijay R. Ramakrishnan
Abstract: Artificial intelligence has reshaped medical imaging, yet the use of AI on clinical data for prospective decision support remains limited. We study pre-operative prediction of clinically meaningful improvement in chronic rhinosinusitis (CRS), defining success as a more than 8.9-point reduction in SNOT-22 at 6 months (MCID). In a prospectively collected cohort where all patients underwent surgery, we ask whether models using only pre-operative clinical data could have identified those who would have poor outcomes, i.e. those who should have avoided surgery. We benchmark supervised ML (logistic regression, tree ensembles, and an in-house MLP) against generative AI (ChatGPT, Claude, Gemini, Perplexity), giving each the same structured inputs and constraining outputs to binary recommendations with confidence. Our best ML model (MLP) achieves 85 % accuracy with superior calibration and decision-curve net benefit. GenAI models underperform on discrimination and calibration across zero-shot setting. Notably, GenAI justifications align with clinician heuristics and the MLP's feature importance, repeatedly highlighting baseline SNOT-22, CT/endoscopy severity, polyp phenotype, and physchology/pain comorbidities. We provide a reproducible tabular-to-GenAI evaluation protocol and subgroup analyses. Findings support an ML-first, GenAI- augmented workflow: deploy calibrated ML for primary triage of surgical candidacy, with GenAI as an explainer to enhance transparency and shared decision-making.
Authors: Deming Chen, Vijay Ganesh, Weikai Li, Yingyan Celine Lin, Yong Liu, Subhasish Mitra, David Z. Pan, Ruchir Puri, Jason Cong, Yizhou Sun
Abstract: This report distills the discussions and recommendations from the NSF Workshop on AI for Electronic Design Automation (EDA), held on December 10, 2024 in Vancouver alongside NeurIPS 2024. Bringing together experts across machine learning and EDA, the workshop examined how AI-spanning large language models (LLMs), graph neural networks (GNNs), reinforcement learning (RL), neurosymbolic methods, etc.-can facilitate EDA and shorten design turnaround. The workshop includes four themes: (1) AI for physical synthesis and design for manufacturing (DFM), discussing challenges in physical manufacturing process and potential AI applications; (2) AI for high-level and logic-level synthesis (HLS/LLS), covering pragma insertion, program transformation, RTL code generation, etc.; (3) AI toolbox for optimization and design, discussing frontier AI developments that could potentially be applied to EDA tasks; and (4) AI for test and verification, including LLM-assisted verification tools, ML-augmented SAT solving, security/reliability challenges, etc. The report recommends NSF to foster AI/EDA collaboration, invest in foundational AI for EDA, develop robust data infrastructures, promote scalable compute infrastructure, and invest in workforce development to democratize hardware design and enable next-generation hardware systems. The workshop information can be found on the website https://ai4eda-workshop.github.io/.
Authors: Injin Kong, Hyoungjoon Lee, Yohan Jo
Abstract: Post-training pretrained Autoregressive models (ARMs) into Masked Diffusion models (MDMs) has emerged as a cost-effective strategy to overcome the limitations of sequential generation. However, the internal algorithmic transformations induced by this paradigm shift remain unexplored, leaving it unclear whether post-trained MDMs acquire genuine bidirectional reasoning capabilities or merely repackage autoregressive heuristics. In this work, we address this question by conducting a comparative circuit analysis of ARMs and their MDM counterparts. Our analysis reveals a systematic "mechanism shift" dependent on the structural nature of the task. Structurally, we observe a distinct divergence: while MDMs largely retain autoregressive circuitry for tasks dominated by local causal dependencies, they abandon initialized pathways for global planning tasks, exhibiting distinct rewiring characterized by increased early-layer processing. Semantically, we identify a transition from sharp, localized specialization in ARMs to distributed integration in MDMs. Through these findings, we conclude that diffusion post-training does not merely adapt model parameters but fundamentally reorganizes internal computation to support non-sequential global planning.
Authors: Baojun Che, Yifan Chen, Daniel Zhengyu Huang, Xinying Mao, Weijie Wang
Abstract: Black-box variational inference (BBVI) with Gaussian mixture families offers a flexible approach for approximating complex posterior distributions without requiring gradients of the target density. However, standard numerical optimization methods often suffer from instability and inefficiency. We develop a stable and efficient framework that combines three key components: (1) affine-invariant preconditioning via natural gradient formulations, (2) an exponential integrator that unconditionally preserves the positive definiteness of covariance matrices, and (3) adaptive time stepping to ensure stability and to accommodate distinct warm-up and convergence phases. The proposed approach has natural connections to manifold optimization and mirror descent. For Gaussian posteriors, we prove exponential convergence in the noise-free setting and almost-sure convergence under Monte Carlo estimation, rigorously justifying the necessity of adaptive time stepping. Numerical experiments on multimodal distributions, Neal's multiscale funnel, and a PDE-based Bayesian inverse problem for Darcy flow demonstrate the effectiveness of the proposed method.
Authors: Garrett G. Wen, Buxin Su, Natalie Collina, Zhun Deng, Weijie Su
Abstract: Machine learning and artificial intelligence conferences such as NeurIPS and ICML now regularly receive tens of thousands of submissions, posing significant challenges to maintaining the quality and consistency of the peer review process. This challenge is particularly acute for best paper awards, which are an important part of the peer review process, yet whose selection has increasingly become a subject of debate in recent years. In this paper, we introduce an author-assisted mechanism to facilitate the selection of best paper awards. Our method employs the Isotonic Mechanism for eliciting authors' assessments of their own submissions in the form of a ranking, which is subsequently utilized to adjust the raw review scores for optimal estimation of the submissions' ground-truth quality. We demonstrate that authors are incentivized to report truthfully when their utility is a convex additive function of the adjusted scores, and we validate this convexity assumption for best paper awards using publicly accessible review data of ICLR from 2019 to 2023 and NeurIPS from 2021 to 2023. Crucially, in the special case where an author has a single quota -- that is, may nominate only one paper -- we prove that truthfulness holds even when the utility function is merely nondecreasing and additive. This finding represents a substantial relaxation of the assumptions required in prior work. For practical implementation, we extend our mechanism to accommodate the common scenario of overlapping authorship. Finally, simulation results demonstrate that our mechanism significantly improves the quality of papers selected for awards.
