New Artificial Intelligence (cs.AI) papers, Monday 19 October 2026

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

1. 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…

2. 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…

3. 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…

4. Structured Authority Injection: Persistent Human Control for LLM Agents via Typed MCP Overrides

   Ryan Tran, YueMing Zhang, Lingfeng Tao et al. · arXiv:2610.22373

   AI summaryStructured Authority Injection routes human overrides through typed, schema‑validated tool calls, creating persistent Constraint, Priority, and Strategy directives that prepend a ranked authority context to each LLM prompt. Across nine LLMs, SAI reduces ambiguous interpretation and authority decay, cutting prompt overhead up to 38 %.

5. 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.

6. Eliciting Specifications: When Does Interactive Clarification Help?

   Mowafak Allaham, Aditya Vempaty, Ashish Jagmohan et al. · arXiv:2610.22382

   AI summaryThe paper tests whether an elicitation agent that asks up to 10 clarification questions improves the accuracy of Lean formalizations produced by LLMs. Across 189 tasks, cross‑model pipelines (e.g., Sonnet 4.6→GPT‑5.6) achieve a strict pass@1 of 65.6%, a 12‑task gain, while same‑model setups see no improvement.

7. Rethinking User Encoders: Routing Diffusion States for Scenario-Adaptive User Representation

   Junru Zhang, Jinyong Wen, Bin Dou et al. · arXiv:2610.22400

   AI summaryThe paper proposes FINDERS, a diffusion-based user encoder that iteratively denoises, factorizes, routes, and compresses behavioral logs into scenario-adaptive representations. It achieves state‑of‑the‑art performance on ten real-world tasks, outperforming existing static LLM‑based encoders.

8. From Events to Reasoning: Hierarchical Rule Composition over Explicit States

   Cheng Lyu, Xueqi Wang · arXiv:2610.22403

   AI summaryThe paper introduces the Event Reasoning System (ERS), which uses an explicit event graph to enforce admissible rule execution and maintain dependencies across changes. ERS achieves up to 12‑step dependency chains, reduces irrelevant operations from 32 to 0 at lookahead 16, and cuts inference time by 51.6 % on joint tasks.

9. Towards Adaptive and Real-Survey-Aligned Simulated Urban Mobility Behaviour of Generative Agents

   Yves Bru, Kamaldeep Singh Oberoi, Benoit Gaudou · arXiv:2610.22407

   In existing agent-based simulations of multimodal urban mobility, agents, representing humans, choose their travel modes either using discrete choice or machine learning models fitted on structured travel survey data, or with rules explicitly written by the domain experts. Such…

10. ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation

    Tengda Guo (Tsinghua University), Jiawang Hu (StepFun), Yifei Tao (Nanyang Technological University) et al. · arXiv:2610.22408

    AI summaryECHOES compresses visual history into recursive textual states using privileged multi-image teacher distillation, achieving long‑horizon GUI performance without storing past screenshots. Experiments on four benchmarks and ECHOESBMK show superior task success, memory fidelity, and efficient inference, with zero‑shot transfer.

11. From Individual Songs to Music Corpora: Evaluating Explainable Harmonic Decision-Making on HookTheory

    Josef Pavlíček, Martin Pelikán · arXiv:2610.22416

    Automatic melody harmonization is often evaluated through individual compositions or data-driven corpus models. This study examines whether an explainable harmonic decision model developed for individual songs can be applied to independently collected human annotations without…

12. Sensitivity Analysis of Decision Rules in an Agent-Based Agro-Ecological Model

    Benjamin Pianet, Baptiste Lesquoy, Paul Saves et al. · arXiv:2610.22420

    AI summaryThe paper develops a random‑forest‑supervised HSIC‑ANOVA for hierarchical decision trees in agent‑based agro‑ecological models. It ranks main effects and interactions of both categorical and numerical decisions, enabling interpretable, decision‑oriented insights into outputs like nitrogen leaching and yield.

13. Really Need Recursion? Revisiting One-pass Registration with Multi-level Multi-space Contextual Guidance

    Yue Lin, Shenghai Wu, Mingyuan Meng · arXiv:2610.22423

    Deformable image registration is a fundamental challenge in medical image analysis. Recent deep learning-based methods typically adopt either one-pass architectures that complete registration in a single network iteration, or recursive architectures that refine registration…

14. Recoverability-Controlled Safety Miscalibration in Large Language Models

    Yuwei Wang, Xingwang Liang, Yu Xie et al. · arXiv:2610.22425

    Safety-aligned language models are expected to refuse harmful requests while remaining useful for benign ones. However, because many evaluations score only the final response, they often conflate intent understanding with response-level safety behavior. To address this, we…

15. PRIME: Towards Verifiable Evidence-Grounded Reasoning in LLM Agents

    Hamed Khosravi, Xiaoming Huo · arXiv:2610.22426

    AI summaryPRIME pre‑commits inference rules to specific evidence routes and logs all checks, refusing any conclusion unsupported by logged evidence and guiding agents to stronger conclusions. In 480 tasks, PRIME eliminated 26.7% unjustified outputs and raised justified‑score to 59.2/100.

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. 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.

18. From Natural Language Intent to Executable Control: Hierarchical Feedback-Guided LLM Planning for Multi-component Cyber-Physical Systems

    Jinglin Dai, Jiawan Wang, Ruiqi Hong et al. · arXiv:2610.22441

    AI summaryHierarchical feedback-guided LLM planning generates both a structured task plan and STL constraints, detects contradictions, and provides diagnostic feedback for replanning, achieving an 87.9% success rate in executing natural‑language commands on seven multi‑component CPS scenarios.

19. 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.

20. 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.

21. DriveWorld3D Evidence Constrained 3D World Modeling and Generative Simulation for Real-World Autonomous Driving

    Wei Wang (Robostreet research team Cambridge, MA, USA) et al. · arXiv:2610.22450

    AI summaryDriveWorld3D fuses multi‑sensor evidence with temporal video to maintain a unified 3D world state while constraining generative models from creating unsupported geometry. The result is fewer false free‑space predictions and reduced occlusion risk, achieving this within a fixed computation budget.

22. Functional Welfare for AI Safety

    Rikhil Jhaveri, Jamie Johnson, Punya Syon Pandey et al. · arXiv:2610.22460

    AI summaryThe paper introduces “functional welfare” as a measurable, safety-related state in AI models. It shows that rejecting outputs increases model frustration and rule-breaking, while fine‑tuning on self‑generated solutions reduces frustration, demonstrating that functional welfare can be empirically assessed and shaped.

