New Machine Learning (cs.LG) papers, Monday 19 October 2026

224 papers announced on arXiv. Get them ranked for you, free →

1. Property Estimation in Geotechnical Databases Using Labeled Random Finite Sets

   Changbeom Shim, Youngho Kim, Craig Butterworth · arXiv:2410.22659

   The sufficiency of accurate data is a core element in data-centric geotechnics. However, geotechnical datasets are essentially uncertain, whereupon engineers have difficulty with obtaining precise information for making decisions. This challenge is more apparent when the…

2. CausalEvo: Causal Skill Self-Evolution for Health LLM Agents

   Yuchen Zhang · arXiv:2610.21016

   Large language model (LLM) agents for health can accumulate reusable textual skills from task feedback, such as calculation criteria, measurement time points, and formula units. Deciding which skills to retain is consequential: useful rules improve later answers, while faulty…

3. How language model agents reshape price competition

   Martin Bichler, Philip Hierhager, Kassian Köck et al. · arXiv:2610.21711

   Repeated Bertrand competition is a stringent benchmark for algorithmic pricing: in homogeneous-goods markets, standard learning procedures typically select the stage-game Nash outcome with prices near marginal cost. We show that LLM pricing agents can escape this benchmark,…

4. Encoded but Exposed: Language Embeddings can Reveal More Than They Should in Recommender Systems

   Linda Beková, Rodrigo Alves · arXiv:2610.21713

   Language embeddings are increasingly used in recommenders to represent item content, but their privacy risks remain poorly understood. We study whether item embeddings can reveal item identities when an adversary has access to auxiliary item text from external sources. We show…

5. Causal Integrated Gradients: Scalable Feature Attribution Along Causal Paths

   Joran Michiels, Anis Ismail, Maarten De Vos et al. · arXiv:2610.22126

   Existing feature attribution methods that account for feature dependencies do not scale to high dimensions. We propose causal integrated gradients (CIG) and causal expected gradients (CEG), which extend integrated gradients to incorporate prior causal structure among input…

6. Physics-Transfer Learning: A Physics-Constrained Theory for Zero-Shot Generalization Along the Ladder of Complexity

   Yingjie Zhao, Zhiping Xu · arXiv:2610.22376

   Reliable out-of-distribution generalization remains a central challenge in machine learning, particularly when models are deployed in data-scarce high-fidelity domains. Existing approaches often rely on statistical distribution alignment, or invariance assumptions that can fail…

7. Memory Bandwidth, Not FLOPs: Profiling and Accelerating Query-Based Segmentation Decoders on Edge GPUs

   Oliver Busemann (Osnabrück University, CLAAS E-Systems GmbH), Torben Töniges (CLAAS E-Systems GmbH) et al. · arXiv:2610.22377

   AI summaryThe paper shows query‑based segmentation decoders are memory‑bandwidth‑bound on edge GPUs, not FLOPs, and introduces CompactInst, a parameter‑free spatial pooling scheme that cuts decoder traffic by 4×, yielding 2–5× speedups with minimal accuracy loss.

8. When is a large language model a good calibration agent? Physics-labeled prompting for budgeted engineering simulator calibration

   Ruo-Qian Wang · arXiv:2610.22379

   AI summaryLLMs guided by physics‑labeled prompts produce superior early parameter updates for expensive simulator calibration, reducing error 21.9× faster than physics‑free methods in the first five evaluations. A hybrid LLM‑warm‑start followed by conventional optimization reaches target accuracy in just nine runs.

9. Trust but Verify: Designing Region-Inhomogeneous Forward Dynamics for Discrete Diffusion

   Gunmay Jhingran · arXiv:2610.22380

   AI summaryThe paper proposes a region‑inhomogeneous forward diffusion where missing tokens use absorbing dynamics while context undergoes low‑rate random substitution, allowing the model to learn to detect and correct corrupted context. Fine‑tuning on a 170 M masked diffusion improves infill F1 by 3.8–6.4 points at 10–20% corruption while keeping clean performance intact.

10. Binding-Specific Relearning Savings for Test-Time Continual Learning

    Zhe Zhou · arXiv:2610.22381

    AI summaryThe study shows that learning a specific factual binding in a language model yields a measurable relearning advantage at test time, speeding recovery by ~18% over controls and reducing updates needed, indicating current accuracy metrics underestimate lasting experience effects.

11. Deep Learning for Forest Plantation Mapping in Godavari Districts of Andhra Pradesh, India

    Snehal More, Anuj Karpatne, Valerie A. Thomas et al. · arXiv:2610.22415

    Small-area forest plantations play a vital role in the socioeconomic well-being of farmers in Southeast Asia. Most plantations are established on former agricultural land, often on land less suitable for agriculture. Plantations that are converted from natural forest have…

12. CGDT-CESN: Correlation-Guided and Dynamics-Trained Clustered Echo State Network for Multivariate Time-Series Prediction

    Hariharan Srinivasan, Masoom Bhargava, Sayan Gupta · arXiv:2610.22430

    AI summaryCGDT‑CESN trains a clustered Echo State Network where each cluster specializes in one variable and inter‑cluster links are guided by empirical correlations, then fine‑tunes those links. This yields up to 15‑20 % better multivariate forecasts on Twitter cascades and solar wind data, even with limited training data.

13. GPU-Accelerated Robust MA--RIS Pre-Optimization for Communication and Target Illumination

    Tran Le Dung, Ho Viet Duc Luong · arXiv:2610.22431

    AI summaryGPU-accelerated RAMP pre-optimizes MA geometry and RIS phases for ISAC, yielding a reusable Pareto set that balances expected worst-user data rate and transmit-side worst-target illumination across channel regimes. It achieves competitive robust Pareto quality without online re‑optimization.

14. An Adaptive-Horizon Graph Neural Operator for coupled surrogate morphodynamic modelling

    Ying Cao, Yifan Yang, Junqiang Xia et al. · arXiv:2610.22432

    AI summaryAdaptive-Horizon Graph Neural Operator (AH‑GNO) learns bed‑elevation changes from recent hydrodynamic and topographic states on unstructured meshes, allowing the history length and update interval to vary with flow. It matches Telemac2D/GAIA predictions while accelerating bed‑update from 12–1204 s to 0.08–3.34 s, a 150–360× speed‑up.

15. BOLT: Multi-Bottleneck Bound-Guided Learning for Flexible Job-Shop Scheduling with Time Lags and Resource-Coupled Transport

    Ziheng Zhang, Wei Zhang, Yunyi Zhao et al. · arXiv:2610.22433

    AI summaryBOLT constructs a tight lower bound from chain, capacity, and fleet relaxations for transport‑coupled job‑shop scheduling and uses its stepwise increase as a reward, guaranteeing a provable makespan floor. It outperforms penalty and dispatching‑rule baselines on large construction instances, achieving up to 13 % makespan savings and sub‑hour re‑planning time.

16. TRACE: Transformer Residual Activation Codebook Encoding for Communication-efficient Distributed Edge Inference

    Muhammad Azlan Qazi, Alexandros Iosifidis, Qi Zhang · arXiv:2610.22435

    AI summaryTRACE replaces post‑attention residuals with a vector quantizer, sending only codebook indices between edge devices so receivers can look up keys/values directly, eliminating downstream computation. This reduces communication by up to 22,624× and boosts ViT/ViViT inference speed by 1.27× with <1.9% accuracy loss.

17. Generative adversarial networks with Monte Carlo dropout uncertainty quantification for structural health monitoring: Fibre Bragg grating data augmentation comparison in Hawk T1A aircraft

    Alvaro Peiro, Marios Impraimakis · arXiv:2610.22436

    AI summaryGenerative adversarial networks with Monte‑Carlo dropout produce realistic structural response signals and quantify uncertainty. On a Hawk T1A aircraft dataset, the synthetic data match experimental measurements and improve damage‑classification performance, providing a useful digital‑twin for safety‑critical monitoring.

18. CF-NODE: Counterfactual Spatial Node Intervention for Test-Time Adaptation Under Domain Shift, Evaluated Across Five Diabetic Retinopathy Datasets

    Sabarish G, Deepak Parashar · arXiv:2610.22437

    AI summaryCF‑NODE adapts a frozen DR grading model to each test image by counterfactually attenuating the most influential spatial nodes and calibrating the grade distribution based on the stability of the response. It achieves higher AUC and macro‑F1 than TENT, EATA, SAR, and T3A across five external datasets.

19. Rank Measures of Fine-Tuning Updates Need a Noise Baseline

    Muhammad Usman Altaf · arXiv:2610.22438

    AI summaryThe paper shows that fine‑tuning weight updates have ranks only slightly below random noise, unlike previously thought. It demonstrates that rank‑based metrics lack meaningful baselines, and alignment claims need controls against non‑alignment data.

20. Sparse Autoencoders Can Induce Local Geometry for Mitigating Catastrophic Forgetting

    Kamil Faber, Katarzyna Filus, {\L}ukasz Pawela · arXiv:2610.22439

    AI summarySAGA uses a Sparse Autoencoder to create local low‑rank constraints on representation drift at each task boundary, guiding replayed examples and penalizing leading Jacobian directions. It outperforms ER, DER++, ER‑ACE, and SGD, improving Class‑IL accuracy by 12.15–4.79–3.22 points on CIFAR‑100, CUB‑200, and ImageNet‑R.

21. Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Guards for Scientific Discovery and Interpretability

    Tianyu Wu (Project Aether) · arXiv:2610.22445

    AI summaryThe paper introduces four operations to modify physical models and evaluates acceptance tests (NAR, PAR, CAR). It shows that null models can pass descriptive tests yet fail consequence checks, highlighting gaps that must be addressed for reliable automated scientific discovery.

22. Learnable Splines for Unstructured Neural Operators

    Vignesh Gopakumar, Zhuoyuan Wang, Ander Gray et al. · arXiv:2610.22446

    AI summaryLearnable spline neural operator that maps irregular meshes to a latent Cartesian grid via a sparse B‑spline transfer; achieves high accuracy on 3‑D engineering benchmarks with lower memory, supports zero‑shot super‑resolution, and provides rigorous error bounds via Bernstein hulls.

23. Lean4AgenticRL: Advancing Agentic RL with Verifiable Multi-level Process Rewards

    Ruida Wang, Gerald Friedland, Tong Zhang et al. · arXiv:2610.22448

    AI summaryLean4AgenticRL introduces multi‑level process rewards grounded in Lean4 formalization, enabling verification of low‑level toolcalls, mid‑level workflow stages, and high‑level outcomes. Its GRPO algorithm trains LAR‑27B, achieving 77.3% on TerminalBench‑2.1 and surpassing similar‑sized baselines by ~10%.

24. Where Prefill Expert Reduction Hurts MoE Models: A Few Prompt Tokens Carry Most of the Short-Answer F1 Loss

    Zhiyuan Wang · arXiv:2610.22454

    AI summaryReducing the number of routed experts during prompt prefill mainly hurts short‑answer QA because early prompt tokens become sensitive; at half the experts, F1 drops by roughly one‑third of the loss when only the first few prompt tokens are routed. Restoring full‑routing only for those few tokens recovers most of the F1 loss.

25. Progressive Misalignment Can Emerge from Benign Recursive Self-Distillation

    Karim Abdel Sadek, Federico Di Gennaro, Fanny Yang · arXiv:2610.22467

    AI summaryThis paper demonstrates that recursively training models on their own safe generations using benign auxiliary cues causes progressive misalignment, leading to increased vulnerability to harmful queries across multiple model families.

26. Has GPT-6-Astra Achieved Embodied Intelligence?

    SenseNova, Joint Research Team, : et al. · arXiv:2610.22471

    AI summaryGPT‑6‑Astra shows unprecedented performance on 12 embodied robotics benchmarks, achieving high success rates with in‑context learning and a System 2/1 control loop. However, it still struggles with costly demonstration collection, real‑time constraints, and long‑horizon robustness, limiting practical embodied intelligence.