Authors: Moad Abudia, Tejasvi Channagiri, Joel A. Rosenfeld, Rushikesh Kamalapurkar
Abstract: This manuscript presents an algorithm for obtaining an approximation of a nonlinear high order control affine dynamical system. Controlled trajectories of the system are leveraged as the central unit of information via embedding them in vector-valued reproducing kernel Hilbert space (vvRKHS). The trajectories are embedded as the so-called higher order control occupation kernels which represent an operator on the vvRKHS corresponding to iterated integration after multiplication by a given controller. The solution to the system identification problem is then the unique solution of an infinite dimensional regularized regression problem. The representer theorem is then used to express the solution as finite linear combination of these occupation kernels, which converts an infinite dimensional optimization problem to a finite dimensional optimization problem. The vector valued structure of the Hilbert space allows for simultaneous approximation of the drift and control effectiveness components of the control affine system. Several experiments are performed to demonstrate the effectiveness of the developed approach.
Authors: Gengyuan Zhang, Jisen Ren, Jindong Gu, Volker Tresp
Abstract: Video-Text Retrieval (VTR) is a crucial multi-modal task in an era of massive video-text data on the Internet. A plethora of work characterized by using a two-stream Vision-Language model architecture that learns a joint representation of video-text pairs has become a prominent approach for the VTR task. However, these models operate under the assumption of bijective video-text correspondences and neglect a more practical scenario where video content usually encompasses multiple events, while texts like user queries or webpage metadata tend to be specific and correspond to single events. This establishes a gap between the previous training objective and real-world applications, leading to the potential performance degradation of earlier models during inference. In this study, we introduce the Multi-event Video-Text Retrieval (MeVTR) task, addressing scenarios in which each video contains multiple different events, as a niche scenario of the conventional Video-Text Retrieval Task. We present a simple model, Me-Retriever, which incorporates key event video representation and a new MeVTR loss for the MeVTR task. Comprehensive experiments show that this straightforward framework outperforms other models in the Video-to-Text and Text-to-Video tasks, effectively establishing a robust baseline for the MeVTR task. We believe this work serves as a strong foundation for future studies. Code is available at https://github.com/gengyuanmax/MeVTR.
Authors: Robin Young
Abstract: We propose an information-theoretic formalization of the distinction between two fundamental AI safety failure modes: deceptive alignment and goal drift. While both can lead to systems that appear misaligned, we demonstrate that they represent distinct forms of information divergence occurring at different interfaces in the human-AI system. Deceptive alignment creates entropy between an agent's true goals and its observable behavior, while goal drift, or confusion, creates entropy between the intended human goal and the agent's actual goal. Though often observationally equivalent, these failures necessitate different interventions. We present a formal model and an illustrative thought experiment to clarify this distinction. We offer a formal language for re-examining prominent alignment challenges observed in Large Language Models (LLMs), offering novel perspectives on their underlying causes.
Authors: S\"oren Christensen, Jan Kallsen, Claudia Strauch, Lukas Trottner
Abstract: We introduce a new class of generative diffusion models that, unlike conventional denoising diffusion models, achieve a time-homogeneous structure for both the noising and denoising processes, allowing the number of steps to adaptively adjust based on the noise level. This is accomplished by conditioning the forward process using Doob's $h$-transform, which terminates the process at a suitable sampling distribution at a random time. The model is particularly well suited for generating data with lower intrinsic dimensions, as the termination criterion simplifies to a first-hitting rule. A key feature of the model is its adaptability to the target data, enabling a variety of downstream tasks using a pre-trained unconditional generative model. These tasks include natural conditioning through appropriate initialisation of the denoising process and classification of noisy data.
Authors: Emanuele Iannone, Quang-Cuong Bui, Riccardo Scandariato
Abstract: Software vulnerabilities are often detected via taint analysis, penetration testing, or fuzzing. They are also found via unit tests that exercise security-sensitive behavior with specific inputs, called vulnerability-witnessing tests. Generative AI models could help developers in writing them, but they require many examples to learn from, which are currently scarce. This paper introduces VuTeCo, an AI-driven framework for collecting examples of vulnerability-witnessing tests from Java repositories. VuTeCo carries out two tasks: (1) The "Finding" task to determine whether a unit test case is security-related, and (2) the "Matching" task to relate a test case to the vulnerability it witnesses. VuTeCo addresses the Finding task with UniXcoder, achieving an F0.5 score of 0.73 and a precision of 0.83 on a test set of unit tests from Vul4J. The Matching task is addressed using DeepSeek Coder, achieving an F0.5 score of 0.65 and a precision of 0.75 on a test set of pairs of unit tests and vulnerabilities from Vul4J. VuTeCo has been used in the wild on 427 Java projects and 1,238 vulnerabilities, obtaining 224 test cases confirmed to be security-related and 35 tests correctly matched to 29 vulnerabilities. The validated tests were collected in a new dataset called Test4Vul. VuTeCo lays the foundation for large-scale retrieval of vulnerability-witnessing tests, enabling future AI models to better understand and generate security unit tests.
Authors: Jim W. Barrett, Nils Erlanson, Joana F\'elix China, G. Niklas Nor\'en
Abstract: Objectives: To advance state-of-the-art for duplicate detection in large-scale pharmacovigilance databases and achieve more consistent performance across adverse event reports from different countries. Background: Unlinked adverse event reports referring to the same case impede statistical analysis and may mislead clinical assessment. Pharmacovigilance relies on large databases of adverse event reports to discover potential new causal associations, and computational methods are required to identify duplicates at scale. Current state-of-the-art is statistical record linkage which outperforms rule-based approaches. In particular, vigiMatch is in routine use for VigiBase, the WHO global database of adverse event reports, and represents the first statistical duplicate detection approach in pharmacovigilance deployed at scale. Originally developed for both medicines and vaccines, its application to vaccines has been limited due to inconsistent performance across countries. Methods: This paper extends vigiMatch from probabilistic record linkage to predictive modelling, refining features for medicines, vaccines, and adverse events using country-specific reporting rates, extracting dates from free text, and training separate support vector machine classifiers for medicines and vaccines. Recall was evaluated using 5 independent labelled test sets. Precision was assessed by annotating random selections of report pairs classified as duplicates. Results: Precision for the new method was 92% for vaccines and 54% for medicines, compared with 41% for the comparator method. Recall ranged from 80-85% across test sets for vaccines and from 40-86% for medicines, compared with 24-53% for the comparator method. Conclusion: Predictive modeling, use of free text, and country-specific features advance state-of-the-art for duplicate detection in pharmacovigilance.