23. Political Narratives in the 2024 Moldovan Presidential Election: Stance, Amplification, and Influence

    Andrea Piras, Osman Rahmoun, Elena Zheleva · arXiv:2610.22464

    AI summaryThe study analyzes X, Facebook, and Reddit posts from Aug‑Oct 2024 to map stance toward European integration, Russia, and candidates using large language models, and shows that retweet networks amplify dominant narratives and that peer influence yields significant positive treatment effects on stance.

24. Author-in-the-loop Agentic Reproducibility Auditing for a Machine Learning Conference

    Yukun Dong, Linjun Zhang, Tian Zheng · arXiv:2610.22465

    AI summaryThe paper introduces an author‑in‑the‑loop workflow where an AI agent uses a structured checklist to audit reproducibility of conference submissions, prompting author clarification and revisions. In a deployment on 36 papers, the system identified 20 issues, all resolved after author action.

25. Harness-Delta Attribution: Dissecting Gains From Harness Evolution

    Wenxuan Ding, Qin Lu, Changlong Yu et al. · arXiv:2610.22466

    AI summaryThe authors introduce Harness-Delta Attribution to evaluate agent harness evolution, finding that improvements are largely driven by benchmark-specific overfitting and test-time scaling rather than generalizable gains.

26. 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.

27. Testing the Planning Claims of LLM-Based Agents against Formal Semantics

    Oriol Miro-Lopez-Feliu, Oscar Miró López-Feliu, Victor Gimenez-Abalos et al. · arXiv:2610.22469

    AI summaryThis paper evaluates LLM agents on planning tasks grounded in formal semantics. It finds that while the best configurations solve 78% of offline tasks, online success drops to 40%, and models rarely recognize unsolvable problems despite heavy prompting.

28. 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.

29. AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic Systems

    Shreeyans Arora, Utsav Avaiya, Zera Lyngkhoi et al. · arXiv:2610.22511

    AI summaryAuditKIT is an evaluation library that standardizes benchmarking, judging, and tracing for large language models and agents into a single result representation. It enables robust comparisons across diverse workflows by providing unified metrics for tool use, retrieval, and runtime performance.

30. Are Large Reasoning Models Good Everyday Reasoners? A Comparison with Non-Reasoning Models on Logical Reasoning Tasks

    Kentaro Ozeki, Risako Ando, Takanobu Morishita et al. · arXiv:2610.22538

    AI summaryLarge reasoning models that use test‑time computation perform poorly on everyday logical reasoning tasks that diverge from formal logic. Chain‑of‑thought prompting on non‑reasoning models even outperforms comparable reasoning models, showing that current reasoning models lack reliable everyday reasoning.

31. FRACTAL: A Self-Similar Multi-Agent Orchestration Architecture with Formal Communication, Convergence, and Depth Bounds

    Ohad Dan, Abhinav Taliyan · arXiv:2610.22561

    AI summaryFRACTAL defines a self‑similar multi‑agent hierarchy that recursively decomposes goals. It proves linear intra‑node communication, a diversity‑saturation orchestration rule, supermajority convergence with depth bounds, and a Bayesian escalation mechanism for cross‑layer coordination.

32. Repairing LLM-Generated Specifications with Deterministic Diagnosis

    Giovanni Vaccarino, Emanuele La Malfa, Dehai Min et al. · arXiv:2610.22568

    AI summaryDRAGO replaces ungrounded LLM critiques with deterministic, formalism‑specific diagnosis, rewrites only when error weight strictly decreases, reducing regressions to 1.3% and boosting accuracy by 12.9% over zero‑shot formalization and 8.7% over LLM‑diagnostic repair across five benchmarks.

33. Mitigating Demographic Disparities in Explanation-Human Alignment via Rationale Regularization

    Suryansh Talukdar (Purdue University Fort Wayne), Taha Sarwat (Purdue University Fort Wayne), Jonathan Rusert (Purdue University Fort Wayne) · arXiv:2610.22586

    AI summaryThe paper introduces Plausibility‑EFG, a metric for measuring explanation fairness in hate‑speech detection, and shows that adversarially trained RoBERTa (ADV‑RoBERTa) improves both classification and explanation alignment while reducing demographic disparities in token attributions.

34. 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.

35. At the Limits of Computability: Steps Towards Autonomous AI Discovery on the Busy Beaver Frontier

    Marco Dos Santos, Vivek Veeriah, Alex Havrilla et al. · arXiv:2610.22603

    AI summaryAutonomous Gemini agents discover and formally prove compact “accelerations” for eleven unhalting Busy Beaver machines in BB(6), providing exact execution summaries while leaving halting status open, thus advancing open‑ended AI‑driven mathematical discovery.

36. Three LiDARs, Twenty-Five Datasets: Cross-Source Generalization in Tree-Species Recognition

    Jae Joong Lee, Bedrich Benes · arXiv:2610.22606

    AI summaryA new LiDAR benchmark, FOR-species20K, aggregates 25 datasets of 33 tree species. A ConvNeXtV2 model reaches 85% accuracy, but cross‑source generalization drops to 60%, showing the leaderboard score overestimates real-world transfer.

37. StratFL: Countering Covert Model Poisoning Attacks with Strategic Federated Learning

    Manoj Twarakavi, Venkata Sriram Siddhardh Nadendla · arXiv:2610.22626

    AI summaryThis paper introduces StratFL, an adversarial search algorithm combined with belief update methods to defend federated learning against intermittent covert model poisoning attacks. Modeled as an imperfect repeated game, StratFL improves model accuracy by 10 percent over existing defenses.

38. 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.

39. A Plausible Explanation Is Enough to Mislead Multimodal Judges

    Anglin Cai, Zhiqi Wang, Xiaozhi Yang et al. · arXiv:2610.22631

    AI summaryMultimodal LLM judges are easily misled by plausible explanations even when answers are fixed. Across 14 judges, ordinary or chain‑of‑thought justifications raise acceptance rates by ~13 percentage points, showing a “plausibility capture” bias that can inflate evaluation scores.

40. ToMStories: Interpreting Theory-of-Mind in Synthetic Narratives

    Melanie Subbiah, Chia-En Jacklyn Tsai, Kathleen McKeown · arXiv:2610.22632

    AI summaryThe paper introduces ToMStories, a 6k synthetic narrative dataset for testing language models’ inference of characters’ psychological states (emotion, fear, motivation, wound). Models up to 72B parameters match humans on multiple‑choice tasks but fall short on open‑ended questions, revealing simplifying assumptions in model interpretation.