27. Upper bounds on the number of causal pieces in spiking neural networks

    Dominik Dold, Dennis Elbrächter, Philipp Petersen · arXiv:2610.22484

    AI summaryThe paper upper‑bounds the number of causal pieces a feedforward single‑spike neural network can generate. It shows a general poly‑Bernoulli combinatorial bound and tighter hyperplane‑based bounds for various integrate‑and‑fire and ReLU‑type neuron models.

28. Generative Value Classifier

    Alexander Swerdlow, Perry Dong, Chelsea Finn · arXiv:2610.22497

    AI summaryThis paper introduces the Generative Value Classifier, which reformulates Q-learning by predicting action likelihoods given returns rather than just predicting returns from state-action pairs. This approach improves performance on both online and offline reinforcement learning tasks compared to standard discriminative methods.

29. Predicting Privileged Information Effectiveness for On-Policy Self-Distillation

    Kaijian Zou, Emiliano Penaloza, MohammadHossein Rezaei et al. · arXiv:2610.22512

    AI summaryThis paper evaluates privileged information configurations for on-policy self-distillation in LLM agents and introduces KL-based metrics to predict their training utility. The proposed metrics reliably select effective privileged information, improving downstream agent performance over random selection and standard baselines.

30. Towards Reconstructing Geographically Diverse Architecture with 3D Foundation Models

    Nicky Kriplani, Yiwen Zhang, Angelina Wang et al. · arXiv:2610.22514

    AI summaryThe authors introduce ArchWorld, a benchmark of geographically diverse architectural landmarks, to analyze performance disparities in 3D foundation models. They find that model errors stem from region-specific architectural features and propose an intervention scheme that improves overall accuracy while reducing geographic performance gaps.

31. Source-Aware Dynamic Pricing with Strategic Buyers

    Jinjiang Zou, Yang Li, Zemin Zheng · arXiv:2610.22516

    AI summaryThis paper introduces a source-aware dynamic pricing framework that handles both verified and strategically manipulated self-reported buyer features. The proposed SALE policy achieves a sublinear expected regret bound by combining localized exploration with strategic correction.

32. Learning smooth dynamics via Kinematic Interpolants

    Yiheng (Roger), Jiang, Youssef Marzouk et al. · arXiv:2610.22519

    AI summaryThe paper introduces Kinematic Interpolants, a framework leveraging position and velocity measurements to infer nonlinear dynamics. It achieves superior reconstruction of dynamical systems while requiring a fraction of the training time compared to existing baselines.

33. Learning the Shared Latent Structure of Cryo-EM Inverse Problems

    Nithin Chalapathi, Theo Sternlieb, Ethan Han et al. · arXiv:2610.22523

    AI summaryThis paper introduces EMber, a family of vision transformers pre-trained on millions of cryo-EM images using complementary global and masked patch objectives. EMber improves conformational and compositional discrimination by over 14 percentage points and enables reference-quality reconstructions in fewer epochs.

34. Don't Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks

    Md Farhan Ishmam, Minh Pham Dinh, Kenneth Marino · arXiv:2610.22539

    AI summaryWebMix uses a mixture hypernetwork that composes pre‑trained low‑rank adapters via a scouting run to adapt web agents to unseen interfaces without retraining. It outperforms in‑context learning and LoRA baselines, improving TimeWarp domain transfer by ~8 % and matching specialist performance.

35. HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments

    Ron Keinan, Amir David Nissan Cohen, Reut Tsarfaty · arXiv:2610.22542

    AI summaryHeRE builds a 1.5‑million‑instance Hebrew relation‑extraction benchmark by denoising noisy knowledge‑graph alignments with LLM‑guided NLI. Gemma‑4‑31B‑it scores F1 = 0.915, matching GPT‑5.4, and the dataset enables scalable, language‑agnostic RE evaluation.

36. On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction

    Eliezer Shlomi, Eliron Rahimi, Avi Mendelson et al. · arXiv:2610.22543

    AI summarySpeculative Decoding for Diffusion Language Models (SPOT) predicts the order of token generation from prior model outputs, enabling a tree of candidate continuations that the target verifies in a single pass. This reduces target passes by up to 6.8× and boosts throughput by 3.2× over standard decoding.

37. When Plans Change, Remember: Replanning-Guided Visual Memory for Vision-Language-Action Models

    Victoria Smelova, Daniil Kazachkov, Georgiy Kazantsev et al. · arXiv:2610.22556

    AI summaryThis paper introduces RePlanMem-VLA, a memory management method for vision-language-action models that uses policy plan revisions to selectively store task-relevant visual observations. The resulting system successfully preserves critical information while eliminating redundancy across long-horizon robotic tasks.

38. Deterministic Bounds on Accuracy and Stability of Metric-Space Regression and Classification

    Nina Maria Gottschling, Jakob Gawlikowski, Michele Caprio · arXiv:2610.22564

    AI summaryThe paper proposes a geometric framework that uses the average kernel size to give deterministic, non‑asymptotic bounds on empirical error and accuracy for regression and classification, independent of the model class, and demonstrates these bounds on tabular and image datasets.

39. Coverage Cliffs in Off-Policy Quantile Certification

    Pauline Bourigault, Emmanuelle Bourigault · arXiv:2610.22565

    AI summaryTailCert certifies a policy’s upper-tail quantile from logged bandit data, providing a PAC‑Bayes lower‑confidence envelope over thresholds; it guarantees sample complexity linear in overlap W, logarithmic in policy‑search, and quadratic in the inverse safety margin, matching lower bounds.

40. Structure-Aware Spatial Kernels for Gaussian Process Distributed Port-Hamiltonian Systems

    Tianyi Yang, Thomas Beckers · arXiv:2610.22575

    AI summaryStructure‑aware spatial kernels for Gaussian‑process distributed port‑Hamiltonian systems decouple state and spatial dependencies, removing irrelevant coordinates and allowing a modular spatial kernel choice. Experiments on a nonlinear string show that this decoupling improves reconstruction, and matching the spatial kernel to stiffness profiles yields large accuracy gains.

41. Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations

    Devin Han, Hershraj Niranjani · arXiv:2610.22576

    AI summaryThis paper introduces embodiment-classifier guidance to adapt frozen flow-based vision-language-action policies to new robot bodies without demonstrations or parameter updates. Using gradients from a lightweight classifier trained on a few calibration tasks, the method improves average target success from 72% to 80% on simulation benchmarks.

42. Noise-Robust Forced Alignment for Spontaneous Speech

    Laurin Wagner, Mario Zusag · arXiv:2610.22583

    AI summaryImproves forced alignment on spontaneous noisy speech by replacing MFCCs with upsampled WavLM features, adding explicit filler/non‑lexical phone units, and a discriminative projection trained on clean/noisy pairs. Achieves 20.6 ms (Buckeye) and 76.9 ms (FluencyBank) MAE, 19–17 % better than 10 baselines.

43. Initiating Removal from Operational Symptoms in Federated Vehicular Networks

    Zihao Ding, Jun Huang, Qiang Duan · arXiv:2610.22584

    AI summaryThe paper introduces \method{}, a federated unlearning method that neutralizes operational symptoms in vehicular FL by projecting errors into a subspace of recent increments and backward‑scoring contributions. It reduces residual harm by 43–61% versus baselines while keeping utility within 0.36 points of full retraining.

44. Flat Loss, Evolving Features: Grokking in Modular Addition

    Akash Kumar, Karthik Sankaranarayanan · arXiv:2610.22593

    AI summaryReLU networks on modular addition can maintain 100 % training accuracy while training cross‑entropy stays flat above zero, and test accuracy rises above 90 %; this is provable for one‑hidden‑layer networks and observed empirically in deeper nets and transformers.

45. Diagnosing GNN Failure Modes with Layer-Wise Probe Entropy

    Shanzi Bao, Mar Gonzàlez I Català, Pietro Liò · arXiv:2610.22595

    AI summaryThey propose freezing a trained GNN, fitting linear probes per layer, and tracking probe accuracy with normalized entropy. The entropy curves reveal when smoothing is helpful, neutral, or harmful, distinguishing heterophily mismatch early and oversmoothing in deep models.

46. Can LLMs predict the outcomes of machine learning experiments?

    Ayush Bansal, Aldo Pareja, Sean Klein et al. · arXiv:2610.22599

    AI summaryLarge language models can predict the performance of their own proposed ML experiments better than chance but explain only ~50 % of outcome variance; they systematically underestimate effect sizes, show task‑dependent accuracy, and are largely indifferent to the coding harness used.

47. Rarity of Singletons Matters & Supplementary Materials

    Pragatheeswaran Vipulanandan, Kamal Premaratne, Dilip Sarkar · arXiv:2610.22604

    AI summaryThe paper introduces a fragility‑aware framework that measures the instability of singleton counts and adaptively regularizes missing‑mass and unseen‑species estimators. It achieves closed‑form, concave solutions that outperform classical methods like Good–Turing and Good–Toulmin in small‑sample, long‑tailed settings.

48. Shift-Reliance Alignment: Ranking Vision-Language Models by the Cues They Trust

    Faisal Mohammad, Mohamed Younis · arXiv:2610.22610

    AI summaryShift-Reliance Alignment measures how much a VLM relies on cues like context or text by editing source images, then uses that reliance scaled by the shift (proportion of target questions misled) to rank models. SRA picks models that reduce error on misleading target sets, achieving up to 18.8% error versus 30.4% with source-only ranking.

49. Noise-shaped Sampling and Annealing for Scalable On-the-edge Training of Binary and Ternary AI Models

    Zihao Chen, Zhili Xiao, Rui-Jie Zhu et al. · arXiv:2610.22621

    AI summaryThe paper introduces $\Sigma\Delta$-Sampler, an optimizer that embeds stochastic sampling and annealing within a noise‑shaping loop to align Ising solvers with generalization‑oriented training. It achieves near full‑precision accuracy on large ternary models while cutting training memory use by 50% and speeding training by an order of magnitude.

50. Understanding Stability of Outer Nesterov Momentum on Stochastic Pseudo-Gradients

    Nishant Sharma, Laura Gomezjurado Gonzalez, Ganesh Talluri et al. · arXiv:2610.22623

    AI summaryThe paper analyzes how outer Nesterov momentum, local steps, and worker averaging affect stability of distributed training with pseudo-gradients. It derives a mean‑square stability criterion for random quadratics, showing inner batch size and worker count trade‑offs, and validates the theory on neural networks.

51. Network-Wide Freeway Speed Nowcasting and Forecasting from Sparse Fixed Sensors with a Probe-Supervised Graph Convolutional Recurrent Network

    Ran Tu, Zihao Li, Arash Mirzaei · arXiv:2610.22630

    AI summaryA graph‑convolutional recurrent network called GCRN+ predicts hourly freeway speeds on all 4,127 links of the Dallas–Fort Worth network using only sparse fixed sensor data and probe‑vehicle supervision, achieving RMSEs of 2.21 mph (nowcasting) and 2.39 mph (next‑hour) on unsensed links.

52. Alert-List Concordance for Evaluating Imputation Strategies in ICU Sepsis Prediction

    Priyanka Siddappa, Muntaqim Ahmed Raju, Shifat Haider Al Amin et al. · arXiv:2610.22635

    AI summaryThe study introduces alert‑list concordance, measuring Jaccard similarity of top‑risk patient lists across imputation methods. It finds that many methods have similar AUROC yet differ significantly in patient selection, with ForwardFill showing the lowest overlap.

53. The Ephemeral Cost of Switching Optimizers Across LLM Training Stages

    Xinyu Zhao, Davis Wertheimer, Ashutosh Sharma et al. · arXiv:2610.22650

    AI summaryThe study trains 1.2B Llama‑3 models with varying mixes of AdamW and Muon optimizers, finding no significant penalty for switching optimizers between pretraining and fine‑tuning. Muon remains competitive on downstream benchmarks even when the model was pretrained with AdamW.