Authors: Berk Tinaz, Zalan Fabian, Mahdi Soltanolkotabi
Abstract: Diffusion models have become the go-to method for text-to-image generation, producing high-quality images from pure noise. However, the inner workings of diffusion models is still largely a mystery due to their black-box nature and complex, multi-step generation process. Mechanistic interpretability techniques, such as Sparse Autoencoders (SAEs), have been successful in understanding and steering the behavior of large language models at scale. However, the great potential of SAEs has not yet been applied toward gaining insight into the intricate generative process of diffusion models. In this work, we leverage the SAE framework to probe the inner workings of a popular text-to-image diffusion model, and uncover a variety of human-interpretable concepts in its activations. Interestingly, we find that even before the first reverse diffusion step is completed, the final composition of the scene can be predicted surprisingly well by looking at the spatial distribution of activated concepts. Moreover, going beyond correlational analysis, we design intervention techniques aimed at manipulating image composition and style, and demonstrate that (1) in early stages of diffusion image composition can be effectively controlled, (2) in the middle stages image composition is finalized, however stylistic interventions are effective, and (3) in the final stages only minor textural details are subject to change.
Authors: Daniel Goldstein, Eric Alcaide, Janna Lu, Eugene Cheah
Abstract: We present Rapid Attention Distillation to Linear Attention Decoders at Scale (RADLADS), a protocol for rapidly converting softmax attention transformers into linear attention decoder models, along with two new RWKV-variant architectures, and models converted from popular Qwen2.5 open source models in 7B, 32B, and 72B sizes. Our conversion process requires only 350-700M tokens, less than 0.005% of the token count used to train the original teacher models. Converting to our 72B linear attention model costs less than \$2,000 USD at today's prices, yet quality at inference remains close to the original transformer. These models achieve state-of-the-art downstream performance across a set of standard benchmarks for linear attention models of their size. We release all our models on HuggingFace under the Apache 2.0 license, with the exception of our 72B models which are also governed by the Qwen License Agreement. Models at https://huggingface.co/collections/recursal/radlads-6818ee69e99e729ba8a87102 Training Code at https://github.com/recursal/RADLADS-paper
URLs: https://huggingface.co/collections/recursal/radlads-6818ee69e99e729ba8a87102, https://github.com/recursal/RADLADS-paper
Authors: Adam \v{S}torek, Mukur Gupta, Samira Hajizadeh, Prashast Srivastava, Suman Jana
Abstract: Large language models (LLMs) are increasingly deployed for understanding large codebases, but whether they understand operational semantics of long code context or rely on pattern matching shortcuts remains unclear. We distinguish between lexical recall (retrieving code verbatim) and semantic recall (understanding operational semantics). Evaluating 10 state-of-the-art LLMs, we find that while frontier models achieve near-perfect, position-independent lexical recall, semantic recall degrades severely when code is centrally positioned in long contexts. We introduce semantic recall sensitivity to measure whether tasks require understanding of code's operational semantics vs. permit pattern matching shortcuts. Through a novel counterfactual measurement method, we show that models rely heavily on pattern matching shortcuts to solve existing code understanding benchmarks. We propose a new task SemTrace, which achieves high semantic recall sensitivity through unpredictable operations; LLMs' accuracy exhibits severe positional effects, with median accuracy drops of 92.73% versus CRUXEval's 53.36% as the relevant code snippet approaches the middle of the input code context. Our findings suggest current evaluations substantially underestimate semantic recall failures in long context code understanding.
Authors: Manuela Gonz\'alez-Gonz\'alez, Soufiane Belharbi, Muhammad Osama Zeeshan, Masoumeh Sharafi, Muhammad Haseeb Aslam, Marco Pedersoli, Alessandro Lameiras Koerich, Simon L Bacon, Eric Granger
Abstract: Ambivalence and hesitancy (A/H), a closely related construct, is the primary reasons why individuals delay, avoid, or abandon health behaviour changes. It is a subtle and conflicting emotion that sets a person in a state between positive and negative orientations, or between acceptance and refusal to do something. It manifests by a discord in affect between multiple modalities or within a modality, such as facial and vocal expressions, and body language. Although experts can be trained to recognize A/H as done for in-person interactions, integrating them into digital health interventions is costly and less effective. Automatic A/H recognition is therefore critical for the personalization and cost-effectiveness of digital behaviour change interventions. However, no datasets currently exists for the design of machine learning models to recognize A/H. This paper introduces the Behavioural Ambivalence/Hesitancy (BAH) dataset collected for multimodal recognition of A/H in videos. It contains 1,427 videos with a total duration of 10.60 hours captured from 300 participants across Canada answering predefined questions to elicit A/H. It is intended to mirror real-world online personalized behaviour change interventions. BAH is annotated by three experts to provide timestamps that indicate where A/H occurs, and frame- and video-level annotations with A/H cues. Video transcripts, cropped and aligned faces, and participants' meta-data are also provided. Since A and H manifest similarly in practice, we provide a binary annotation indicating the presence or absence of A/H. Additionally, this paper includes benchmarking results using baseline models on BAH for frame- and video-level recognition, zero-shot prediction, and personalization using source-free domain adaptation. The data, code, and pretrained weights are available.
Authors: Dionysis Christopoulos, Sotiris Spanos, Eirini Baltzi, Valsamis Ntouskos, Konstantinos Karantzalos
Abstract: We introduce SLIMP (Skin Lesion Image-Metadata Pre-training) for learning rich representations of skin lesions through a novel nested contrastive learning approach that captures complex relationships between images and metadata. Melanoma detection and skin lesion classification based solely on images, pose significant challenges due to large variations in imaging conditions (lighting, color, resolution, distance, etc.) and lack of clinical and phenotypical context. Clinicians typically follow a holistic approach for assessing the risk level of the patient and for deciding which lesions may be malignant and need to be excised, by considering the patient's medical history as well as the appearance of other lesions of the patient. Inspired by this, SLIMP combines the appearance and the metadata of individual skin lesions with patient-level metadata relating to their medical record and other clinically relevant information. By fully exploiting all available data modalities throughout the learning process, the proposed pre-training strategy improves performance compared to other pre-training strategies on downstream skin lesions classification tasks highlighting the learned representations quality.