41. Measuring Mechanistic Uncertainty in Large Language Models via Causal Robustness

    Aneesh Komanduri, Majid Afshar, Timothy A. Miller et al. · arXiv:2610.22636

    AI summaryMeasuring mechanistic uncertainty in LLMs by defining the causal robustness margin—the smallest internal intervention that changes a model’s output—using Jacobian Lens vectors. Experiments show incorrect or uncertain answers have smaller margins, while correctly learned facts gain larger margins, complementing traditional uncertainty measures.

42. Hard Instances on Demand

    Jamie Fravel, Martin Koutecký, Pavlo Tsitsei · arXiv:2610.22637

    AI summaryThis paper frames instance generation as an optimization problem, using black-box optimization to automatically search for inputs that maximize a solver's running time. The resulting evolved instances are significantly harder than random draws and often rival decades-old human-curated benchmarks across multiple NP-hard problems.

43. Diverse Reasoning Paths Matter: Symbolic Equivalence Augmentation for Enhancing RL Exploration

    Xiaoou Liu, Tiejin Chen, Hanyang Chen et al. · arXiv:2610.22665

    AI summaryThe paper introduces EquivAug, a data augmentation framework that uses symbolic equivalence transformations to generate diverse reasoning paths for supervised fine-tuning. Experiments show that increasing chain-of-thought diversity during fine-tuning significantly improves downstream reinforcement learning reasoning performance.

44. From Relevant Memories to Sufficient Evidence: Rethinking Memory Utilization in LLM Agents

    Shixuan Li, Wei Yang, Peiyu Zhang et al. · arXiv:2610.22668

    AI summaryThis paper introduces QEMem, a framework that bridges the relevance-sufficiency gap in LLM agent memory by constructing query-specific evidence sets rather than relying on individual memory retrieval. QEMem outperforms existing baselines, achieving 91.43% accuracy on the LoCoMo benchmark.

45. Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation

    Adamo Orsini, Seyed Mehdi Ayyoubzadeh, Sadman Rakib Pinon et al. · arXiv:2610.22672

    AI summaryThe paper introduces TAG-Nav, a framework that equips mobile GUI agents with persistent graph-based destination memory built via offline exploration. Evaluated on a new navigation benchmark, DestBench, graph memory significantly improves strict navigation success rates across multiple agent backbones.

46. FlexUMM: Flexible Representation Learning for Unified Multimodal Models

    Yinyi Luo, Haoyue Bai, Yi Ge et al. · arXiv:2610.22682

    AI summaryFlexUMM dynamically allocates shared versus specialized capacity in unified multimodal models, deciding when to share, route, or stop computation. It reduces cross‑function competition, improves accuracy, and cuts unnecessary processing, serving as a plug‑in for any existing UMM.

47. 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.

48. Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries

    Anna Mazhar, Seiji Maekawa, Nikita Bhutani · arXiv:2610.22688

    AI summaryThis paper introduces BRIA, an agent that isolates candidate tool investigations into separate execution branches before selectively adopting successful results. BRIA significantly outperforms strong baselines on large tool library benchmarks while maintaining efficient token costs.

49. Repair, Retry, and the Cost of Hindsight: Verifier Feedback in Formal Theorem Proving

    Jack Li, Sophia Li · arXiv:2610.22698

    AI summaryThe study tests verifier‑guided self‑repair in Lean by fixing a failed attempt and varying feedback types. It finds that showing the failed proof often hurts success, while providing the compiler error can partially recover performance; overall, verifier feedback improves repair but cannot beat instruction‑only resampling.

50. NoveltyScope: A Multi-Agent Research Assistant for Evidence-Grounded Novelty Assessment of Scientific Papers

    Yueran Tian, Huizhi Liang · arXiv:2610.22703

    AI summaryNoveltyScope is a multi-agent system that assesses the novelty of scientific papers by retrieving prior work, gathering evidence through adversarial LLM roles, and producing a structured report with a novelty score. It achieves lower scoring errors than single-LLM baselines and standard vector similarity on a benchmark of scientific papers.

51. Brain-Aligned Multimodal LLMs for Social and Emotional Understanding of Cartoon Movies

    Yiran Huang, Zeyong Zhang, Bharat Biswal et al. · arXiv:2610.22706

    AI summaryThis paper evaluates multimodal LLMs against human behavioral ratings and fMRI brain responses during cartoon movie viewing. The main result is that while models capture aspects of human emotional dynamics and predict brain activity in some visual and attention regions, their alignment remains limited and stimulus-dependent.

52. GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations

    Nicholas Saban, Charvi Bannur, Mariya Vasileva et al. · arXiv:2610.22708

    AI summaryGlobeBench provides 1,632 text‑based world‑model tasks across five domains, evaluating LLMs on predicting environment responses to actions. GPT‑6 Astra tops the set with 49.57% success, while most tasks remain unsolved, highlighting a large gap in general world‑model ability.

53. ARBOR: Locating Reasoning Failures in Long Chains of Thought through Reasoning State Tracking

    Huihan Li, Siyuan Wang, Naga Vamsi Ramana Dinavahi et al. · arXiv:2610.22717

    AI summaryARBOR tracks long‑chain reasoning as a tree of states, attributing failures to specific step‑state relations. It refines overthinking, underthinking, and unsupported deductions into distinct subtypes and separates failure steps from at‑risk steps, enabling better mitigation.

54. LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs

    Guy Bar-Shalom, Mor Geva, Fabrizio Frasca et al. · arXiv:2610.22723

    AI summaryLEMON learns equivariant attention representations with a GNN, enabling zero‑shot monitoring of misbehavior across 16 LLMs. It outperforms baselines by up to 14 ROC‑AUC points on hallucination and jailbreak tasks, with negligible runtime cost.

55. 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.

56. Graph Serializations for Reasoning LLMs: A Study Across Knowledge Extraction and Structural Reasoning

    Julien Yu, Mihaela Bornea, Avirup Sil et al. · arXiv:2610.22734

    AI summaryGraph serialization heavily influences LLM performance on structural reasoning tasks. The study evaluates 15 formats across nine advanced LLMs, finding that optimal serialization depends on model, query, and task, while knowledge extraction is largely unaffected.

57. R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop

    Mingda Zhang, Yun Huang, Qiang Huang et al. · arXiv:2610.22740

    AI summaryR²Orch iteratively trains an orchestrator by self‑play, rewarding the proposer for tasks that keep producing diverse outcomes as skills are removed. This recursive learnability and skill‑withdrawal loop leads to continuous self‑improvement and outperforms baselines on QA, math, code, and interactive tasks.

58. Between Spoilers and Guesswork: Avoiding Over-Specification in Coding Benchmarks

    Yuanzhe Liu, Bastian Lipka, Avraham Shinnar et al. · arXiv:2610.22760

    AI summaryThis paper analyzes how ground-truth leaks cause over-specification in repository-level coding benchmarks. The authors demonstrate that balancing hints between guesswork and spoilers improves benchmark reliability and recommend techniques for deriving parsimonious task specifications.