54. Is a Linear Probe Evidence of a Linear Representation?

    Nadir Kutluozen, Christian Internò, David Klindt · arXiv:2610.22655

    AI summaryLinear probes can be misleading: they succeed on any rich feature map, even random kernels, because they only capture a linear subspace. The paper introduces reverse predictivity, spectral concentration, and OOD tests to show pretrained transformers encode more complex, curved structures, not a purely linear representation.

55. Edit Discovery in Human-AI Interactions

    Alireza Mirrokni, Shayan Kiyani, George Pappas et al. · arXiv:2610.22661

    AI summaryThe paper defines Edit Discovery Optimization (EDO) to decide when a costly editor should inspect AI responses, formulating a cost-adjusted threshold policy. Its EDIT algorithm learns online, respects an editor budget, and outperforms batch LLM selection on QA benchmarks.

56. AlphaBet: Reinforcement Learning for LLM Event Forecasting Agents

    Liqin Ye, Rongzhi Zhang, Xin Liu et al. · arXiv:2610.22684

    AI summaryAlphaBet is a reinforcement learning framework that trains LLM agents to iteratively search for evidence and forecast outcomes across multiple dates. By supervising models with proper scoring rules over resolved events, it significantly reduces forecasting error on benchmark datasets.

57. Quantum-Classical Neural Networks for Solving Lindblad Master Equations

    Rafael Florencio Díaz, Julio Guerrero García, Elvira Romera Gutiérrez · arXiv:2610.22691

    AI summaryThis paper introduces a hybrid quantum-classical neural network to solve Lindblad master equations while naturally preserving density matrix properties. The authors prove that the classical component enlarges the hypothesis class and improves trainability compared to purely quantum models.

58. Optimal approximation of smooth functions by ReLU neural networks

    Thomas Nagler · arXiv:2610.22693

    AI summaryThis paper establishes optimal ReLU network constructions for uniformly approximating Hölder functions across all dimensions and smoothness levels. The resulting architectures achieve the information-theoretic optimal approximation error and minimax regression rates without extraneous logarithmic factors.

59. Chance-Constrained Personalized LLM Routing with Online Preference Learning

    Ameya Anjarlekar, Rasoul Etesami, R Srikant · arXiv:2610.22695

    AI summaryThis paper introduces a constrained contextual bandit algorithm for personalized LLM routing that minimizes inference cost while guaranteeing response quality based on online preference feedback. The method achieves sublinear cost regret and high-probability satisfaction guarantees compared to an optimal oracle.

60. Data Provenance Aggregation and Structural Descriptors for Query-Template Identification

    Paulo Pintor, Rogério Costa, José Moreira · arXiv:2610.22705

    AI summaryThis paper uses aggregated data provenance to build query‑level profiles and compact Provenance Structural Descriptors (PSD). Experiments on 12,500 TPC‑DS queries show that these provenance footprints accurately identify query templates, with PSD achieving high accuracy even with limited training.

61. Normalizing Flow Maps

    Francesco M. Ruscio, Timothy James Hitge, Fred Zhangzhi Peng et al. · arXiv:2610.22711

    AI summaryNormalizing Flow Maps learn invertible dynamical system integrators that give exact likelihoods. On CIFAR‑10, ImageNet‑64 and Boltzmann‑generator benchmarks they match or beat state‑of‑the‑art normalizing flows in sample quality while enabling tractable likelihood evaluation.

62. Spectral Limits of Polynomial Graph Filters on the Long-Range Ising Model Benchmark

    Amin Najafgholizadeh, Mohsen Afsharchi · arXiv:2610.22713

    AI summaryPolynomial‑adjacency filters hit an irreducible error floor on the Long‑Range Ising benchmark, independent of degree; a 50‑coefficient strided filter trained on labels reduces test error 52× below prior state‑of‑the‑art, while a dilated convolution surpasses published results at 256 and 1024 nodes and sets new in‑domain benchmarks for larger lattices.

63. CoLU: Confidence-Weighted Layerwise Unlearning for Large Language Models

    Jingkai Guo, Ziyi Zhao, Chaitali Chakrabarti et al. · arXiv:2610.22721

    AI summaryCoLU applies confidence‑weighted forgetting and retain regularization to a single transformer layer, and uses Brief Layer Screening with a normalized forgetting‑retention score to select layers, achieving 60–80% faster runtime and superior forgetting and retention on TOFU, MUSE Books, and News while saving 77–80% GPU memory.

64. The LLM-Attacker: A Unified Red-Team Framework for Privacy Evaluation of Split and Distributed LLM Systems

    Georgios Politis, Evangelos Pappas · arXiv:2610.22722

    AI summaryThis paper introduces LLM-Attacker, a unified framework for standardizing privacy evaluations and text reconstruction attacks in split and distributed language model systems. Experiments using the framework demonstrate that granting attackers access to training gradients significantly increases both token and exact-document recovery rates.

65. OrderFlowDiT: Transport-Regularized Diffusion Transformer for Probabilistic Limit-Order Supply Forecasting

    Jung Hun Phee, Tucker Balch · arXiv:2610.22727

    AI summaryOrderFlowDiT trains a diffusion transformer on bid/ask volume fields using an unbalanced optimal‑transport loss, improving joint price–volume prediction. Across 30 U.S. equities, UOT reduces transport cost by 45.8 % and energy score by 26.1 % over MSE‑only training.

66. Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns

    Kevin Foley, Jonathan Hartadi, Shivesh Prakash et al. · arXiv:2610.22745

    AI summaryThe authors introduce NoLBERT, a financial news encoder trained strictly on historical text to prevent hindsight bias. They find that while NoLBERT yields stable topics, text-based models ultimately fail to outperform price-only benchmarks in out-of-sample stock return forecasting.

67. Native Geometry vs. Linear Recoverability: How Alignment Changes Platonic Scaling

    Angus Hoy, Austin Lutterbach, Jasper Chong et al. · arXiv:2610.22749

    AI summaryThis paper investigates how model scale affects representation similarity across different datasets and architectures. The main result is that larger models improve linear recoverability of representations much more clearly than they establish convergence toward a shared native geometry.

68. Refusal Is Not Robustness: Measuring Harmful Reachability in Autoregressive Language Models

    Parv Kapoor, Teresa Fortes, Claudia Mamede et al. · arXiv:2610.22754

    AI summaryThis paper introduces bounded continuation analysis to measure the probability that an autoregressive language model reaches harmful generations despite outputting a safe response. The method reveals substantial hidden risk in models that appear robust under standard attack success rate evaluations.

69. Shape-Aware Neural Time Fields for Rigid-Body Motion and Manipulation Planning

    Yanming Zhou, Yuchen Liu, Ahmed H. Qureshi · arXiv:2610.22762

    AI summaryThis paper introduces Shape-Aware Neural Time Fields, a physics-informed neural framework that learns reusable cost-to-go representations for rigid-body motion planning. The method achieves higher success rates and inference efficiency than existing neural baselines while matching classical planner performance at significantly lower computation times.

70. Acquisition Function-Guided Generation for Bayesian Optimization over Discrete Spaces

    Tommaso Ancilli, Thomas Wedenig, Henry Moss et al. · arXiv:2610.22763

    AI summaryThis paper introduces acquisition function-guided generation for discrete Bayesian optimization, using a diffusion bridge to enable gradient-based sampling from a pretrained generative model. The proposed method outperforms or matches state-of-the-art baselines across various molecular design benchmarks.

71. One Communicator, Three Vendors: Extending the MoA Predictive Theory to a Real NVIDIA+AMD+CPU MPI Ensemble

    Lenore Mullin, Gaetan Hains · arXiv:2610.22769

    AI summaryThis paper extends Mixture-of-Agents predictive theory to a multi-vendor MPI ensemble uniting NVIDIA, AMD, and CPU ranks. The authors demonstrate successful cross-architecture communication and validate launch-geometry models, revealing significant performance and timing discrepancies across hardware vendors.

72. Unit-level aggregation breaks coverage in causal foundation models

    Gabe Schulman, Lucius E. J. Bynum, Yoav Wald · arXiv:2610.22770

    AI summaryCausal foundation models struggle with group-level uncertainty quantification because they assume unit-level errors are independent, ignoring correlations from posterior uncertainty over the data-generating process. Jointly predictive neural process models fix this failure and restore correct coverage.

73. Reusing Calibration Gains across Model Representations

    Xu Cao, Wenxuan Xiao · arXiv:2610.22771

    AI summaryThis paper shows that calibration updates learned for one model representation can be directly transferred to other representations of the same pretrained model without retraining. The method retains up to 99% of the text-loss improvements while preserving native quantization formats and storage layouts.

74. Anytime Contracts for AI Services

    Boris Velasevic, Eden Saig, Eric Mazumdar · arXiv:2610.22781

    AI summaryThe paper adapts contract theory to AI services that can be terminated anytime, showing that optimal payment contracts must be proportional to anytime-valid hypothesis tests using e-values. It presents an efficient algorithm for computing such contracts and demonstrates its effectiveness on synthetic and real LLM data.

75. Next Token or Vector? Learning How Far to Decode

    Sai Sreenivas Chintha, Chidera Biringa, Rishab Sharma et al. · arXiv:2610.22788

    AI summaryCADENCE learns both how many tokens to draft and what to draft, using Gaussian latent vectors and a router that adapts span length to context; it achieves up to 1.3× higher throughput than greedy decoding while maintaining high draft acceptance.

76. Ori: Drift-Aware Streaming $k$-PCA for Very Low-Rank LLM Training

    Xinyu Luo, Jincheng Zhou, Mengbo Wang et al. · arXiv:2610.22794

    AI summaryThis paper introduces Ori, a low-rank LLM optimizer that uses streaming PCA to track slow gradient subspace drift in real time. Ori significantly reduces the performance gap between low-rank and full-rank training across various language model tasks.

77. Reasoning Reranking with Semantic IDs as the Sole Item Representation

    Zhenwei Tang, Jintao Liu, Di Bai et al. · arXiv:2610.22797

    AI summaryThis paper introduces ReRankID to evaluate reasoning rerankers using only compact Semantic IDs instead of text. Results show that targeted post-training enables SID-only models to match text-based LLMs with much shorter prompts, despite producing less faithful item descriptions.

78. Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition

    Se Un Park, Kyungkeon Chung, Huigeol Kim et al. · arXiv:2610.22811

    AI summaryChannel replay pruning removes 60–70% of YOLOv5/8 parameters without accuracy loss and outperforms other methods; a 5‑layer 18k‑parameter reader with 4‑bit weights/8‑bit activations matches FP32 accuracy, yielding 95.2% sign detection at 94 fps on an automotive SoC.

79. Causal Edit Serialization

    Louis Robinson · arXiv:2610.22820

    AI summaryCausal Edit Serialization lets language models express revisions (insert, delete, replace, move) in a single causal stream using pointer tokens with source addresses. CES models match autoregressive quality while performing complete edits and improving downstream prediction.

80. Partial-Revelation Panel Forecasting: Predictive Information Gain, Allocation, and Coarsening-Consistent Operators

    Bin Wen, Dai-Qiao Zhang, Tien-Ping Tan · arXiv:2610.22831

    AI summaryPartial-revelation panel forecasting predicts full count distributions for the next dialogue turn from only a subset of annotator votes, using coarsening-consistent Dirichlet-multinomial operators. It achieves modest NLL improvements (≈0.4 nats on IEMOCAP) and shows that cross‑turn and same‑item gains are distinct.

81. Benchmarking and Enhancing Agentic Causal Discovery

    Pingchuan Ma, Rui Ding, Shuai Wang · arXiv:2610.22846

    AI summaryResearchers benchmark LLM agents on causal discovery, showing that agents can improve results beyond standard algorithms by inspecting data. GPT‑6.1 Sol outperforms the best algorithm on all six synthetic families, raising adjacency F1 by up to 14 % on the hardest datasets.