Authors: Lloyd Pellatt, Fotios Drakopoulos, Shievanie Sabesan, Nicholas A. Lesica
Abstract: The mapping from sound to neural activity that underlies hearing is highly non-linear. The first few stages of this mapping in the cochlea have been modelled successfully, with biophysical models built by hand and, more recently, with DNN models trained on datasets simulated by biophysical models. Modelling the auditory brain has been a challenge because central auditory processing is too complex for models to be built by hand, and datasets for training DNN models directly have not been available. Recent work has taken advantage of large-scale high resolution neural recordings from the auditory midbrain to build a DNN model of normal hearing with great success. But this model assumes that auditory processing is the same in all brains, and therefore it cannot capture the widely varying effects of hearing loss. We propose a novel variational-conditional model to learn to encode the space of hearing loss directly from recordings of neural activity in the auditory midbrain of healthy and noise exposed animals. With hearing loss parametrised by only 6 free parameters per animal, our model accurately predicts 62% of the explainable variance in neural responses from normal hearing animals and 68% for hearing impaired animals, within a few percentage points of state of the art animal specific models. We demonstrate that the model can be used to simulate realistic activity from out of sample animals by fitting only the learned conditioning parameters with Bayesian optimisation, achieving crossentropy loss within 2% of the optimum in 15-30 iterations. Including more animals in the training data slightly improved the performance on unseen animals. This model will enable future development of parametrised hearing loss compensation models trained to directly restore normal neural coding in hearing impaired brains, which can be quickly fitted for a new user by human in the loop optimisation.
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, Dawn Song, Christopher Summerfield, Audrey Tang, Jay J. Van Bavel, Sander van der Linden, Jonas R. Kunst
Abstract: Advances in AI offer the prospect of manipulating beliefs and behaviors on a population-wide level. Large language models and autonomous agents now let influence campaigns reach unprecedented scale and precision. Generative tools can expand propaganda output without sacrificing credibility and inexpensively create 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, a 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 efficiently. By adaptively mimicking human social dynamics, they threaten democracy. Because the resulting harms stem from design, commercial incentives, and governance, we prioritize interventions at multiple leverage points, focusing on pragmatic mechanisms over voluntary compliance.
Authors: Chong Di, Jinglin Zhang, Zhenjiang Li, Jean-Marie Mirebeau, Da Chen, Laurent D. Cohen
Abstract: Optimal curve methods provide a fundamental framework for tubular centerline tracking. Point-wise approaches, such as minimal paths, are theoretically elegant but often suffer from shortcut and short-branch combination problems in complex scenarios. Nonlocal segment-wise methods address these issues by mapping pre-extracted centerline fragments onto a segment-proposal graph, performing optimization in this abstract space, and recovering the target tubular centerline from the resulting optimal path. In this paradigm, graph construction is critical, as it directly determines the quality of the final result. However, existing segment-wise methods construct graphs in a static manner, requiring all edges and their weights to be pre-computed, i.e. the graph must be sufficiently complete prior to search. Otherwise, the true path may be absent from the candidate space, leading to search failure. To address this limitation, we propose a dynamic exploration scheme for constructing segment-proposal graphs, where the graph is built on demand during the search for optimal paths. By formulating the problem as a Markov decision process, we apply Q-learning to compute edge weights only for visited transitions and adaptively expand the action space when connectivity is insufficient. Experimental results on retinal vessels, roads, and rivers demonstrate consistent improvements over state-of-the-art methods in both accuracy and efficiency.
Authors: Uri Cohen, M\'at\'e Lengyel
Abstract: Neural networks storing multiple discrete attractors are canonical models of biological memory. Previously, the dynamical stability of such networks could only be guaranteed under highly restrictive conditions. Here, we derive a theory of the local stability of discrete fixed points in a broad class of networks with graded neural activities and in the presence of noise. By directly analyzing the bulk and the outliers of the Jacobian spectrum, we show that all fixed points are stable below a critical load that is distinct from the classical \textit{critical capacity} and depends on the statistics of neural activities in the fixed points as well as the single-neuron activation function. Our analysis highlights the computational benefits of threshold-linear activation and sparse-like patterns.
Authors: Lei Qian, Wu Su, Yanqi Huang, Song Xi Chen
Abstract: We propose a Likelihood Matching approach for training diffusion models by first establishing an equivalence between the likelihood of the target data distribution and a likelihood along the sample path of the reverse diffusion. To efficiently compute the reverse sample likelihood, a quasi-likelihood is considered to approximate each reverse transition density by a Gaussian distribution with matched conditional mean and covariance, respectively. The score and Hessian functions for the diffusion generation are estimated by maximizing the quasi-likelihood, ensuring a consistent matching of both the first two transitional moments between every two time points. A stochastic sampler is introduced to facilitate computation that leverages both the estimated score and Hessian information. We establish consistency of the quasi-maximum likelihood estimation, and provide non-asymptotic convergence guarantees for the proposed sampler, quantifying the rates of the approximation errors due to the score and Hessian estimation, dimensionality, and the number of diffusion steps. Empirical and simulation evaluations demonstrate the effectiveness of the proposed Likelihood Matching and validate the theoretical results.
Authors: Jiajie He, Min-Chun Chen, Xintong Chen, Xinyang Fang, Yuechun Gu, Keke Chen
Abstract: Large language models (LLMs) based recommender systems (RecSys) can adapt to different domains flexibly. It utilizes in-context learning (ICL), i.e., prompts, to customize the recommendation functions, which include sensitive historical user-specific item interactions, encompassing implicit feedback such as clicked items and explicit product reviews. Such private information may be exposed by novel privacy attacks. However, no study has been conducted on this important issue. We design several membership inference attacks (MIAs) aimed to revealing whether system prompts include victims' historical interactions. The attacks are \emph{Similarity, Memorization, Inquiry, and Poisoning attacks}, each utilizing unique features of LLMs or RecSys. We have carefully evaluated them on five of the latest open-source LLMs and three well-known RecSys benchmark datasets. The results confirm that the MIA threat to LLM RecSys is realistic: inquiry and poisoning attacks show significantly high attack advantages. We also discussed possible methods to mitigate such MIA threats. We have also analyzed the factors affecting these attacks, such as the number of shots in system prompts, the position of the victim in the shots, the number of poisoning items in the prompt,etc.