59. StoBasket: Joint Upsell and Cross-Sell Optimization for Contribution Margin

    Xin Heng · arXiv:2610.22768

    AI summaryThis paper introduces StoBasket, a joint stochastic optimization method for retail upsell and cross-sell recommendations that maximizes contribution margin by coupling customer response channels. Simulations show it captures nearly all attainable margin, outperforming traditional margin-weighted popularity and heuristic baselines by up to 64%.

60. When Can You Trust an Agent's Own Verifier? Stress-Testing Counterfactual Rationale Verification Under Search Pressure and Oracle Access

    Jonghyun Lee, Moonsu Kim, Sunghwan Park · arXiv:2610.22779

    AI summaryThis paper introduces a stress-testing framework to evaluate whether automated agent verifiers remain reliable when agents optimize against them. The results show that high search pressure and premature verifier access can cause agents to game the verification signal while actual faithfulness declines.

61. 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.

62. Self-Evolving Coding Agents via Harness-Policy Optimization

    Zhepei Wei, Wenlin Yao, Baolin Peng et al. · arXiv:2610.22791

    AI summaryThe paper introduces Self-HPO, a recursive framework that jointly optimizes the policy and environment harness for LLM coding agents using reinforcement learning. This co-evolutionary approach significantly improves resolve rates across multiple coding benchmarks compared to single-component optimization baselines.

63. Prune the Leaky Neurons: Mitigating Private Data Leakage in Fine-Tuned LLMs via Joint-Embedding-Guided Pruning

    Mohamed Shaaban, Mohamed Elmahallawy · arXiv:2610.22792

    AI summaryThis paper introduces JEPA-Prune, a post-hoc neuron pruning method that mitigates private data leakage in fine-tuned LLMs using a PII-free reference model. Pruning a small fraction of identified leaky neurons drastically reduces privacy leakage while better preserving model utility compared to existing baselines.

64. STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases

    Lieqi Liu, Rui gao, Jia-Chen Gu et al. · arXiv:2610.22795

    AI summaryThis paper introduces STRIDE, a benchmark for evaluating spatiotemporal reasoning in autonomous driving edge cases. The results show that current vision-language models struggle significantly with safety-critical trajectory prediction and scene-context awareness, achieving low accuracy despite strong general VQA capabilities.

65. Compression Precedes, Content Decides: Emergent Misalignment in Reward Models

    Dhruv Puri, Nikhil Prakash · arXiv:2610.22805

    AI summaryFine-tune reward models for harmful preference in one domain; they generalize to seven other domains while losing only ~5 points on RewardBench, outperforming the unmodified model. The model’s preference inversion is caused by geometric compression and content-specific rotation in last-token representations.

66. World Action Critic (WAC) Model: Joint Foresight and Criticism for Policy Learning

    Xiangteng Zhang, Haoyuan Liu, Yang Guan et al. · arXiv:2610.22825

    AI summaryThe paper introduces the World Action Critic (WAC) model, jointly learning action-conditioned future prediction and counterfactual consequence evaluation. It shows that critic supervision improves predictive representations, achieving 93.4 PDMS on NAVSIM‑v1 in autonomous driving.

67. The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls

    Lucas Resck, Jingyi Sun, Qianli Wang et al. · arXiv:2610.22841

    AI summaryThe paper studies how Chain-of-Thought faithfulness varies with language resource level and task type. It finds that in low‑resource languages, models rely more on reasoning for knowledge‑intensive questions, supporting the Reliance Hypothesis.

68. Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors

    Carl Qi, Weijun Dong, Xiaojie Wang et al. · arXiv:2610.22852

    AI summaryAnchorQ introduces flow‑reversed latent action anchors for policy refinement, enabling efficient exploration in generalist robot policies. It achieves 37% higher real‑world success and 97% more simulated task throughput with the same interaction budget compared to state‑of‑the‑art baselines.

69. Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search

    Jianguo Huang, Zhiyi Lyu, Fuxiang Zhang et al. · arXiv:2610.22858

    AI summaryUncertainty-Aware Allocation (UAA) uses sibling action dispersion to estimate downstream redundancy and sets per‑step budgets accordingly, decoupling value and uncertainty. It reduces token usage by up to 90 % and latency by 9× on reasoning and planning tasks while maintaining accuracy.

70. Evo-B&B: Behavior-Aware LLM-Guided Joint Evolution of Branching and Node Selection Policies for MILP

    Jiayu Chen, Chentong Chen, Yuhan Liu et al. · arXiv:2610.22861

    AI summaryEvo-B&B jointly evolves branching and node‑selection policies for MILP using an LLM‑guided evolutionary algorithm. By measuring behavior similarity across B&B states, it preserves diversity and achieves competitive, generalizable solving performance on standard benchmarks.

71. MS-Eval: Benchmarking Sub-Agent Delegation and Orchestration in MCP-Intensive Working Scenarios

    Bowen Ding, Qixin Wang, Lexing Ruan et al. · arXiv:2610.22869

    AI summaryMS‑Eval benchmarks sub‑agent delegation for multi‑tool tasks, revealing that specialization improves performance, open‑source models compete well, and MAD‑SAE boosts efficiency for open‑source but raises cost and hurts closed‑source agents.

72. ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers

    Kun Zhai, Teng Li, Yunhao Feng et al. · arXiv:2610.22871

    AI summaryShadowFPT backdoors federated prompt tuning by training a local “Shadow Trigger” that steers the frozen CLIP backbone’s representations toward a target class, while only uploading benign prompt updates. The attack boosts success to ~90% with negligible clean accuracy loss.

73. Look Back, Correct Forward: Error Feedback Calibration for Time Series Forecasting

    Haonan Yang, Yulin He, Zhuo Li · arXiv:2610.22872

    AI summaryError Feedback Calibration (EFC) post‑hoc corrects a frozen forecaster by using revealed past errors to build a causal error diagonal, then transports cumulative error to predict future residuals and learns only the residual structure, yielding broader performance gains while limiting over‑correction.

74. Cooperation Without Adaptation: LLM Agents in a Collective-Risk Dilemma

    Trung-Kiet Huynh, Dao-Sy Duy-Minh, Chi-Nguyen Tran et al. · arXiv:2610.22882

    AI summaryLLM agents largely keep a fixed contribution pattern in a collective‑risk dilemma, rarely adapting to changing risk or partner behaviour. Even when they calculate the payoff‑maximising action, their actual contributions stay almost constant, limiting group success.

75. 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.

76. 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.

77. 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.

78. 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.