82. ToPA: Block-wise Toeplitz Adaptation for Expressive and Efficient Fine-Tuning

    Sicong Li, Qianqian Xu, Zhiyong Yang et al. · arXiv:2610.22883

    AI summaryToPA fine‑tunes large models by inserting block‑wise Toeplitz matrices into weight updates, allowing shorter, stable multiplicative chains that approximate arbitrary matrices and outperform low‑rank or sparse PEFTs across 21 NLP and CV benchmarks.

83. Learning to Be Guided: Shaping Latent Geometry for Classifier-Free Guidance

    Haoming Liu, Junhao Zhu, Shaoyi Zheng et al. · arXiv:2610.22885

    AI summaryThey analyze how the arrangement of discrete class embeddings in latent space—“class geometry”—affects classifier‑free guidance, and propose Prototype Regularization, a lightweight cosine‑prototype classifier that improves guided image generation, lowering guided gFID on ImageNet from 3.31 to 2.28.

84. Batched Contextual Reinforcement II: Does RLVR Need Long Rollout Windows?

    Bangji Yang, Hongbo Ma, Ruihan Guo et al. · arXiv:2610.22887

    AI summaryBatched Contextual Reinforcement explores concurrent verifiable problems to avoid long rollout windows in RLVR. CCR trains models with a shared budget, cutting training tokens by ~80% while maintaining accuracy, enabling efficient short‑window reinforcement learning.

85. What Accuracy Hides In Learned Transformer Programs

    Jonathan Lee, Narmeen Oozeer, Philip Quirke · arXiv:2610.22891

    AI summaryTransformer Programs expose learned logic as readable code, revealing that token‑level accuracy can mask wrong outputs. Program inspection guided data augmentation, improving sequence accuracy and uncovering hidden failures that standard metrics missed.

86. Competence-Based Evaluation of Logical Reasoning in Large Language Models

    Jize Jiang, Chengxiang Zhai · arXiv:2610.22894

    AI summaryThe paper proposes a competence‑based evaluation for logical reasoning in LLMs, using parameterized generators and answer‑preserving variations to test generalization. It finds that even strong GPT and Qwen models fail to consistently apply the same reasoning competence across problem families.

87. When to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents

    Chenle Chen, Yangbo Wei, Zhijie Zheng et al. · arXiv:2610.22897

    AI summaryReCAP decides when a long-horizon agent should keep, compress, or discard past failures. It outperforms full-history KEEP, reducing prompt cost by 2.13% while preserving 91.9% of continuation success across five backbones.

88. A Feature Selection Layer Approach for Multi-Class Classification on Neural Networks

    João Pedro Silveira e Silva, Márcio Dorn, Bruno Iochins Grisci · arXiv:2610.22898

    AI summaryThe paper extends the Feature Selection Layer to embed label‑specific relevance learning in neural networks. Experiments show it matches existing methods for global feature importance and outperforms them in identifying how each feature influences individual class predictions.

89. FederatedBytes: Tokenizer-Free, Byte-Level Small Language Models for Efficient Multimodal Federated Learning

    Ali Sarabadani, Massimo Piccardi, Mehran Abolhasan et al. · arXiv:2610.22906

    AI summaryFederatedBytes trains a byte‑level language model with a linear attention mechanism for federated learning of NLP and multimodal tasks. It achieves higher accuracy than ByT5 and LSTM on multilingual sentiment and ScienceQA datasets while cutting memory and communication costs.

90. Where's the Indirect Evidence? Investigating Grammatical Generalization with Influence Functions

    William Dai, Zhenghao Herbert Zhou, Maya Viswanathan et al. · arXiv:2610.22919

    AI summaryUsing influence functions, the paper identifies 60,000 “indirect evidence” sentences that significantly affect grammatical generalization in language models, showing they improve held‑out accuracy by about 12 points more than random subsets, and revealing linguistic category structure.

91. Scalable Levenberg-Marquardt Training Framework for Deep Learning

    Xianghao Wang, Johan Barthélemy, John Taylor · arXiv:2610.22925

    AI summaryLeMA rederives LMA to avoid a parameter‑sized Jacobian, shards the Jacobian across GPUs, and enables exact second‑order steps for billion‑parameter networks. On 8 GPUs it attains 7.97× speedup and 99.7% scaling for a 1B‑parameter CNN.

92. Conformal Policy Selection: When to Deploy AI Decisions for Targeted Intervention

    Junyu Cao, Ruijiang Gao, Maytal Saar-Tsechansky et al. · arXiv:2610.22930

    AI summaryConformal Policy Selection tests each individual’s predicted outcome against a threshold and selects those exceeding it while controlling false discovery rate. It guarantees finite‑sample FDR control and is doubly robust, offering a safety gate for deploying AI policies in high‑stakes settings.

93. Many Biases, One Circuit: Post-Training Tunes a Pretrained Cue-Routing Mechanism

    Prakhar Gupta, Terry Jingchen Zhang, Zhijing Jin · arXiv:2610.22937

    AI summaryThe paper identifies a shared cue‑routing circuit in transformer attention heads that drives multiple-choice biases. Post‑training fine‑tuning and RL strengthen these heads, suggesting that mitigation can target this specific circuit.

94. Aligning Language Models Towards Counterfactual Invariance

    Basar Demir, Emily Shen, Ksheeraj Sai Vepuri et al. · arXiv:2610.22942

    AI summaryThe paper shows that reinforcement learning alignment increases counterfactual bias in language models. It proposes a policy‑gradient method with group‑level rewards on matched counterfactual prompts, reducing demographic differences while maintaining overall response quality.

95. Large Diffusion Models Need More Than EMA

    Chuyu Hu, Huanran Chen, Shitong Shao et al. · arXiv:2610.22945

    AI summaryLarge diffusion models benefit from learning‑rate decay and Schedule‑Free optimization rather than constant LR with EMA alone. The study shows decay+EMA improves baselines but is timing‑sensitive, while Schedule‑Free consistently reduces parameter variability and yields better objectives and generation quality.

96. On the tradeoff between robustness and geometric accuracy in adversarial training

    Rachel Morris, Ryan Murray, Kerrek Stinson · arXiv:2610.22955

    AI summaryShows that for high-complexity non‑parametric classifiers, robustness to an adversary of size ε only reduces geometric accuracy by at most O(√ε) (or O(ε) with stronger regularity). This gives a tight trade‑off between robustness and accuracy for any norm.

97. What Shapes Brain Alignment of Large Language Models during Dialogue? Insights from Language Production and Comprehension

    Yuko Nakagi, Yuya Chiba, Yasue Kishino · arXiv:2610.22961

    AI summaryThe study compares fMRI-based brain alignment of LLMs during dialogue, finding distinct patterns for production versus comprehension that correlate with the models’ intrinsic dimension in relevant brain areas, and shows that linguistic and communicative feature contributions differ by processing mode, layer, and region.

98. Towards Hyperparameter Transfer for Differentially Private Optimization

    Tianze Wang, Zhiqi Bu, Linjun Zhang · arXiv:2610.22976

    AI summaryDP‑µP extends maximal update parametrization to differentially private training, deriving a signal‑to‑noise ratio condition for layer‑wise learning‑rate scaling. It yields practical scaling rules for DP‑SGD, DP‑Adam, and DP‑Muon that transfer learning rates across widths, matching or surpassing tuned baselines.

99. How Murky Before a VLM Knows It Cannot See? Accuracy, Confidence and Abstention on a Physical Visibility Axis

    Mochamad Rizal Fauzan, Irgi Surya · arXiv:2610.22981

    AI summaryThe study measures VLM accuracy, internal, verbalized confidence, and abstention on a physical visibility axis (fog and underwater) for 1000 COCO scenes. Accuracy drops to chance at 12–16 m fog, yet high confidence remains, revealing a blind‑spot width of at least 6× visibility.

100. PersonaCUABench: Benchmarking Computer-Use Agents Serving Diverse Personas

     Haoming Li, Defu Cao, Xu Hu et al. · arXiv:2610.22990

     AI summaryPersonaCUABench evaluates computer‑use agents by letting 44 personas rephrase tasks and pose constraints, revealing a 36% drop in completion and lower satisfaction. The study shows agents often misinterpret or ignore user‑specified requirements.

101. TALLOP: Joint Optimization of Local Training and Communication Compression for Time-Constrained Federated Learning

     Honoka Anada, Yafei Ou, Shinya Takamaeda-Yamazaki · arXiv:2610.22996

     AI summaryTALLOP jointly optimizes local training workloads and communication compression under per-round time constraints in federated learning. By predicting training progress and compression degradation, it minimizes post-compression loss to achieve faster time-to-accuracy in heterogeneous environments.

102. A Label Noise Robust Algorithm for Graph Data using Neighborhood Label Distribution

     Pintu Kumar, Nandyala Hemachandra · arXiv:2610.22998

     AI summaryThis paper proposes LENS, an end-to-end graph neural network framework that identifies and corrects noisy labels by measuring changes in neighborhood entropy. LENS achieves state-of-the-art node classification accuracy while running significantly faster than existing robust methods.

103. Controlling the Diversity-Quality Trade-Off of LLMs by Merging Models Before and After Alignment

     Yasuaki Sumita, Tomoharu Iwata, Toshiyuki Tanaka · arXiv:2610.22999

     AI summaryThis paper introduces a training-free layer-wise model merging technique that combines aligned and base LLMs to control the diversity-quality trade-off. By mixing lower and higher layers, the merged models achieve superior semantic and lexical diversity compared to standard alignment without sacrificing quality.

104. A Minimal ReLU Classifier on Non-Linearly Separable Data: Convergence and Implicit Bias at and beyond the Edge of Stability

     Quan Nguyen, Robin Yadav, Param Mody et al. · arXiv:2610.23007

     AI summaryThis paper analyzes a two-neuron ReLU network trained on non-linearly separable data with large learning rates. It proves that gradient descent converges to a max-margin solution in the edge-of-stability regime and introduces an adaptive schedule achieving a faster loss decay rate.

105. "Deep lifts": Neural networks can generalize even with the dumbest optimizer

     Sunil Madhow, Yu-Xiang Wang · arXiv:2610.23009

     AI summaryThis paper proves that randomly sampled interpolating neural networks generalize well despite using no standard optimizer, achieving a rate dependent on the true underlying network size rather than the overparameterized architecture. This occurs because simpler networks witness an exponentially larger volume of interpolating parameters in the overparameterized space.

106. Distinguishability and Recovery for a Doubly Sparse Spiked Wigner Model

     Ioana Dumitriu, JD Flynn, Zhichao Wang · arXiv:2610.23033

     AI summaryThe paper shows that a leave‑one‑out quadratic thresholding method recovers the support of a sparse rank‑one signal in a spiked Wigner model with independent sparse noise, and sign synchronization or spectral refinement yields exact or strong recovery, giving sparsity‑SNR tradeoffs.

107. Towards Looped Models Done Right, Part I: Topology, Input Injection, Recurrent-State Design

     Benhao Huang, Chufan Shi, Junlin Chen et al. · arXiv:2610.23036

     AI summaryThe paper studies looped Transformer designs, isolating recurrence placement, input injection, and state layout. It shows that the Huginn-style prelude‑recurrent‑coda architecture outperforms the Ouro architecture and feedforward MoE baselines, achieving better accuracy with fewer resident parameters.

108. GR-OPD: Group-Relative Sequence Credit for On-Policy Distillation

     Yushu Wang, Renkai Yang, Jingcao Xu et al. · arXiv:2610.23042

     AI summaryGR-OPD introduces a group-relative sequence credit method for on-policy distillation that decouples teacher feedback from credit assignment. It outperforms vanilla OPD across six mathematical reasoning benchmarks and maintains stability under mismatched thinking patterns.

109. DIFFLET: A Fast Inference Engine for Diffusion Transformer Models on AWS Trainium

     Bin Ma, Min Yu, Vijay Kodamalla et al. · arXiv:2610.23047

     AI summaryDifflet compiles diffusion transformer pipelines on AWS Trainium2 into a shared backend, splits attention heads across cores, and uses host‑controlled caching to skip redundant denoiser runs. It achieves 1.2–4.3× speedups over PyTorch/XLA and 1.5–3.4× faster resident serving.