Authors: Abhijnan Nath, Carine Graff, Nikhil Krishnaswamy
Abstract: As Large Language Models (LLMs) get integrated into diverse workflows, they are increasingly being regarded as "collaborators" with humans, and required to work in coordination with other AI systems. If such AI collaborators are to reliably coordinate their actions and behaviors with humans or other AIs, their properties and behaviors over multi-turn interactions must be known and predictable. This paper examines how different alignment methods affect LLM agents' effectiveness as partners in multi-turn, multi-party collaborations. We study this question through the lens of intervention agents that insert themselves into group dialogues not to provide answers, but to encourage the collaborative group to slow down and reflect upon their reasoning for deliberative decision-making. Common alignment techniques are typically developed under simplified single-user settings and assume the optimality of the underlying token MDP. Using the theoretical lens of the modified-action MDP, we show how they do not account for the dynamics of long-horizon multi-party interactions. We present a novel roleplay simulation methodology, where we align LLMs according to different methods and then deploy them in collaborative task dialogues to quantify how interventions affect the trajectory of group collaboration, belief alignment, and coordination. Our results show that an intervention agent that is robust to action modification significantly outperforms common alignment baselines in supporting correct task outcomes.
Authors: Gokul Karthik Kumar, Rishabh Saraf, Ludovick Lepauloux, Abdul Muneer, Billel Mokeddem, Hakim Hacid
Abstract: Large language models (LLMs) have transformed NLP, yet their integration with audio remains underexplored despite audio's centrality to human communication. We introduce Falcon3-Audio, a family of Audio-Language Models (ALMs) built on instruction-tuned LLMs and Whisper encoders. Using a remarkably small amount of public audio data, less than 30K hours (5K unique), Falcon3-Audio-7B matches the best reported performance among open-weight models on the MMAU benchmark, with a score of 64.14, matching R1-AQA, while distinguishing itself through superior data and parameter efficiency, single-stage training, and transparency. Notably, our smallest 1B model remains competitive with larger open models ranging from 2B to 13B parameters. Through extensive ablations, we find that common complexities such as curriculum learning, multiple audio encoders, and intricate cross-attention connectors are not required for strong performance, even compared to models trained on over 500K hours of data.
Authors: Yujian Ma, Xikun Lu, Jinqiu Sang, Xianquan Jiang, Ruizhe Li
Abstract: Large pre-trained speech models such as Whisper offer strong generalization but pose significant challenges for resource-efficient adaptation. Low-Rank Adaptation (LoRA) has become a popular parameter-efficient fine-tuning method, yet its underlying mechanisms in speech tasks remain poorly understood. In this work, we conduct the first systematic mechanistic interpretability study of LoRA within the Whisper encoder for speech emotion recognition (SER). Using a suite of analytical tools, including layer contribution probing, logit-lens inspection, and representational similarity via singular value decomposition (SVD) and centered kernel alignment (CKA), we reveal two key mechanisms: a delayed specialization process that preserves general features in early layers before consolidating task-specific information, and a forward alignment, backward differentiation dynamic between LoRA's matrices. Our findings clarify how LoRA reshapes encoder hierarchies, providing both empirical insights and a deeper mechanistic understanding for designing efficient and interpretable adaptation strategies in large speech models. Our code is available at https://github.com/harryporry77/Behind-the-Scenes.
Authors: Ke Wang, Zishuo Zhao, Xinyuan Song, Zelin Li, Libin Xia, Chris Tong, Bill Shi, Wenjie Qu, Eric Yang, Lynn Ai
Abstract: Decentralized inference provides a scalable and resilient paradigm for serving large language models (LLMs), enabling fragmented global resource utilization and reducing reliance on centralized providers. However, in a permissionless environment without trusted nodes, ensuring the correctness of model outputs remains a core challenge. We introduce VeriLLM, a publicly verifiable protocol for decentralized LLM inference that achieves security with incentive guarantees while maintaining practical efficiency. VeriLLM combines lightweight empirical rerunning with minimal on-chain checks to preclude free-riding, allowing verifiers to validate results at approximately 1% of the underlying inference cost by exploiting the structural separation between prefill and autoregressive decoding. To prevent verification bottlenecks, we design an isomorphic inference--verification architecture that multiplexes both inference and verification roles across the same GPU workers. This design (i) improves GPU utilization and overall throughput, (ii) enlarges the effective validator set, enhancing robustness and liveness, and (iii) enforces task indistinguishability to prevent node-specific optimizations or selective behavior. Through theoretical analysis and system-level evaluation, we show that VeriLLM achieves reliable public verifiability with minimal overhead, offering a practical foundation for trustworthy and scalable decentralized LLM inference.
Authors: Fiona Victoria Stanley Jothiraj, Arunaggiri Pandian Karunanidhi, Seth A. Eichmeyer
Abstract: In semiconductor manufacturing, early detection of wafer defects is critical for product yield optimization. However, raw wafer data from wafer quality tests are often complex, unlabeled, imbalanced and can contain multiple defects on a single wafer, making it crucial to design clustering methods that remain reliable under such imperfect data conditions. We introduce DECOR, a deep clustering with orientation robustness framework that groups complex defect patterns from wafer maps into consistent clusters. We evaluate our method on the open source MixedWM38 dataset, demonstrating its ability to discover clusters without manual tuning. DECOR explicitly accounts for orientation variations in wafer maps, ensuring that spatially similar defects are consistently clustered regardless of its rotation or alignment. Experiments indicate that our method outperforms existing clustering baseline methods, thus providing a reliable and scalable solution in automated visual inspection systems.