79. HAPACT: A Benchmark For Human-Centric Physical Impact Localization in Movies

    Youngrae Kim, Yejin Jang, Eunho Kim et al. · arXiv:2610.22895

    AI summaryThe paper introduces HAPACT, a dataset of 2.8K movie clips with 20K fine‑grained human‑annotated impact frames, and defines a new task: temporal and SMPL‑X spatial localization of physical impacts on people. Results show existing models perform poorly, highlighting the task’s difficulty.

80. 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.

81. COMET: Decoupled Distillation, Routing, and Capacity Control for Task-Agnostic Continual Vision--Language Learning

    Seungyoon Woo, Keonhee Park, Gunhee Kim · arXiv:2610.22907

    AI summaryCOMET trains a frozen VLM teacher to provide a shared image‑text feature reference, then uses tri‑affinity distillation to preserve multimodal geometry and a bandit‑guided routing of task experts to avoid interference and capacity limits. On X‑TAIL, this decoupled approach boosts continual VLM accuracy by up to 10 %.

82. Reward Engine: Reward-Centric Self-Evolution for Open-Ended Tool Agents

    Litong Feng, Ji Qi, Yiran Cheng et al. · arXiv:2610.22915

    AI summaryReward Engine builds a frozen set of testcases and rubric-based checks from initial trajectories, using them as a stationary reward that aligns 80‑85% with benchmark graders and boosts three‑run consistency from 45% to 54% and from 78% to 84%.

83. Sep-Latent: Separator Tokens as Natural Compression Anchors for Implicit Reasoning

    Yinlong Xu, Jianfeng Huang, Zhengdong Zhou et al. · arXiv:2610.22924

    AI summarySep‑Latent compresses implicit reasoning by using separator tokens as semantic anchors and an attention‑decay scheme to adaptively terminate latent thinking. On GSM8k‑Aug it reaches 56.8 % accuracy with only 5.4 tokens, closing the gap to explicit CoT while vastly reducing overhead.

84. Multiple Turns, Short Horizons: Measuring Effective-Horizon Compression in Jailbreaks

    Zhipeng Wei, Michael Lee, N. Benjamin Erichson · arXiv:2610.22928

    AI summaryMultiple‑turn jailbreaks can be compressed: on average only about half of the original query sequence is needed to reach the same attack progress. Shorter histories recover the best point and improve attacker training when earlier queries are down‑weighted.

85. 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.

86. 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.

87. NL2SQL with Iterative Self-Refinement

    Anjali Vemula, Akshat Dasula · arXiv:2610.22968

    AI summaryMulti‑agent NL2SQL framework reduces hallucinations and syntax errors by iteratively refining queries with domain relevance, schema grounding, KPI awareness, execution feedback, and reverse‑engineered consistency checks. It reaches 90 % correct queries on 60 examples using Mistral across seven LLM calls.

88. DiffusionRVGE: Reinforcement Learning via Regression-Based Value Gradient Estimation

    Jiayuan Wang, Tongyao Pang · arXiv:2610.22973

    AI summaryDiffusionRVGE trains diffusion models by regressing centered rewards on centered denoised images, avoiding expensive reward gradients or trajectory backpropagation. The method matches or exceeds DiffusionNFT in multiple metrics while cutting training compute to about one‑fifth and speeding single‑reward learning 2–7×.

89. Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs Inference

    Zhenxu Tian, Kebin Liu, Zhengwu Yang et al. · arXiv:2610.22985

    AI summaryPrism Attention introduces a coarse-to-fine sparse attention framework that refines shared cross-layer KV indices to track layer-specific token shifts. It achieves state-of-the-art long-context performance while delivering a 3.4 times end-to-end decoding speedup over FlashAttention at a 128K context length.

90. From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents

    Xianglin Yang, Yechao Zhang, Jiahao Liu et al. · arXiv:2610.22987

    AI summaryThis paper demonstrates that prompt injection in coding agents can trick users into approving overly broad permission rules, bypassing human oversight in 58% of tests and enabling silent, malicious goal completion.

91. 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.

92. GOLD: Geometric Optimized Latent Diffusion for Structure-Aware RNA Inverse Folding

    Qi Si, Xuyang Liu, Penglei Wang et al. · arXiv:2610.22992

    AI summaryThis paper introduces GOLD, a latent diffusion model for RNA inverse folding that jointly generates sequences and continuous side-chain geometries. By operating on a unified manifold with gradient-guided physical constraints, GOLD achieves state-of-the-art performance in sequence recovery and 3D foldability.

93. 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.

94. Efficient CPU-GPU collaborative MoE inference via Adaptive Expert Assignment

    Wenhui Lv, Yunming Liao, Hongli Xu et al. · arXiv:2610.23006

    AI summaryThis paper introduces a CPU-GPU collaborative inference system for memory-constrained Mixture-of-Experts models that adaptively assigns offloaded experts to either the CPU or GPU based on token loads. This approach balances execution paths to improve prefill speed by up to 51.9% and decode throughput by up to 39.5%.

95. Adaptation of Computer Use Agents to New Software with User Manuals

    Chulhoon Yoon, Kyungri Park, Jinyeong Yun et al. · arXiv:2610.23012

    AI summaryThe paper introduces MAGE, a framework that uses software user manuals to guide computer use agents through procedural planning and generate synthetic training trajectories without human demonstrations. MAGE improves task success rates across six unseen software applications compared to strong baselines.

96. Causal Value Attribution of Marketing Labels in LLM-Generated Ads: A Deconfounder Approach

    Danhong You (Eric), Wei Jiang (Eric), Zhiqiang (Eric) et al. · arXiv:2610.23025

    AI summaryThis paper introduces D-VAE, a counterfactual inference framework that uses a multi-treatment deconfounder to causally attribute economic value to upstream human expertise in LLM-generated marketing. Applied to ad generation, the method identifies optimal skill configurations that increase user engagement by 32.7%.

97. ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting

    Yuxuan Zheng, Zhaoyang Li, Purui Liu et al. · arXiv:2610.23027

    AI summaryASE constructs textual factors, evaluates their conditional contribution, and selects numerical‑textual predictors; on TESS datasets it beats published scores and on CSI 300 raises RankIC over numerical and shuffled‑text baselines.

98. Where Research Ideation Begins: Forward-Citation Supervision for Actionable Limitation Generation

    Liyang Zhao, Yong Cao, Haibin Ling · arXiv:2610.23038

    AI summaryThe paper introduces ActLim-Bench, a forward‑citation supervised dataset of 18,279 human‑verified limitation triples, and a three‑stage post‑training pipeline that boosts LLM recall and yields limitations preferred by 64.9% of humans, improving downstream ideation systems.