110. Unifying Embodied Foundation Models via a Model-to-Policy Formulation

     Xinyi Hong, Mingyang Ren · arXiv:2610.23053

     AI summaryUnifies VLA and WAM robot learning by modeling them as stochastic kernels with conditional output distributions and action‑realization kernels that produce task‑conditioned policies. Demonstrates the framework on OpenVLA and DreamZero, handling both autoregressive action tokens and joint video‑action generation.

111. Path to Prior-Data Fitted Relational Models via Learned Latent Multi-table Representations

     Mohammad Sadeq Abolhasani, Viswanath Ganapathy · arXiv:2610.23073

     AI summaryPretrains a graph‑based relational model, Griffin, on synthetic databases with curriculum learning, then freezes the encoder and feeds its representations into a frozen in‑context predictor. On 16 real relational tasks, curriculum‑pretrained encoders outperform real‑data pretraining and baselines.

112. Where Latents Belong and How They Flow: Joint Identification of Structural Roles

     Xintong Li, Kun Zhang, Yujia Zheng · arXiv:2610.23074

     Latent factors in dynamical systems can play different roles across observations and time. For example, an idea in a group meeting may be expressed by only some participants, inherit information from earlier discussions, and shape what is said next. We formalize this notion as…

113. Towards a GraphNARX model for structural dynamics

     George Tsialiamanis, Keith Worden · arXiv:2610.23103

     AI summaryGraphNARX is a graph‑based NARX model that uses nodal attributes of a structure’s graph to predict next‑step displacements, enabling system identification and virtual sensing. It generalises across different geometries and outperforms baselines on diverse structural sizes.

114. Domain-specific language alignment of multimodal embeddings towards explainable predictions

     Gabrielė Tijūnaitytė, Thijs L. van der Plas, Mohamed-Bachir Belaid et al. · arXiv:2610.23110

     AI summaryAligns any multimodal embedding to language using captioning of auxiliary data, a soft contrastive loss, and concept retrieval. Applied to Earth observation embeddings, it outperforms prior models on species‑distribution metrics and yields interpretable habitat associations.

115. LED: Improving Masked Diffusion Language Models by Lookahead Editing

     Yuyan Zhou, Zeyuan Yang, Yuxuan Song et al. · arXiv:2610.23122

     AI summaryLED adds a rewritable proposal state to masked diffusion, allowing iterative lookahead updates before committing tokens, improving expressivity while keeping simple masks. Fine‑tuned Qwen models achieve +11.4/+7.7 points over standard MDM SFT and beat LLaDA2.1‑16B on benchmarks.

116. Information-Geometric Blind Spots of Conformal Test Martingales

     Johan Hallberg Szabadváry · arXiv:2610.23131

     Conformal test martingales (CTMs) are e-processes for sequentially testing exchangeability via p-values produced by a conformity function $A$. However, any distribution shift that preserves the push-forward $P\circ A^{-1}$ leaves the p-value distribution invariant, rendering it…

117. MetaCurate: Evolving Curation Filters for High-Quality Agentic Coding Data

     Donghoon Kim, Seunghyuk Oh, Kevin Galim et al. · arXiv:2610.23134

     AI summaryMetaCurate autonomously evolves executable curation filters for agentic coding data using pairwise LLM judgments, eliminating hand‑written rules. It achieves the best downstream coding performance on the Nemotron v3 SWE dataset, outperforming random, hand‑authored, and training‑loop baselines with lower search cost.

118. Local proxies for memory-efficient training

     Wu S. Zihan, Brian A. Yuan, Guillaume Bellec · arXiv:2610.23145

     AI summaryLocal proxy networks per transformer block predict global outputs, allowing gradients to be computed locally and reducing memory. On CIFAR‑100 and LLM fine‑tuning, this method achieves higher accuracy than other memory‑efficient local algorithms while using comparable memory.

119. Grounded Credit Assignment in Reinforcement Learning for Multi-Turn Search Agents

     Qirui Liu, Hao Chen, Yi Wen et al. · arXiv:2610.23148

     AI summaryGrounded Credit Assignment assigns separate rewards for evidence retrieval and answer correctness, correcting bias from GRPO. It boosts EM by 5.7–6.3 points and raises evidence-supported correct answers from 76.9% to 93.4% on QA benchmarks.

120. CIIR - Class-Indexed Interface Representation Learning

     Doo Yon Kim, Christian Wallraven · arXiv:2610.23151

     AI summaryCIIR learns a shared 128‑dimensional address space linking heterogeneous datasets without paired samples. It attains 94.7% address realization accuracy and enables cross‑dataset transfer (e.g., 90.2% MNIST→SVHN) and arithmetic coding for 100 CIFAR‑100 classes.

121. Nearly Linear Private Learning of Littlestone Classes

     Shay Moran, Chao Yan · arXiv:2610.23160

     AI summaryThis paper establishes a nearly linear upper bound of $O(d \log d)$ on the differentially private sample complexity for learning binary concept classes with Littlestone dimension $d$. This improves significantly over prior polynomial bounds and yields a tight linear dependence for learning subspace indicators.

122. Explaining Chess and Go in Natural Language

     Ori Press, Alexander Golubev · arXiv:2610.23166

     AI summaryThe paper introduces BinProp, a training‑free method that lets language models learn reusable chess concepts to generate natural‑language move explanations. BinProp nearly triples explainability scores over direct prompting and generalizes to Go, draughts, and shogi.

123. Structural Affinity: An Algorithm for Coherent Cage-Based Shape Design

     Aviraj Newatia, Shengan Yang, Hantao Zhong et al. · arXiv:2610.23173

     AI summaryStructural Affinity precomputes an affinity matrix linking a fine inner cage to a coarse outer cage, directly coupling handle motions and reducing cage self‑intersections and mesh distortion. Experiments on co‑design, neural transfer, and aerodynamic wings show smoother deformations and lower distortion metrics.

124. Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning

     Heming Zou, Qi Wang, Yixiu Mao et al. · arXiv:2610.23181

     AI summaryTreeOPSD is a critic‑free, tree‑aware posterior self‑distillation method that back‑ups terminal rewards through alternative continuations, producing a bounded teacher distribution. It reduces policy drift and sparsity, boosting long‑horizon agentic RL performance (e.g., +4.97 on Deep Research).

125. A Stronger Teacher Within: Strengthening MoE Models via On-Policy Bidirectional Distillation

     Haoyu Chen, Jun Bai, Yang Liu et al. · arXiv:2610.23185

     AI summaryMoEvolve expands MoE activation budgets to create a stronger internal teacher, then distills its gains back to a low‑budget model. The method boosts performance by up to 22.2 domain‑average points, exceeding same‑budget RLVR baselines by 6.5 points.

126. LangNCO: Compiling Natural-Language Constraints into Inference-Time Interventions for Neural Combinatorial Optimization

     Hironori Ohigashi · arXiv:2610.23190

     AI summaryLangNCO compiles natural‑language constraints into structured inference‑time masks for a frozen neural combinatorial optimizer. It achieves 97–100 % feasibility on CVRP with various constraints and 1.6–3.3 % mean gap, enabling post‑training adaptation without retraining.

127. Proportionality in Ranking Compression

     Daniel Halpern, Xinyu Liu, Evi Micha et al. · arXiv:2610.23193

     Motivated by applications in reinforcement learning from human feedback, we consider a setting where a given collection of rankings needs to be compressed into a smaller collection of rankings. This can also be seen as a combination of two fundamental social choice frameworks:…

128. Non-Parametric Optimization for Scalable Learning in Stochastic Decision Problems

     Mohsen Amidzadeh, Lauri Viitasaari, Mario Di Francesco · arXiv:2610.23204

     AI summaryDevelops a Malliavin‑calculus–based stochastic variational framework for time‑varying stochastic optimization, yielding non‑parametric optimality conditions and a dimension‑robust deep‑learning algorithm (Stochastic Path Follower) that outperforms current methods on drifting least‑squares and logistic regression problems.

129. Process Reward Modeling with Reliability- and Conservatism-Aware Bayesian Belief

     Zhongjian Qiao, Zhixiang Zhou, Jiafei Lyu et al. · arXiv:2610.23223

     AI summaryBayesianPRM models process rewards as random variables, updating a belief over a reward ensemble to handle PRM uncertainty (via reliability‑aware updates) and policy uncertainty (via conservatism‑aware calibration). It improves reward accuracy and test‑time scaling on multimodal reasoning tasks.

130. Characterising and Analysing Retrieval-Augmented Generation Techniques

     Emmanuel Igbekele, Olawande Daramola · arXiv:2610.23226

     AI summaryThe paper proposes a conceptual framework with dimensions and sub‑dimensions for categorising and evaluating Retrieval‑Augmented Generation methods, enabling researchers to filter suitable RAG techniques for specific tasks before costly empirical benchmarking.

131. Response-Informed Value Estimation for Cost-Aware Multi-Turn LLM Routing

     Haotian Yan, Jiaqi Liu, Bingyi Jing et al. · arXiv:2610.23239

     AI summaryRIVET uses logged responses to train two value estimators: a deployable pre‑call routing value and a training‑only response‑informed value. It reduces estimation error and achieves 74.7% lower cost while solving more MINT benchmark tasks than previous baselines.

132. Anisotropic Representation Dynamics in Transformers

     Nguyen Do, My T. Thai · arXiv:2610.23242

     AI summaryThe paper introduces a Mixture‑Gated Theory of Anisotropic Representation Dynamics, decomposing attention into value transport, query gating, and key gating. It provides a spectral criterion that predicts when attention routing causes representation differences to contract or expand, and validates the theory on several Transformers.

133. Think More or Read More? A Token-Budget Theory of Reasoning with Shared Context

     Manh Nguyen, Sunil Gupta, Hung Le · arXiv:2610.23246

     AI summaryThe paper formalizes a token‑budget trade‑off for reasoning systems that share a fixed context window: each token spent reading others reduces the reasoning budget. It provides a closed‑form prediction for the optimal group size in multi‑agent debate, which matches empirical results and shows self‑refinement performs similarly, especially under tight token limits.

134. Can the Parts Fool the Test? Counterfactual Pairing Cycles for Relational OOD

     Yibo Zhou, Yirong Xiang · arXiv:2610.23248

     AI summaryThis paper introduces counterfactual pairing cycles, a blocked evaluation and training primitive that holds image and caption parts constant while swapping their pairings to eliminate superficial shortcuts. This approach successfully isolates relational out-of-distribution errors and improves vision-language model accuracy on complex pairings.

135. FLARE: Physics-Grounded Evaluation of Language-Model Agents for Feeder Restoration with Degraded Observations

     BHVSP Subrahmanyam, Neeraj Choudhary · arXiv:2610.23251

     AI summaryThe paper introduces FLARE, a physics-grounded benchmark for evaluating language-model agents on power grid restoration under degraded observations. Evaluating five models across 800 cases, it finds that while median performance is strong, all models exhibit significant failure tails where external safety checks are essential.

136. Sample, Mix, or Verify? Allocating Test-Time Compute under Imperfect Verifiers

     Srinivasan Iyengar · arXiv:2610.23255

     AI summaryThe authors introduce SPLIT, a method that uses a small pilot to optimally allocate test-time compute across sampling, mixing models, and verification. SPLIT achieves near-optimal performance across diverse benchmarks and imperfect verifiers, outperforming heuristic budgeting rules.

137. Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To

     Uriel Dolev, Idan Schwartz, Yoav Goldberg · arXiv:2610.23257

     AI summaryReasoning models struggle to follow chain-of-thought constraints when given in user prompts, but achieve significantly higher controllability when the constraints are placed as self-instructions inside their reasoning process.

138. ADT: All-Atom Diffusion Tokenizer for 3D Molecular Structures

     Xuantuo Huang, Encheng Su, Aoran Wang et al. · arXiv:2610.23265

     AI summaryThis paper introduces the All-Atom Diffusion Tokenizer (ADT), which uses compact discrete structural tokens combined with a diffusion-based decoder to represent and reconstruct both small molecules and all-atom protein structures. Experiments show that ADT achieves high-fidelity reconstruction and enables quality autoregressive generation.

139. Beyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents

     BHVSP Subrahmanyam, Neeraj Choudhary · arXiv:2610.23266

     AI summaryThis paper introduces JointCoreBench to study interaction-aware memory selection for LLM agents, and proposes JointCore, a baseline that recovers minimum-cost sufficient memory subsets while significantly reducing evaluations and active memory.

140. NestedVLA: Learning to Consolidate and Generate Skills for Vision-Language-Action Models

     Guodong Du, Wanyu Lin · arXiv:2610.23271

     AI summaryNestedVLA learns individual VLA skills separately, then uses a hypernetwork to generate task‑specific action‑expert parameters that compose with a frozen base VLA, avoiding interference and parameter conflicts. Across MetaWorld, ManiSkill, and CALVIN, it beats the best model‑merging baseline by 41 pp on average and matches per‑task fine‑tuning on ManiSkill.

141. DeDiCa: A Scalable Deconfounding Discrete Causal Generative Model

     Kunchangtai Liang, Adrián Javaloy, Antonio Vergari et al. · arXiv:2610.23272

     AI summaryDeDiCa introduces a tensorized causal Bayesian network that uses maximum likelihood to handle hidden confounding in discrete variables. It matches state-of-the-art neural models in treatment effect estimation while enabling exact, millisecond-scale inference and efficient CPU training.

142. Data-Efficient Classifier Pretraining Through Diffusion Models

     Qinyu Li, Yee Whye Teh, Chris Holmes et al. · arXiv:2610.23289

     AI summaryThe paper introduces DiffPT, a diffusion-based pretraining method for classifier backbones in low-data regimes. The main result shows that DiffPT outperforms self-supervised baselines on medical imaging and time-series tasks with limited training data.

143. Distributed Performance Boosting through Neural Control for Plug-and-Play Networks

     Danilo Saccani, Leonardo Massai, Giancarlo Ferrari-Trecate · arXiv:2610.23292

     AI summaryDistributed neural controllers are designed by parametrizing local neural operators so that an auxiliary operator satisfies an $\ell_2$‑gain bound, guaranteeing closed‑loop stability during plug‑and‑play subsystem additions/removals. The resulting unconstrained training improves voltage transients in DC‑microgrid experiments.

144. The Optimal Ranking Tradeoffs Between Two Ranking Scores in Two-Class Classification

     Sébastien Piérard, Adrien Deliège, Marc Van Droogenbroeck · arXiv:2610.23294

     AI summaryThe paper shows that many common classification metrics are special cases of a single parameterized ranking score. It derives an exact formula for the optimal tradeoff between any pair of such scores, minimizing Kendall‐distance variance over all possible rankings.

145. Learning angles for quantum approximate optimization

     Johannes Jakubik, Isabelle Wittmann, Zaheed Gaffoor et al. · arXiv:2610.23296

     AI summaryWe train neural models to predict high‑quality QAOA angles from the problem graph, replacing costly per‑instance optimization. The predictions run in milliseconds, match solver quality, generalize to new topologies, and perform equivalently on quantum hardware.

146. How the elastic axis folds the transonic flutter boundary: shock energy transfer and conformally calibrated active learning

     Ashray Saxena, Sharad Shrivastava · arXiv:2610.23298

     AI summaryMaps the folded transonic flutter boundary of a pitch‑plunge aerofoil with the elastic axis ahead of the leading edge using 193 Euler simulations, a GP surrogate trained by active learning, and conformal calibration. The surrogate predicts stability with <1.6% error, revealing the fold exists only when the axis is >0.4 chord ahead, and explaining its closure via shock‑energy budgets.

147. Look, Reason, Draw a Region: Recursive Refinement for Spatial Context Prediction

     Wajeeha Tahir, Anna Tegon, Lionnus Kesting et al. · arXiv:2610.23302

     AI summaryThe paper introduces Spatial-TRM, a compact, weight-tied recursive model that uses frozen vision-language features for spatial context prediction. It achieves a 29.51% context accuracy on the RoboSpatial-Home benchmark, significantly outperforming single-step models with the same parameter count.

148. Untrained CNN and coefficient-space regularization for joint reconstruction in photoacoustic tomography

     Alfio Bozi, Anwesa Dey, Souvik Roy · arXiv:2610.23315

     AI summaryPaper uses fixed wavelet bases and untrained CNN features with Tikhonov regularization to reconstruct pressure, then learns a coefficient‑space prior for sound speed, achieving accurate, noise‑robust results with ~6× less computation than supervised CNN.

149. The Loss or the Architecture? What Matters in Survival Analysis with Competing Risks

     Julie Alberge, Gaël Varoquaux, Judith Abécassis · arXiv:2610.23322

     AI summaryThe study independently tests loss functions and backbones for deep competing‑risk survival models across tabular and image data, finding that a cumulative‑incidence loss dominates over architecture choice; the best performance comes from combining a proper loss with a modality‑matched backbone, yielding state‑of‑the‑art Brier scores.

150. Peak-Memory Schedules You Can Trust: Exact Realization, Certified Optimality, and the Limits of Recomputation for Inference

     Ivan Fomin, Ivan Koryabkin, Boris Sheludko · arXiv:2610.23328

     AI summaryThe paper presents an exact, certifiable memory‑scheduling method for inference that models a compiled program’s memory use, guarantees bit‑exact results, and proves optimality for 94 % of graphs, while also proving selective recomputation NP‑hard and showing modest gains on certain models.

151. HiPredict: Predicting the missing HiRISE RED4 spectral channel with machine learning

     Ardan Suphi, Ben Moseley, Gareth S Collins et al. · arXiv:2610.23335

     AI summaryHiPredict uses a U‑Net to reconstruct missing HiRISE RED4 channel imagery from IR10, BG12, neighboring red images and metadata, achieving a 1.1% NMAE versus 3.1% for the synthetic SYN4 method, improving colour fidelity and feature detection.

152. Minimax-optimal Scaling Laws in Ridge Regression via Transfer Learning

     Yalda Shabanzadeh, Diyuan Wu, Marco Mondelli · arXiv:2610.23337

     AI summaryThe paper analyzes ridge regression transfer learning: a pretrained source model regularized by ridge and a target model regularized by its distance to the source. It derives deterministic risk equivalents, shows minimax‑optimal scaling under power‑law covariance spectra, and confirms the result with experiments.

153. SETAREH Transform: A Sorting-Assisted Haar-Wavelet Sparsification Method for Approximate Inner-Product Computation

     Seyed Kamaledin Setarehdan · arXiv:2610.23338

     AI summaryThe paper introduces SETAREH, a Haar‑wavelet sparsification that preserves exact inner products before truncation by applying identical permutations and transforms to fixed coefficients and inputs. It shows that far fewer transformed pairs are needed to approximate inner products, reducing storage and communication costs.

154. How Finite-Rank Bottlenecks Shape the Low-Rank Adaptation Landscape

     Long Nguyen-Chi, Quynh Nguyen, Thanh Nguyen-Cung et al. · arXiv:2610.23343

     AI summaryLoRA updates are low rank because, once the backbone is frozen, the loss depends on each adapted matrix only through a finite‑rank linear map induced by frozen activations. Trace‑norm regularization forces the optimizer to choose a matrix whose rank is bounded by this bottleneck rank, not the dataset size.

155. RTSM: Adapting Recursive Models to Multivariate Time-Series Classification

     Georg Tenzing, Anna Tegon, Luca Benini et al. · arXiv:2610.23353

     AI summaryRTSM uses a learned convolutional tokenizer, a single weight‑tied refinement state, and pooled readout for multivariate time‑series classification. It attains 72.81% mean accuracy on UEA datasets, 1.60 points above MambaSL, showing temporal encoding and state design are critical.

156. Beyond Success Rate: Interventional Evaluation of Memory in Vision-Language-Action Systems

     Keyuan Luo, Yi Xie, Wanyun Ling et al. · arXiv:2610.23372

     AI summaryThis paper introduces an Interventional Memory Evaluation Protocol (IMEP) for vision‑language‑action systems and shows that providing correct historical facts raises task success from 39.1% to 74.3% on ChainBench, while mere memory changes can alter outcomes by up to 11.9% in cross‑episode tests.

157. Data-Driven Prediction of Optimal MPI Rank Placement and Node Count for Memory-Bound CFD on AMD EPYC Clusters

     Marcin Lawenda, {\L}ukasz Szustak, Aleksandra Krasicka · arXiv:2610.23383

     AI summaryThis study measures 560 OpenFOAM runs on AMD EPYC to map memory‑bound performance vs. cost. It finds that the nearest‑neighbor mesh predicts optimal MPI placement 75–85% of the time, and two recommended configs catch 85–90% of speed and cost trade‑offs, outperforming ML models.

158. Sentence-Level Topic Modeling with Lexical and Contextual Evidence

     Federica Granese, Serena Villata, Charles Bouveyron · arXiv:2610.23388

     Topic modeling faces a fundamental granularity dilemma: probabilistic topic models assign topics at the word level, whereas transformer-based methods operate at the document level. The latter overlooks that real-world documents typically cover multiple topics, whereas the former…

159. Forecasting Voltage Time Series under Anomalies Using a WGAN Model With Lipschitz Control and Progressive Kernel Critic Strategy

     Najmieh Sadat Safarabadi · arXiv:2610.23394

     In time-series forecasting tasks, the use of recursive forecasting on high-dimensional data substantially increases the complexity of the prediction process. The main challenge is that long-term dynamics becomes harder to model, since the sequences show greater temporal…

160. Contracting with an Agent under Model Mismatch

     Elif Y{\i}lmaz, Jakub T{\l}uczek, Victor Villin et al. · arXiv:2610.23402

     In the classical principal-agent setting, the principal delegates some task to the agent, who receives a payment conditioned on the outcome of the task, which in turn benefits the principal. To align their objectives, the principal designs a contract that incentivizes the agent…

161. Information Estimation with Discrete Diffusion

     Alberto Foresti, Giulio Franzese, Pietro Michiardi · arXiv:2610.23404

     Information-theoretic measures, such as Mutual Information (MI), play a crucial role in understanding non-linear relationships between random variables and are widely used across scientific disciplines. Yet, their use on real-world discrete data remains challenging. Existing…

162. Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation

     Mahmoud Ashraf · arXiv:2610.23409

     AI summaryFlash RNN‑T reduces training memory by combining custom kernels with optional tiled recomputation. On H100 it delivers 24–39 % higher throughput and up to 56 % less memory than baselines, achieving ninefold memory savings on medium workloads while staying within 16 % of peak speed.

163. Residual Stochastic Concept Bottleneck Models

     Stephen Marcou, Mateo Espinosa Zarlenga, Sonia Laguna et al. · arXiv:2610.23421

     Concept-based models (CMs) learn human-interpretable intermediate representations that enable users to understand the model's reasoning and steer its predictions through test-time concept interventions. However, in realistic settings, their performance can be limited by concept…

164. Bridging Compressed Communication and Local Steps via Spectral Method in Federated Learning

     Dmitriy Bystrov, George Chirkov, Daniil Medyakov et al. · arXiv:2610.23434

     AI summaryThis paper introduces a spectral federated learning method that integrates local steps and communication compression. The approach provides rigorous convergence guarantees without requiring bounded data heterogeneity and demonstrates strong empirical performance on image classification benchmarks.

165. Generative Policy Improvement from Observational Data

     Jonas Schweisthal, Dennis Frauen, Stefan Feuerriegel et al. · arXiv:2610.23436

     We study how to improve decision policies from observational data to generate better actions when actions are complex, structured objects, such as text messages or images. This task is subject to two major challenges: (i) action-level overlap is often too weak for reliable…

166. RECREG: Recovery-based Shielding for Regular Safety Properties in Reinforcement Learning

     Alexander W. Goodall, Francesco Belardinelli · arXiv:2610.23445

     AI summaryRecReg uses recovery-based shielding to enforce non-Markovian regular safety properties in RL, offering a unified framework with theoretical safety and convergence guarantees in tabular settings and empirical validation via MASA-Safe-RL.