Authors: Moran Yanuka, Paul Dixon, Eyal Finkelshtein, Daniel Rotman, Raja Giryes
Abstract: Speculative decoding accelerates autoregressive speech generation by letting a fast draft model propose tokens that a larger target model verifies. However, for speech LLMs that generate acoustic tokens, exact token matching is overly restrictive: many discrete tokens are acoustically or semantically interchangeable, reducing acceptance rates and limiting speedups. We introduce Principled Coarse-Graining (PCG), which verifies proposals at the level of Acoustic Similarity Groups (ASGs) derived from the target model's embedding space. By splitting each token's probability mass across the overlapping groups that contain it, we define an overlap-aware coarse-grained distribution and perform rejection sampling on the resulting group variable. This yields an exactness guarantee at the group level while allowing the accepted draft token to stand in for any member of the group in practice. On LibriTTS, PCG increases acceptance and throughput relative to standard speculative decoding and prior speech-specific relaxations while maintaining intelligibility and speaker similarity. These results suggest acoustically aware, group-level acceptance as a simple and general way to accelerate speech token generation while maintaining speech quality.
Authors: Jack Cook, Junxian Guo, Guangxuan Xiao, Yujun Lin, Song Han
Abstract: As large language models have grown larger, interest has grown in low-precision numerical formats such as NVFP4 as a way to improve speed and reduce memory usage. However, quantizing models to NVFP4 remains difficult as the lack of precision generally degrades model performance. In this work, we address this issue with Four Over Six (4/6), a modification to the block-scaled NVFP4 quantization algorithm that yields reduced quantization error. Unlike integer formats, floating point formats have non-uniform step sizes which create larger quantization error on larger values. 4/6 takes advantage of this by adaptively scaling some blocks to smaller FP4 values, making the distribution of representable values more uniform and reducing quantization error for near-maximal values. We show that 4/6 can be implemented efficiently on NVIDIA Blackwell GPUs, resulting in performance gains during both pre-training and inference with minimal computational overhead. In pre-training experiments with the Nemotron 3 Nano 30B-A3B model architecture, we find that 4/6 brings training loss closer to BF16 compared to models trained with current state-of-the-art NVFP4 training recipes. Our code is available at http://github.com/mit-han-lab/fouroversix.
Authors: Sofiia Chorna, Davide Tisi, Cesare Malosso, Wei Bin How, Michele Ceriotti, Sanggyu Chong
Abstract: The past few years have seen the development of ``universal'' machine-learning interatomic potentials (uMLIPs) capable of approximating the ground-state potential energy surface across a wide range of chemical structures and compositions with reasonable accuracy. While these models differ in the architecture and the dataset used, they share the ability to compress a staggering amount of chemical information into descriptive latent features. Herein, we systematically analyze what the different uMLIPs have learned by quantitatively assessing the relative information content of their latent features with feature reconstruction errors, and observing how the trends are affected by the choice of training set and training protocol. We find that uMLIPs encode the chemical space in significantly distinct ways, with substantial cross-model feature reconstruction errors. When variants of the same model architecture are considered, trends become dependent on the dataset, target, and training protocol of choice. We also observe that fine-tuning of a uMLIP retains a strong pre-training bias in the latent features. Finally, we discuss how atom-level features, which are directly output by MLIPs, can be compressed into global structure-level features via concatenation of progressive cumulants, each adding significantly new information about the variability across the atomic environments within a given system.
Authors: Devansh Arpit
Abstract: We identify test prediction variance (TPV)-- the first-order sensitivity of model outputs to parameter perturbations around a trained solution-- as a unifying quantity that links several classical observations about generalization in deep networks. TPV is a fully label-free object whose trace form separates the geometry of the trained model from the specific perturbation mechanism, allowing a broad family of parameter perturbations like SGD noise, label noise, finite-precision noise, and other post-training perturbations to be analyzed under a single framework. Theoretically, we show that TPV estimated on the training set converges to its test-set value in the overparameterized limit, providing the first result that prediction variance under local parameter perturbations can be inferred from training inputs alone, and this stability is decoupled from generalization performance. Empirically, TPV exhibits a striking stability across datasets and architectures even for extremely narrow networks. Further, TPV correlates well with test loss, serving as a training-set based predictive metric for generalization. Code available at github.com/devansharpit/TPV.
Authors: Yu Chen, Hongwei Lin
Abstract: Persistence diagrams (PDs) provide a powerful tool for understanding the topology of the underlying shape of a point cloud. However, identifying which points in PDs encode genuine signals remains challenging. This challenge directly hinders the practical adoption of topological data analysis in many applications, where automated and reliable interpretation of persistence diagrams is essential for downstream decision-making. In this paper, we study automatic significance detection for one-dimensional persistence diagrams. Specifically, we propose Topology Understanding Net (TUN), a multi-modal network that combines enhanced PD descriptors with self-attention, a PointNet-style point cloud encoder, learned fusion, and per-point classification, alongside stable preprocessing and imbalance-aware training. It provides an automated and effective solution for identifying significant points in PDs, which are critical for downstream applications. Experiments show that TUN outperforms classic methods in detecting significant points in PDs, illustrating its effectiveness in real-world applications.
Authors: Ziming Mao, Yihan Zhang, Chihan Cui, Zhen Huang, Kaichao You, Zhongjie Chen, Zhiying Xu, Zhenyu Gu, Scott Shenker, Costin Raiciu, Yang Zhou, Ion Stoica
Abstract: Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication systems such as DeepEP demonstrate strong performance but exhibit poor portability across heterogeneous GPU and NIC platforms. The poor portability is rooted in architecture: GPU-initiated token-level RDMA communication requires tight vertical integration between GPUs and NICs, e.g., GPU writes to NIC driver/MMIO interfaces. We present UCCL-EP, a portable EP communication system that delivers DeepEP-level performance across heterogeneous GPU and NIC hardware. UCCL-EP replaces GPU-initiated RDMA with a high-throughput GPU-CPU control channel: compact token-routing commands are transferred to multithreaded CPU proxies, which then issue GPUDirect RDMA operations on behalf of GPUs. UCCL-EP further emulates various ordering semantics required by specialized EP communication modes using RDMA immediate data, enabling correctness on NICs that lack such ordering, e.g., AWS EFA. We implement UCCL-EP on NVIDIA and AMD GPUs with EFA and Broadcom NICs. On EFA, it outperforms the best existing EP solution by up to $2.1\times$ for dispatch and combine throughput. On NVIDIA-only platform, UCCL-EP achieves comparable performance to the original DeepEP. UCCL-EP also improves token throughput on SGLang by up to 40% on the NVIDIA+EFA platform, and improves DeepSeek-V3 training throughput over the AMD Primus/Megatron-LM framework by up to 45% on a 16-node AMD+Broadcom platform.