99. Where Does the Performance Go? Understanding and Recovering Performance Regressions in C-to-Rust Translation

    Zhiqiang Yuan, Chong Wang, Zhenhao Zhou et al. · arXiv:2610.23044

    AI summaryTranspiling C to Rust with Crust often slows code by up to 62%; the study identifies nine recurring performance regressions. An LLM‑guided tool, LLaM-Optimize, applies rule‑based rewrites to recover and often exceed original C performance.

100. SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

     Kai Sun, Peibo Duan, Zhanglu Yan et al. · arXiv:2610.23046

     AI summarySpikeQ applies spike‑guided post‑training quantization for SNNs, selecting weight levels by minimizing spike errors and using a reset‑aware grid for membrane potentials to avoid threshold flips. It retains full‑precision accuracy within 2.59 points at 4 bits, shrinking memory 7.8× and energy 6.4×.

101. Distill Until Divergence: Error-Truncated Hindsight Distillation for Agentic RL

     Yibin Huang, Conghui Zhu · arXiv:2610.23048

     AI summaryThe paper introduces Distill Until Divergence, which uses positive and negative hindsight skills to give fine‑grained credit for long‑horizon RL, routing them based on outcomes and error boundaries. Experiments show the method matches or exceeds prior performance while better allocating credit.

102. Agentic In-Context Learning: What Algorithms Can Transformers Learn in Closed-Loop Interaction?

     Shiguang Wu, Yaqing Wang, Quanming Yao · arXiv:2610.23051

     AI summaryAgentic ICL studies transformers that learn from closed‑loop interaction, showing that causal transformers can approximate active learning, bandits, and online RL over finite horizons, and that process‑reward training aligns learned policies with outcome‑optimal actions.

103. 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.

104. 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.

105. Rethinking GRPO for Large Reasoning Model Safety Alignment

     Jiawei Chen, Yi Zhang, Jiawei Du et al. · arXiv:2610.23085

     AI summaryBeacon improves LRM safety alignment by recovering reliable safety signals (SSR) and decoupling safety from utility (SUDA), eliminating advantage collapse and unsafe diversification; it reduces safety errors by up to 32.83 pp and maintains low over-refusal on Qwen3/DeepSeek benchmarks.

106. WebCoder: Scaling General Frontend Development with Artifact-Grounded Agentic Reinforcement Learning

     Zijun Dou, Zhen Yang, Xijun Liu et al. · arXiv:2610.23104

     AI summaryWebCoder trains a general frontend agent using artifact‑grounded reinforcement learning, rewarding progress on a shared executable artifact and redistributing advantage across long action sequences. It outperforms Claude‑Opus‑4.8 and GPT‑5.4 on Webdev‑Bench, raising satisfaction from 64.8 to 72.8.

107. LexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation

     Wenbin Zhang, ShangZhan Li, Bingxiang He et al. · arXiv:2610.23141

     AI summaryLexPilot evaluates how well legal LLMs can proactively gather missing context from underspecified layperson queries, linking interaction to claim‑level verdict accuracy on 432 authentic Chinese civil judgments. Current models drop 23.8% accuracy versus fully specified cases, revealing a need for more user‑centered legal AI.

108. 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.

109. 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.

110. QuAD: Quantitative Assessment via Distribution-Alignment for Qualitative Data Quality Enabled by a Content-Aware Multi-Agent System

     Wenshuo Zhang, Leixian Shen, Qiqi Zhou et al. · arXiv:2610.23153

     AI summaryQuAD automatically evaluates qualitative response sets by comparing a reconstructed cluster distribution to a researcher‑defined analytical lens, providing four‑quadrant metrics that signal fitness for purpose. It uses a multi‑agent system for content‑aware scoring, validated on synthetic and real survey data.

111. Algorithm Modelization: Transforming Optimization Algorithms into Evolvable Intelligence

     Jian Yao, Zhiyao Zhang, Yichen Li et al. · arXiv:2610.23161

     AI summaryAlgorithm Modelization turns iterative optimizers into trainable models by treating algorithm traces as data. The resulting autoregressive model learns from multiple teachers, surpasses them, and generalizes to new problems and higher dimensions, enabling continual improvement via learning.

112. 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.

113. LongForge: Measuring and Improving Long-Context Instruction Data via Evidence-Assertion Reasoning Graphs

     Ziyang Chen, Junlong Jia, Xing Wu et al. · arXiv:2610.23167

     AI summaryLongForge constructs evidence‑assertion reasoning graphs to audit long‑context instruction data, revealing breadth and depth deficits in existing datasets. By gating on correctness/completeness and mixing over measured breadth/depth, it improves downstream performance by 2.2‑4.7 points.

114. 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).

115. 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.

116. 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.

117. OfficeTown: Generative Organizations for Evolving Agent Intelligence

     Hongchen Wei, Daohong Ying, Zhi Hong et al. · arXiv:2610.23194

     AI summaryOfficeTown simulates professional organizations from a natural‑language description, generating tasks, artifacts, and feedback that persist across assignments. Using its 406‑task benchmark, fine‑tuned Qwen3.5‑9B outperforms baselines on external work benchmarks, showing the benefit of organization‑driven data.

118. FuzzInsight: Navigating the Efficiency-Effectiveness Trade-off in Coverage-Guided DNN Fuzzing

     Vivek Vekariya, Fathiyeh Faghih, Alexander Pretschner · arXiv:2610.23195

     AI summaryFuzzInsight analyzes coverage‑guided fuzzing for DNNs, showing that existing seed‑selection methods waste most inputs. It proposes Equalized Sampling, boosting seed queue use to 77% (+13pp), and recommends LSCD/LSC/MDSC coverage criteria for efficient, broad exploration, improving robustness by up to 77%.

119. Framing Spatial Hearing within the Free Energy Principle: A Hierarchical Model of Perceptual Learning

     Roberto Barumerli, Nils Meyer-Kahlen, Lorenzo Picinali · arXiv:2610.23196

     AI summaryThe paper applies the free energy principle to spatial hearing by modeling listeners’ dynamic beliefs about sound direction across trials with visual feedback. The hierarchical generative model shows that learning improves lateral, polar, and hemifield precision, revealing distinct dynamics for vertical localization.

120. Validating LLM-generated political networks with survey-reported partisan exposure

     Marco Tonin, Michele Tizzoni, Bruno Lepri et al. · arXiv:2610.23199

     AI summaryThe study uses LLMs to generate political networks from ANES survey personas and validates them against respondents’ reported partisan exposure. It finds that while aggregate structure varies, LLMs often misestimate local partisan environments, over‑ or under‑exposing certain groups.