167. Regularized free deconvolution for spectral repair of resampling distortion in high-dimensional particle filters

     Deming Song, Chi-Kin Lam, Kefeng Li et al. · arXiv:2610.23448

     AI summaryReconstructs the true ensemble covariance spectrum after high‑dimensional resampling in particle filters using a regularized free‑deconvolution estimator; achieves order‑of‑magnitude error reduction and recovers 90% of oracle sampling efficiency with proven consistency and explicit error bounds.

168. Predictable Scaling: Efficient Model Exploration via Scaling-Law Prediction

     Aishwarya Wesanekar, Sheng-Tao Yang, Jiawei Zhang et al. · arXiv:2610.23453

     AI summaryPredictable Scaling predicts full‑scale model performance from small‑scale runs, reducing compute by 79–88%. A Gaussian Process variant yields 15% lower MAE and calibrated uncertainty, enabling accurate ranking of hypotheses across two ranking models.

169. HydraLLM: Adaptive Width Slicing for Elastic Text and Vision-Language Models

     Oscar Kemme, Janek Haberer, Ali Hojjat et al. · arXiv:2610.23455

     AI summaryHydraLLM introduces an elastic MQA architecture that adjusts the number of active attention heads to scale model width for any hardware. It outperforms nested baselines by 2.2 % TriviaQA accuracy at 75 % head usage, matches Gemma‑3 4B CLIPScore with 50 % heads, and boosts throughput by 75 % at 25 % heads.

170. Graphlet-based Decentralized Finance Transaction Method Classification

     Célestin Coquidé, Alessio Brini, Natkamon Tovanich · arXiv:2610.23466

     AI summaryGraphlet features of directed DeFi transaction graphs classify 30 common methods better than event logs alone, boosting accuracy. SHAP analysis reveals distinct structural patterns for each method, providing interpretable explanations.

171. QD-Rec: Open-Ended Evolution of Recommendations via Quality-Diversity

     Takashi Ishida, Taro Makino, Issa Sugiura · arXiv:2610.23470

     Open-ended search combined with foundation models has produced advances such as faster algorithms, new results on open mathematical problems, and self-improving agents. We extend this paradigm to recommendation for decision-making, focusing on decisions where the set of possible…

172. Towards optimal selection schedules for online active learning

     Hugo Cui, Francesca Mignacco, Francesco Mori · arXiv:2610.23477

     We consider the problem of online linear classification, where a stream of unlabeled samples is available, and annotation is costly. We investigate the optimal query policy, that prescribes whether and when a sample should be labeled and used for a training step, or instead…

173. A Topological Signature of Learning under Noisy Labels

     Zhongtian Sun, Fan Mo, Prayag Tiwari et al. · arXiv:2610.23491

     How does noisy supervision change the representations learned by a neural network? Training accuracy measures agreement with training labels but provides little information about representation geometry. We use persistent homology to track loop structure across distance…

174. Adaptive conformal prediction on the Grassmannian

     Dimitrios Loukrezis, Zachary J. Grey, Dimitris G. Giovanis · arXiv:2610.23508

     We introduce a conformal prediction framework for regression with Grassmann-valued responses. The proposed method constructs prediction regions in the form of metric balls directly on the Grassmann manifold, using basis-invariant geometric distances as non-conformity measures…

175. Uncertainty-aware AI-enabled framework for predicting hydrogen dispersion in confined spaces to enhance operator situational awareness

     Niclas Flehmig, Mary Ann Lundteigen, Shen Yin · arXiv:2610.23515

     We present a two-stage AI-enabled framework for enhancing operator situational awareness during compressed hydrogen bunkering in maritime confined spaces. In the event of a leak, the framework first estimates the unknown leakage rate from early sensor readings and then forecasts…

176. Interpreting SSL Representations for Spoof Detection: a WavLM Study

     Mohamed Mallat, Michele Panariello, Massimiliano Todisco et al. · arXiv:2610.23526

     AI summaryWavLM’s layers encode spoof‑relevant cues at different depths: mid‑layers best capture many acoustic features, while late layers encode voice quality metrics. Using five optimally chosen layers gives better spoof detection than pooling all thirteen.

177. Out of Support, Not Out of Reach: In-Context Learning with Unsupported Instructions and Composed Tasks

     Xuhan Tong, Yuchen Zeng, Jiawei Zhang · arXiv:2610.23528

     In-Context Learning (ICL) enables pretrained LLMs to adapt to downstream tasks from a small set of input-output demonstrations, without parameter updates. Yet existing theory primarily explains ICL as identifying a latent task already represented in pretraining, leaving unclear…

178. Causal Discovery beyond Conditional Independence: Identifiability, Algorithms, and Statistical Inference for Directed Acyclic and Cyclic Graphs

     Yang Ni · arXiv:2610.23543

     AI summaryThe paper reviews causal discovery methods that use distributional features beyond conditional independence, separating identification, estimation, and computation, and discusses latent confounding, feedback, and cyclic graphs, offering consistency, uncertainty, and algorithmic insights for diverse data settings.

179. Beyond Latent Personalization: Integrating HRV-Conditioned Personalization via LLM for Interpretable Stress Detection

     Wei-Hsiang Ma, Ya-Tse Wu, Hsing-Hang Chou et al. · arXiv:2610.23555

     Stress detection based on heart rate variability (HRV) has received increasing attention. However, existing personalized approaches integrate personal attributes with HRV through latent representations, limiting the interpretability of their interactions. In this work, we…

180. Beyond Expected Values: Risk-Sensitive Planning with Distributional Monte-Carlo Tree Search

     Quynh Trinh, Trung Nguyen, Nam Nguyen et al. · arXiv:2610.23565

     Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR). We introduce Categorical Thompson Sampling with Optimism (CATSO) and Particle Thompson Sampling with…

181. A general recipe for data-adaptive anytime-valid envelopes for iterative algorithms

     Soham Bonnerjee, Aaditya Ramdas · arXiv:2610.23570

     Most convergence guarantees for iterative stochastic algorithms are stated at a fixed iteration count and do not justify stopping at a data-dependent time. Recent work has constructed anytime-valid envelopes that do, but deterministic envelopes are typically too conservative for…

182. BMuon: Budgeting Muon's Spectral Geometry for Structured Compressibility

     Rida Assalouh, Tom Jacobs, Rebekka Burkholz et al. · arXiv:2610.23571

     Muon has emerged as a strong optimizer for neural network training and admits an interpretation as steepest descent under a spectral-norm geometry. Yet we find that Muon-trained Transformers can degrade sharply when their MLP hidden channels are pruned after training. To…

183. Boundary Detection for High-Dimensional Point Clouds

     Abigail Hickok, Valerio Galanti, Osama Al Dalahmah et al. · arXiv:2610.23590

     We introduce two new methods for boundary detection in high-dimensional point cloud data, intended to work in situations with non-uniform density, high ambient dimension, and substantial noise. Our first method fits a separating hyperplane on local neighborhoods after doing…

184. When Magnitude Suffices: Reconstructing Time-Domain-Gated Microwave Signals Without Phase on a Thoracic Phantom

     Maciej Cichoń, Patrycja Go\l{}aś, Maksymilian Bartosik et al. · arXiv:2610.23608

     Time-domain gating separates the arrivals that form a microwave S-parameter measurement. It requires the measurement of complex values, so this application almost always uses a vector network analyzer (VNA). We show that, on one calibrated measurement configuration, the gated…

185. How Many, but How Sure? Class Competition as a Confidence Proxy in Multi-Class Density-Based Counting

     Villanelle O'Reilly, Georgios Leontidis, Marc Hanheide · arXiv:2610.23611

     Multi-class density estimators can count objects in crowded scenes, but their class-wise outputs offer no indication of how reliably density mass has been allocated. We introduce a training-free confidence proxy based on local competition between the strongest class channels.…

186. Trajectory-Triggered Sparse Collaborative Decoding

     Amit Ranjan Trivedi, Sruthikesh Surineni, Kyong Ho Lee et al. · arXiv:2610.23613

     Dense assistance methods such as speculative decoding repeatedly verify draft tokens with a target model, keeping target computation coupled to generation. We study sparse collaborative decoding, where a draft model generates most tokens and a target model is invoked only for…

187. The Sign Code: The Hidden Binary Nature of Deep Networks

     Niclas Göring, Shuofeng Zhang, Roi Holtzman et al. · arXiv:2610.23616

     Every neuron in a trained network outputs a number that can be separated into its sign (positive or negative) and its magnitude (absolute size). We call the set of all neuron signs the \textit{sign code}. This raises a key question: does classification depend more on the sign…

188. ThetreStageAI: Multi-View Retrieval-Augmented Generation for Structured Bengali Theatre Production Planning

     Md. Mehedi Hasan Naeem, Mst. Kamrunnahar Ruma, Shakila Sultana et al. · arXiv:2610.23618

     AI summaryThetreStageAI builds a multi‑view retrieval‑augmented system for Bengali theatre planning, using separate scene, blocking, and lighting indexes, then Gemini to generate validated JSON plans. It achieves better plan quality than single‑view retrieval, with a 1,500‑document synthetic dataset.

189. GRAFT-EEG: Gauge-Aware Physical Field Transport for Cross-Dataset Motor-Imagery EEG

     Ying Zhou, Mengting Liu · arXiv:2610.23620

     Cross-dataset EEG decoding often treats montage, reference, and active-sensor changes as representation shift, although they also alter the observation operator. We study target-statistics-blind transfer: fitting receives no target signal, label, or empirical statistic, and…

190. Deconstructing Multi-Task Active Learning: The Paradox of Gradient Conflict and Orthogonal Decomposition

     Juho Jung, Seongjae Kang, Jinyoung Han · arXiv:2610.23623

     Existing multi-task active learning (MTL-AL) methods predominantly rely on linear scalarization of heterogeneous objectives such as uncertainty, diversity, and gradient conflict. We argue that this entangled formulation is fundamentally limited, as linear combination fails to…

191. Deep LLM layers expose passage-level context in EEG during silent reading

     Ingo Marquardt, Anthilia Alchanat, Priyanka Jain · arXiv:2610.23638

     In a previous report we showed that a decoder can recover open-vocabulary word-level information from 19-channel EEG recorded during silent reading, by aligning short EEG windows with hidden-state embeddings of the presented word taken from a large language model (LLM). That…

192. No-Regret Language Generation

     Jiaxun Li, Vinod Raman, Ambuj Tewari · arXiv:2610.23641

     We introduce no-regret generation, which requires a generator to make a sublinear number of mistakes uniformly over all demonstration sequences and compatible target languages. We also formulate an agnostic variant that allows arbitrary demonstrations that may not be compatible…

193. Scalable long-range machine learning interatomic potentials with electrical response for atomistic material modeling

     Yicheng Chen, Xiaoyun Yu, Hyejeong Song et al. · arXiv:2610.23643

     AI summaryThe paper introduces DPA4C‑LR, a distributed machine‑learning interatomic potential that analytically separates long‑range electrostatics and Born effective charges, enabling scalable atomistic simulations up to millions of atoms. It accurately predicts field‑dependent infrared spectra, polarization hysteresis, ionic conductivity, and interfacial thermal transport.

194. JAXtari: High-Throughput and Easy-to-Modify Arcade Learning Environment

     Quentin Delfosse, Raban Emunds, Paul Seitz et al. · arXiv:2610.23650

     The Arcade Learning Environment (ALE) has been a cornerstone of Deep Reinforcement Learning research, providing a standardized benchmark for evaluating general agents. However, its lack of native support for object-extraction and environment modification restricts research into…

195. A Language-Informed Loss for Visual Generation

     Suchir Agarwal, Keshigeyan Chandrasegaran, Ziyu Chen et al. · arXiv:2610.23653

     AI summaryThe paper introduces LILIE, an invertible image encoder trained with a vision‑language model to generate captions from noised features, enabling a language‑informed loss for text‑to‑image generation. Using this loss, JiT models achieve a 14‑point gFD reduction on GPIC.