Authors: Omer Gazit, Yael Itzhakev, Yuval Elovici, Asaf Shabtai
Abstract: Radio frequency (RF) based systems are increasingly used to detect drones by analyzing their RF signal patterns, converting them into spectrogram images which are processed by object detection models. Existing RF attacks against image based models alter digital features, making over-the-air (OTA) implementation difficult due to the challenge of converting digital perturbations to transmittable waveforms that may introduce synchronization errors and interference, and encounter hardware limitations. We present the first physical attack on RF image based drone detectors, optimizing class-specific universal complex baseband (I/Q) perturbation waveforms that are transmitted alongside legitimate communications. We evaluated the attack using RF recordings and OTA experiments with four types of drones. Our results show that modest, structured I/Q perturbations are compatible with standard RF chains and reliably reduce target drone detection while preserving detection of legitimate drones.
Authors: Johan Hallberg Szabadv\'ary
Abstract: Conformal Test Martingales (CTMs) are a standard method within the Conformal Prediction framework for testing the crucial assumption of data exchangeability by monitoring deviations from uniformity in the p-value sequence. Although exchangeability implies uniform p-values, the converse does not hold. This raises the question of whether a significant break in exchangeability can occur, such that the p-values remain uniform, rendering CTMs blind. We answer this affirmatively, demonstrating the phenomenon of \emph{conformal blindness}. Through explicit construction, for the theoretically ideal ``predictive oracle'' conformity measure (given by the true conditional density), we demonstrate the possibility of an \emph{$A$-cryptic change-point} (where $A$ refers to the conformity measure). Using bivariate Gaussian distributions, we identify a line along which a change in the marginal means does not alter the distribution of the conformity scores, thereby producing perfectly uniform p-values. Simulations confirm that even a massive distribution shift can be perfectly cryptic to the CTM, highlighting a fundamental limitation and emphasising the critical role of the alignment of the conformity measure with potential shifts. By contrasting the predictive oracle with recent results on detection-optimal scores, we emphasise that validity monitoring in safety-critical systems requires careful separation of predictive and diagnostic goals.
Authors: Jiwei Guan, Haibo Jin, Haohan Wang
Abstract: Recent advancements in Large Vision-Language Models (LVLMs) have shown groundbreaking capabilities across diverse multimodal tasks. However, these models remain vulnerable to adversarial jailbreak attacks, where adversaries craft subtle perturbations to bypass safety mechanisms and trigger harmful outputs. Existing white-box attacks methods require full model accessibility, suffer from computing costs and exhibit insufficient adversarial transferability, making them impractical for real-world, black-box settings. To address these limitations, we propose a black-box jailbreak attack on LVLMs via Zeroth-Order optimization using Simultaneous Perturbation Stochastic Approximation (ZO-SPSA). ZO-SPSA provides three key advantages: (i) gradient-free approximation by input-output interactions without requiring model knowledge, (ii) model-agnostic optimization without the surrogate model and (iii) lower resource requirements with reduced GPU memory consumption. We evaluate ZO-SPSA on three LVLMs, including InstructBLIP, LLaVA and MiniGPT-4, achieving the highest jailbreak success rate of 83.0% on InstructBLIP, while maintaining imperceptible perturbations comparable to white-box methods. Moreover, adversarial examples generated from MiniGPT-4 exhibit strong transferability to other LVLMs, with ASR reaching 64.18%. These findings underscore the real-world feasibility of black-box jailbreaks and expose critical weaknesses in the safety mechanisms of current LVLMs
Authors: Benedikt Mayrhofer, Franz Pernkopf, Philipp Aichinger, Martin Hagm\"uller
Abstract: Electro-laryngeal (EL) speech is characterized by constant pitch, limited prosody, and mechanical noise, reducing naturalness and intelligibility. We propose a lightweight adaptation of the state-of-the-art StreamVC framework to this setting by removing pitch and energy modules and combining self-supervised pretraining with supervised fine-tuning on parallel EL and healthy (HE) speech data, guided by perceptual and intelligibility losses. Objective and subjective evaluations across different loss configurations confirm their influence: the best model variant, based on WavLM features and human-feedback predictions (+WavLM+HF), drastically reduces character error rate (CER) of EL inputs, raises naturalness mean opinion score (nMOS) from 1.1 to 3.3, and consistently narrows the gap to HE ground-truth speech in all evaluated metrics. These findings demonstrate the feasibility of adapting lightweight voice conversion architectures to EL voice rehabilitation while also identifying prosody generation and intelligibility improvements as the main remaining bottlenecks.
Authors: Adar Avsian, Christopher Richardson, Anirudh Sundar, Larry Heck
Abstract: Language models have become effective at a wide range of tasks, from math problem solving to open-domain question answering. However, they still make mistakes, and these mistakes are often repeated across related queries. Natural language explanations can help correct these errors, but collecting them at scale may be infeasible, particularly in domains where expert annotators are required. To address this issue, we introduce FLEx ($\textbf{F}$ew-shot $\textbf{L}$anguage $\textbf{Ex}$planations), a method for improving model behavior using a small number of explanatory examples. FLEx selects representative model errors using embedding-based clustering, verifies that the associated explanations correct those errors, and summarizes them into a prompt prefix that is prepended at inference-time. This summary guides the model to avoid similar errors on new inputs, without modifying model weights. We evaluate FLEx on CounterBench, GSM8K, and ReasonIF. We find that FLEx consistently outperforms chain-of-thought (CoT) prompting across all three datasets and reduces up to 83\% of CoT's remaining errors.