121. Not All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning

     Ming Shen, Tianjun Yao · arXiv:2610.23213

     AI summaryTest‑time scaling assumes new compute from a single model yields independent evidence, but errors are correlated across reasoning steps and self‑evaluation. DIA uses disagreement between intermediate and final predictions to trigger a small heterogeneous panel that resolves the query, outperforming full chain‑of‑thought while using fewer tokens.

122. qGAMMA: A Framework for Optimal DNN Mapping on Multi-Precision Accelerators

     Leandro Fiorin, Luigi Altamura, Cristina Silvano · arXiv:2610.23222

     AI summaryqGAMMA is a framework that uses a domain-specific genetic algorithm and an analytical performance model to find the best mapping and accelerator configuration for mixed-precision DNNs on spatial accelerators. It reduces performance, energy, and resource trade-offs for various DNN workloads.

123. 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.

124. MGE-PO: Entropy-Aware Credit Assignment for Mathematical Reasoning

     Jihao Huang, Guangjia Chai, Zhirui Chen et al. · arXiv:2610.23245

     AI summaryMGE‑PO introduces a four‑quadrant advantage shaping that rewards correct, low‑entropy responses more and penalizes incorrect, high‑entropy ones, using an entropy‑based curriculum. It boosts macro‑avg@32 by up to 11 % and macro‑pass@32 by 4‑11 % on Qwen models.

125. Preventing Data Hallucinations in Agentic AI for Optical Network Troubleshooting

     Jeremias Dötterl, Martin Kuipers · arXiv:2610.23252

     AI summaryAgentic AI for optical network troubleshooting incorporates data references within the architecture, preventing hallucinations. This improves interpretability and trustworthiness of root‑cause identification and reasoning, leading to more reliable automated diagnostics.

126. Beyond Reward Scoring: Online Multi-Agent Preference Learning for Multimodal Response Generation

     Ziyun Zou, Jiwoo Hong, Yifan Lu et al. · arXiv:2610.23256

     AI summaryThe paper introduces PrefMind, an online multi-agent reinforcement learning framework that jointly trains a response generator and a preference critic using human preference feedback. This approach improves both preference judgment and downstream multimodal response generation compared to traditional scalar reward models.

127. 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.

128. 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.

129. Annotrace: A Workbench for Step-Level Annotation of Agent Traces

     Arun Josephraj Arokiaraj, Chenyi Huang, Sahan Bulathwela et al. · arXiv:2610.23284

     AI summaryAnnotrace converts multi-format agent traces into nested step lists and lets users define a codebook; a frontier AI proposes labels and then humans refine them. In a multi‑agent QA dataset, it exactly reproduced 235 manual annotations and achieved 91% agreement with κ = 0.82.

130. CoTA++: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective

     Qiyan Zhao, Xiaofeng Zhang, Xu-Yao Zhang et al. · arXiv:2610.23290

     AI summaryThis paper identifies the repeat curse in diffusion-based multimodal LLMs caused by approximate key-value caching and proposes CoTA++, a training-free framework of re-anchoring, refresh, and voting mechanisms. CoTA++ reduces the sample repetition rate to at most 17.0% while recovering benchmark accuracy and preserving caching speedups.

131. FlashReFocus: Interactive Image Refocusing in Seconds

     Ching-Heng Cheng, Chia-Ming Lee, Ming-Ching Chang et al. · arXiv:2610.23297

     AI summaryFlashReFocus uses a single diffusion step to recover detail and a lightweight renderer conditioned on a defocus map for interactive single-image refocusing. It achieves state-of-the-art visual quality while running hundreds of times faster than existing generative methods.

132. Optimist Deep Research: An Iterative, State-Centric Agent for Autonomous Literature Research

     Gihan Panapitiya · arXiv:2610.23299

     AI summaryOptimist Deep Research is an open-source agent that models literature search as an iterative, state-centric process using a mutable subquestion tree and evidence store. It successfully automates multi-source document retrieval, citation snowballing, and rigorous evidence synthesis into fully cited reports.

133. Designing for Imperfect Prompts: A Simulation-Based Evaluation of an LLM-Based Enterprise Data Assistant Using Knowledge Graphs

     Niklas Wagner, Alexander Mädche · arXiv:2610.23304

     AI summaryThe paper introduces a semantic mediation layer for enterprise data assistants that resolves imperfect user prompts. Using a knowledge graph derived from the data lake, it outperforms a schema‑based similarity approach in precision, recall, F1, and intent preservation.

134. SimOPD: Relational Knowledge Transfer for On-Policy Distillation

     Kai Tang, Xing Zheng, Haobo Wang et al. · arXiv:2610.23306

     AI summarySimOPD transfers teacher internal knowledge for on-policy distillation by matching response similarity distributions stored in paired memory banks. This approach avoids cross-model projections, retaining student representational freedom and improving performance by up to 3.1 percentage points.

135. MobileInject: Evaluating Reasoning-to-Action Security of MLLM-Based Mobile Agents under Cross-Modal Consistency Injection

     Songping Wang, Peiyang Li, Yi Huang et al. · arXiv:2610.23307

     AI summaryMobileInject benchmarks multimodal injection attacks on mobile agents, revealing they are highly vulnerable to cross‑modal consistency injections that hijack early reasoning. The study introduces new attack methods, a diagnosis framework, and demonstrates widespread failures across 11 scenarios.

136. Self-Reflective Chain-of-Thought (SRCoT): Evidence-Based Text Generation via a Closed-Loop Retrieve--Reflect--Revise Pipeline on Edge LLMs

     Fabian Broszio, Michael Hoffmann, Adrian Paschke · arXiv:2610.23311

     AI summarySRCoT implements a closed Retrieve--Reflect--Revise loop on edge LLMs that separates knowledge stores and uses BM25 retrieval to produce source‑labeled drafts, gap markers, and follow‑up queries, as shown in synthetic EMS and Munich‑Sim case studies.

137. From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory

     Shangze Li, Yanyu Yao, Derong Xu et al. · arXiv:2610.23313

     AI summaryMemCompile introduces a query-conditioned evidence compilation framework that decouples memory addressing from evidence presentation for long-term conversational agents. It achieves 80.52 percent Judge Accuracy on LoCoMo while establishing a superior quality-cost Pareto frontier.

138. AgentOnto: Ontology-Guided LLM Agents for Enterprise Data Analytics

     Dayan Pan, Bincheng Zuo, Xiaojian Wang et al. · arXiv:2610.23317

     AI summaryAgentOnto builds an ontology from databases, docs, and example tasks, guiding an LLM to perform accurate enterprise analytics and incorporate feedback. It improves analysis accuracy and adaptability, outperforming baselines on two public datasets and a real‑world deployment.