196. An Efficient Approach to Beetle Damage Detection of Banana Using Deep Learning

     Md. Mehedi Hasan Naeem, Ishtiyak Ahmmad Araf, Md. Arefin Haque Mahir et al. · arXiv:2610.23661

     Visible surface damage of banana fruits by Basilepta subcostata (the banana fruit scarring beetle) can reduce their market value. Here, manual inspection of Banana fruits is subjective and difficult to scale. This study presents a lightweight framework based on YOLOv8n deep…

197. TriRouter: Sparse-to-Dense Electrode-Space Routing for Robust EEG Motor Imagery Decoding

     Ying Zhou, Yaoxin, Mengting Liu · arXiv:2610.23665

     Motor imagery brain-computer interfaces require reliable decoding from multichannel EEG, but practical acquisition often yields sparse or corrupted electrode montages. We propose TriRouter, an electrode-space routing framework with discriminative-preservation, mask-robustness,…

198. HardEval: An Execution-Verified Benchmark for AI-Driven Security Hardening and Compliance Remediation

     Manohar Vellala, Soumya Banerjee, Sandip Roy et al. · arXiv:2610.23685

     AI summaryHardEval evaluates language models by having them generate Bash scripts that harden an RHEL 8 system to pass STIG controls, with live execution and OVAL checks grading success. The study shows significant variation in model performance across security domains and highlights realistic failures.

199. Expressivity of IQP Feature-Map Quantum Kernels via Reproducing Kernel Hilbert Spaces

     Valeria Repetto, Stefano Melchionna, Romina D'Aurizio · arXiv:2610.23692

     AI summaryThe paper derives a Fourier representation of the RKHS for IQP quantum kernels, revealing three bandwidth regimes—localized, multimodal, and periodic—whose expressivity depends on accessible frequencies. Numerical tests show each regime matches data structure, guiding bandwidth choice.

200. Equivalence testing and run-count design for stochastic machine-learning experiments

     Dmytro Knopov, Nadiya Shvai · arXiv:2610.23700

     Claims that two training methods perform equally well often rest on a non-significant difference between two or three runs per method. We develop a protocol, with its power theory, for certifying practical equivalence in per-run metrics such as the AUROC (area under the receiver…

201. Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2

     Vinay Saji Mathew, Soundar Kumara, Gretta Kellogg et al. · arXiv:2610.23703

     Diffusion-based protein structure prediction models produce all-atom structures in about 200 denoising steps, and that cost has driven a number of few-step samplers. We present Blitz, a few-step distillation framework that uses a geometric noise schedule to distribute steps more…

202. Tool Recommendation for Multi-Step Agents: Finding Every Tool a Task Needs

     Sam Partee, River Stevens, Thierry Damiba et al. · arXiv:2610.23707

     AI summaryTool Recommendation supplies agents with all needed tools via three fixed calls, covering 93% of required tools in Codex sessions. It improves recall over BM25 by 57%, cuts input tokens by 17–40%, and reduces cost by up to 40% without affecting completion.

203. Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize

     Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino et al. · arXiv:2610.23716

     Deep neural networks demonstrate remarkable generalization despite being highly over-parameterized, challenging traditional statistical learning theories. While several theoretical paradigms such as margin-based bounds, PAC-Bayesian analyses, and neural tangent kernels provide…

204. ECG-SDI: Single-lead Electrocardiogram Enables Continuous Sleep Depth Index Profiling and Scalable Digital Biomarkers for Sleep Health

     Donglin Xie, Qingshuo Zhao, Songchi Zhou et al. · arXiv:2610.23727

     AI summaryThe researchers developed a deep learning framework using single-lead electrocardiography for continuous sleep-depth profiling and biomarker extraction. The method accurately tracked sleep stages, identified a high-mortality disturbed sleep phenotype, and successfully monitored therapeutic responses to continuous positive airway pressure.

205. Escaping Parameter Space: Tight Generalization Bounds via Representation Quality

     Niclas Göring, Shuofeng Zhang, Branton DeMoss et al. · arXiv:2610.23729

     We derive a tight generalization bound for deterministic neural We derive a tight generalization bound for deterministic neural networks based on their last hidden layer representations. By evaluating representation quality with a parameter-free $k$-nearest-neighbor ($k$-NN)…

206. Bridging the Language Gap in Scholarly Data II: Enhancing Gender Inference for Chinese Names

     Lisette Espín-Noboa, Mingrong She, Ana Maria Jaramillo et al. · arXiv:2610.23734

     Perceived gender inference from names is important for inclusive scientometrics research, yet existing methods perform poorly on non-Western scholars. We address this gap for Chinese names, which appear in two forms, Hanzi and Pinyin, its Romanized form (e.g., Zhang Wei, Li Na).…

207. LagNet: Lag-Structured Operators for Long-Term Time Series Forecasting

     Zhenan Yu, Guangxin Jiang, Jiuyun Jiang et al. · arXiv:2610.23735

     Long-term time series forecasting maps a finite history window to an extended future horizon. Many existing models focus primarily on the design of input representations and model structures, while paying less attention to the structural form of the history-to-future forecasting…

208. From Flow-Box Invariants to Global Conservation Laws: Machine Learning-Assisted Detection of Singularities and Near-Integrability

     Jimmie Adriazola, Ioannis G. Kevrekidis, Panayotis G. Kevrekidis et al. · arXiv:2610.23760

     AI summaryThis paper introduces an expanding-domain neural network framework to track the transition from local flow-box invariants to global conservation laws in dynamical systems. The method successfully detects the successive breakdown of local invariants as domains expand, revealing proximity to Liouville integrability and isolating singularities.

209. Universal Semantic Enhancement for Weakly Supervised Anomaly Detection

     Siyu Zhou, Yuqin Xie, Hao Yuan et al. · arXiv:2610.23765

     Weakly supervised anomaly detection on tabular data is an essential task in many label-scarce applications, since tabular data is widely encountered and it is commonly observed in practice that only a limited number of labeled anomalies are available alongside abundant unlabeled…

210. Logic-Guided Graph Denoising Diffusion Models

     Samuel Cognolato, Luciano Serafini, Alessandro Sperduti et al. · arXiv:2610.23766

     Graph diffusion models are competitive approaches for generating complex graph-structured data. However, when such distributions have strict structural constraints, diffusion models may fail to capture them exactly, resulting in the generation of invalid graphs. In this paper,…

211. Mesoscopic learning dynamics of extensive-width neural networks

     Shu Zhou, Jean Barbier, Sebastian Goldt · arXiv:2610.23779

     Understanding how stochastic gradient descent learns internal representations in wide neural networks trained on rich target functions is a central problem in machine learning. We study this question in two-layer neural networks learning target functions represented by networks…

212. Sampling the XY model with an attention-based neural network

     Piotr Bia{\l}as, Christian D. Morales, Tomasz Stebel et al. · arXiv:2610.23783

     AI summaryTransformer-based variational autoregressive network samples 2D XY model near BKT transition; it outperforms previous neural samplers, matches MCMC energy, magnetization, helicity modulus, and covers all configurations on lattices up to 96×96.

213. Parametric Resonance in Primal-Dual Learning for Slowly Non-Stationary Constrained MDPs

     Maxime Baelde · arXiv:2610.23788

     Primal-dual methods for constrained Markov decision processes are analyzed under a two-timescale assumption, with non-stationarity entering through two channels. In the additive channel, where the constraint level oscillates, the dual oscillation saturates at a floor set by the…

214. Task-aware Kernel Stein certificates for Rényi differential privacy

     Leoni Carla Wirth, Tara Trautwein, Gesine Reinert · arXiv:2610.23798

     Rènyi differential privacy, based on Rènyi divergence, has become a notion of choice for privacy guarantees. However, outside the Gaussian setting, Rènyi divergence can be difficult to compute. In particular, Rènyi differential privacy guarantees that exploit task-specific…

215. Explainable Machine Learning Analysis of Temperature Effects on Mechanical Strength of Fiber-Reinforced High-Performance Concrete

     Jagaran Chakma, Zhiguang Zhou, Badhan Chakma et al. · arXiv:2610.23811

     Predicting the compressive strength of fiber-reinforced high-performance concrete subjected to temperature exposure is challenging because to the intricate interdependencies of mixture composition, fiber properties, curing conditions, and temperature. A data-driven framework to…

216. Learning Personalized Memory Gating from Delayed One-Sided Feedback

     Kaixuan Ji, Zixiang Chen, Wenting Zhao et al. · arXiv:2610.23812

     Long-term memory enables language-model agents to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. Existing memory gating policies are typically shared across users and therefore fail to account for user-specific patterns of future…

217. CocoRNA: Collective RNA Design with Cooperative Multi-agent Reinforcement Learning

     Tianmeng Hu, Biao Luo, Ke Li · arXiv:2610.23820

     Designing RNA sequences that reliably fold into specific secondary structures is essential for understanding their biological functions but remains a challenging computational problem. We propose CocoRNA, a cooperative multi-agent reinforcement learning framework for RNA inverse…

218. Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere

     Joshua Williamson, Niall Jeffrey, Lorne Whiteway · arXiv:2610.23821

     Many of the most important datasets in the physical sciences live natively on the sphere, and these data often carry no preferred position or orientation. Using methods that do not reflect this symmetry leads to far higher simulation requirements, which can be catastrophic for…

219. ENFR: Evidence Networks with Flow-Matched References for Amortized Simulation-Based Model Inference

     Kai Lehman, Jameson Dong, Carolina Cuesta-Lazaro et al. · arXiv:2610.23825

     The evidence, $\mathcal{Z} = p(x | \mathcal{M})$, of a data-generating model $\mathcal{M}$ is a powerful tool in Bayesian frameworks: it enables model inference, model misspecification tests, and goodness-of-fit assessment as an alternative to standard discrepancy statistics,…

220. Quality-of-Transmission Estimation in Quantum Networks: Compute, Simulate, or Learn?

     Anis Bekri, Amar Abane, Van Sy Mai et al. · arXiv:2610.23827

     AI summaryThe paper presents a framework to estimate end‑to‑end entanglement fidelity in quantum networks by choosing among analytical calculation, simulation, or learned surrogate models. It finds analytical methods work for simple cases, simulations are needed with stochastic memory effects, and learned surrogates can match simulator outputs at lower runtime cost.

221. Waiting Buys Exactly Two Moments: Endogenous Identification in Prediction-Driven Allocation

     Rama AlHamidi, Rasul Khanbayov, Erchin Serpedin et al. · arXiv:2610.23830

     A planner with a fixed budget of preventive help must decide whom to help and when. Waiting reveals risk but removes intervention opportunities as higher-risk individuals leave first. We characterize this tradeoff under repeated noisy signals and risk-dependent attrition: in the…

222. Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation

     Shaoang Li, Yanhang Shi, Yufei Li et al. · arXiv:2610.23831

     AI summaryReDiST builds student‑conditioned reasoning traces from the learner’s own attempts, revising them with evidence and answer guidance, then selects the best trace via a frozen reader; the distilled student improves over answer‑only and baseline trace methods on multiple QA benchmarks.

223. Task-aligned spectral inductive biases in parameterized quantum circuits

     Felix Frohnert, Adrián Pérez-Salinas, Sebastian Lange et al. · arXiv:2610.23834

     AI summaryThe paper shows that a fixed, task-specific spectral query embedded in a data‑reuploading quantum circuit changes its hypothesis class by reshaping how signal frequencies combine, without new sample data. In small spectral and synthetic aperture radar tasks, this query‑alignment improves learning from limited data.

224. Behavioral Couplings: From Formation to Behavioral Emergence in Language Models

     Dongyoon Hahm, Woogyeol Jin, Dennis Wei et al. · arXiv:2610.23847

     Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding.…