Authors: Rohit Goswami (Institute IMX and Lab-COSMO, \'Ecole polytechnique f\'ed\'erale de Lausanne, Science Institute, University of Iceland, Reykjavik, Iceland), Miha Gunde (Science Institute, University of Iceland, Reykjavik, Iceland, Institute Ru{\dj}er Bo\v{s}kovi\'c, Zagreb, Croatia), Hannes J\'onsson
Abstract: Accurate determination of transition states is central to an understanding of reaction kinetics. Double-endpoint methods where both initial and final states are specified, such as the climbing image nudged elastic band (CI-NEB), identify the minimum energy path between the two and thereby the saddle point on the energy surface that is relevant for the given transition, thus providing an estimate of the transition state within the harmonic approximation of transition state theory. Such calculations can, however, incur high computational costs and may suffer stagnation on exceptionally flat or rough energy surfaces. Conversely, methods that only require specification of an initial set of atomic coordinates, such as the minimum mode following (MMF) method, offer efficiency but can converge on saddle points that are not relevant for transition of interest. Here, we present an adaptive hybrid algorithm that integrates the CI-NEB with the MMF method so as to get faster convergence to the relevant saddle point. The method is benchmarked for the Baker-Chan (BC) saddle point test set using the PET-MAD machine-learned potential as well as 59 transitions of a heptamer island on Pt(111) from the OptBench benchmark set. A Bayesian analysis of the performance shows a median reduction in energy and force calculations of 46% [95% CrI: -55%, -37%] relative to CI-NEB for the BC set, while a 28% reduction is found for the transitions of the heptamer island. These results establish this hybrid method as a highly effective tool for high-throughput automated chemical discovery of atomic rearrangements.
Authors: Wenzhong Zhang, Zheyuan Hu, Wei Cai, George EM Karniadakis
Abstract: The numerical solution of high dimensional partial differential equations (PDEs) is severely constrained by the curse of dimensionality (CoD), rendering classical grid--based methods impractical beyond a few dimensions. In recent years, deep neural networks have emerged as a promising mesh free alternative, enabling the approximation of PDE solutions in tens to thousands of dimensions. This review provides a tutorial--oriented introduction to neural--network--based methods for solving high dimensional parabolic PDEs, emphasizing conceptual clarity and methodological connections. We organize the literature around three unifying paradigms: (i) PDE residual--based approaches, including physicsinformed neural networks and their high dimensional variants; (ii) stochastic methods derived from Feynman--Kac and backward stochastic differential equation formulations; and (iii) hybrid derivative--free random difference approaches designed to alleviate the computational cost of derivatives in high dimensions. For each paradigm, we outline the underlying mathematical formulation, algorithmic implementation, and practical strengths and limitations. Representative benchmark problems--including Hamilton--Jacobi--Bellman and Black--Scholes equations in up to 1000 dimensions --illustrate the scalability, effectiveness, and accuracy of the methods. The paper concludes with a discussion of open challenges and future directions for reliable and scalable solvers of high dimensional PDEs.
Authors: Wenzhi Gao, Chang He, Madeleine Udell
Abstract: This paper introduces a new problem-dependent regret measure for online convex optimization with smooth losses. The notion, which we call the $G^\star$ regret, depends on the cumulative squared gradient norm evaluated at the decision in hindsight $\sum_{t=1}^T \|\nabla \ell(x^\star)\|^2$. We show that the $G^\star$ regret strictly refines the existing $L^\star$ (small loss) regret, and that it can be arbitrarily sharper when the losses have vanishing curvature around the hindsight decision. We establish upper and lower bounds on the $G^\star$ regret and extend our results to dynamic regret and bandit settings. As a byproduct, we refine the existing convergence analysis of stochastic optimization algorithms in the interpolation regime. Some experiments validate our theoretical findings.
Authors: Hunter Heidenreich, Ben Elliott, Olivia Dinica, Yosheb Getachew
Abstract: GutenOCR is a family of grounded OCR front-ends obtained by fine-tuning Qwen2.5-VL-3B and Qwen2.5-VL-7B. The resulting single-checkpoint vision-language models expose reading, detection, and grounding through a unified, prompt-based interface. Trained on business documents, scientific articles, and synthetic grounding data, the models support full-page and localized reading with line- and paragraph-level bounding boxes and conditional ``where is x?'' queries. We introduce a grounded OCR evaluation protocol and show that GutenOCR-7B more than doubles the composite grounded OCR score of its Qwen2.5-VL-7B backbone on 10.5K held-out business and scientific pages (0.40 to 0.82). On Fox and OmniDocBench v1.5, our approach substantially improves region- and line-level OCR as well as text-detection recall, but reveals trade-offs in page-level linearization, color-guided OCR, and formula-heavy layouts.
Authors: Hirofumi Ota, Masaaki Imaizumi
Abstract: We study a linear observation model with an unknown permutation called \textit{permuted/shuffled linear regression}, where responses and covariates are mismatched and the permutation forms a discrete, factorial-size parameter. The permutation is a key component of the data-generating process, yet its statistical investigation remains challenging due to its discrete nature. We develop a general statistical inference framework on the permutation and regression coefficients. First, we introduce a localization step that reduces the permutation space to a small candidate set building on recent advances in the repro samples method, whose miscoverage decays polynomially with the number of Monte Carlo samples. Then, based on this localized set, we provide statistical inference procedures: a conditional Monte Carlo test of permutation structures with valid finite-sample Type-I error control. We also develop coefficient inference that remains valid under alignment uncertainty of permutations. For computational purposes, we develop a linear assignment problem computable in polynomial time and demonstrate that, with high probability, the solution is equivalent to that of the conventional least squares with large computational cost. Extensions to partially permuted designs and ridge regularization are further discussed. Extensive simulations and an application to air-quality data corroborate finite-sample validity, strong power to detect mismatches, and practical scalability.
Authors: Gokul Karthik Kumar, Ludovick Lepauloux, Hakim Hacid
Abstract: Whisper has become the de-facto encoder for extracting general-purpose audio features in large audio-language models, where a 30-second clip is typically represented by 1500 frame features projected into an LLM. In contrast, audio-text embedding models like CLAP-based models have largely relied on alternative audio encoders (e.g., HTS-AT, PaSST), and have not leveraged Whisper effectively. We present WavLink, a compact audio-text embedding model that augments Whisper encoder with a learnable global token, trained jointly with a text encoder. Through a systematic study of design choices, including pretrained text encoders, loss functions, training modes, and data mixtures, we identify configurations that yield state-of-the-art retrieval performance. Our two-stage training recipe across three model sizes, combined with Matryoshka-style supervision, improves scalability, enabling 8x smaller embeddings with minimal performance drop. WavLink also demonstrates competitive performance on AIR-Bench with MCQs and zero-shot classification.