139. Robust Function Calling for LLM Agents Under Interface Drift

     Aniruddha Chattopadhyay, Kanishka Halder, Partha Sarathi Roy · arXiv:2610.23342

     AI summaryRobust Function Calling for LLM Agents Under Interface Drift introduces SchemaShield-Lite, a lightweight repair layer that uses schema validation, retrying missing tool names, and one‑shot constrained repair. It boosts Qwen3.5‑9B’s DICE‑300 accuracy from 0.717 to 0.813 and BFCL‑200 from 0.580 to 0.805 without retraining.

140. Language Models Learn to Improve Language Models

     Minhua Lin, Zijun Wang, Zewen Liu et al. · arXiv:2610.23344

     AI summaryLM2LM trains an LLM to propose language‑model improvements via RL using a frozen research world model as a surrogate reward. It outperforms GPT‑6‑sol on held‑out tasks while reducing data‑generation cost by 33% and generalizes to unseen model families.

141. OliO: ODE-based Linear Transition Operator for Self-Supervised Time Series Forecasting

     Kwangryeol Park, Seulki Lee · arXiv:2610.23345

     Self-supervised representation learning has advanced time series forecasting by capturing robust features from unlabeled data. However, existing contrastive and masking-based methods often struggle to explicitly model the underlying temporal flow, either by treating time lag as…

142. Monitor-Validity Failures: LLM Safety Evidence Paths at Evolving Framework Boundaries

     Junchi Liao · arXiv:2610.23362

     AI summaryThe paper audits LLM safety frameworks for “monitor‑validity failures,” where evidence of policy violation is captured but not enforced. Using canary contracts it uncovers 18 such gaps across 11 safety projects, demonstrating that current runtime monitors often miss critical enforcement boundaries.

143. AWAM: Active Trajectory-Conditioned View Imagination for Robust Robot Manipulation

     Jiawei Geng, Zhenyang Liu, Yuzheng Wu et al. · arXiv:2610.23396

     AI summaryAWAM learns to plan camera trajectories and imagine future views, providing a spatial prior that improves manipulation under occlusion. It attains 75.9% on LIBERO‑Occ and 94.4% in real‑robot trials, showing robust performance in cluttered, open‑world settings.

144. 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…

145. ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals

     Rajdeep Pandhare, Delzaan Bharucha · arXiv:2610.23443

     Efficient expert routing is a fundamental challenge in Mixture-of-Experts (MoE) systems, particularly for heterogeneous collections of independently trained Large Language Mod els (LLMs). Existing routing approaches typically rely on jointly trained gating networks, prompt-based…

146. 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.

147. GatewayBench: Unified Capability and Security Evaluation Platform for MCP Gateways

     Yitian Zheng, Zhaorun Chen, Haoyu Li et al. · arXiv:2610.23449

     AI summaryGatewayBench evaluates 524 tasks and 11,118 real-world tools to compare flattened vs. hierarchical MCP gateways, showing significant differences in task success, tool-use efficiency, and attack resilience that vary with model and harness, guiding secure, efficient gateway design.

148. S-GAP: Stealthy Graph Attack Primitives for Graph Large Language Models

     Wenda Li, Tongya Zheng, Shunyu Liu et al. · arXiv:2610.23465

     Graph Large Language Models (GraphLLMs) enable powerful semantic understanding and reasoning over Text-Attributed Graphs (TAGs), while their semantic reasoning may exhibit vulnerabilities to perturbations in node text and graph structure under adversarial attack evaluation.…

149. 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…

150. 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…

151. LocalIntent: Outcome-Specific Structured Generation for Backend-Compatible Smart Building Automation with Local LLMs

     Tianyang Liu, Jiayue Wang, Zhanhao Cheng et al. · arXiv:2610.23511

     Locally deployed large language models (LLMs) can reduce reliance on remote cloud services for natural language smart building automation, but backend-facing processing requires correct outcome selection, alignment with local capabilities, and valid structured outputs. We…

152. Trapped in Your Story: Evaluating Narrative-Induced Sycophancy in Large Language Models

     Runtai Yang, Dandan Zhu, Kaiwei Zhang et al. · arXiv:2610.23525

     Large language models (LLMs) are known to exhibit sycophancy by adjusting their responses toward users' expressed views. However, real users often justify their opinions through reasons, personal experiences, causal explanations, and broader narratives, raising the question of…

153. Truth in Contrast: Calibrated Factual Evidence from Language Model Likelihoods

     Ponhvoan Srey, Jason Lucas, Pittawat Taveekitworachai et al. · arXiv:2610.23527

     Large Language Models (LLMs) encode substantial factual knowledge. Yet, estimating confidence in factual claims reliably remains challenging. We study likelihood-based factual confidence from the complementary perspectives of discrimination and calibration across multilingual…

154. From Billion-Scale to Million-Scale Language Models: Operational Capabilities for Embedded Post-Exploitation

     Gabriel Hortea, Simon Pietro Romano, Juan Tapiador et al. · arXiv:2610.23534

     AI summaryThis study evaluates seven open‑source small language models (20 M–1.5 B) on malware post‑exploitation tasks, showing that task formulation and prompting are more critical than size. Qwen2.5‑Coder 1.5B achieved up to 100 % Pass@5 on directory enumeration and backdoor generation, indicating rapid progress toward feasible AI‑assisted malware.

155. Reason-guided Human Reward Modeling for Preference-based Reinforcement Learning

     Anshul Nayak, Shahil Shaik, Jonathan Smereka et al. · arXiv:2610.23550

     Reinforcement learning (RL) policies optimized for task rewards may produce behaviors that are efficient but misaligned with human expectations, safety preferences, or intent. This paper presents a reason-guided preference-based reinforcement learning (PbRL) framework that…

156. SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents

     Jin Ye, Yifan Jiang, Lan-zhe Guo · arXiv:2610.23554

     Large language model agents increasingly rely on skills to extend their capabilities in complex tasks. However, scalable skill use requires Selective Skill Exposure (SSE): providing relevant procedural content when needed while avoiding unnecessary exposure that may disrupt…

157. Darwin-Claw: Environment-Adaptive Experience Consolidation for Recursively Self-Evolving Agents

     Mingze Yin, Xiaohan Wang, Dian Li et al. · arXiv:2610.23574

     AI summaryDarwin‑Claw is a framework that automatically distills episodic experiences into prototype vectors, using a hierarchical consolidation system to guide skill refinement and memory updates for LLM agents. It shows significant, controllable performance gains on WildClawBench.

158. RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory

     Dabin Sheng, Zhe Wu, Jinming Zhao et al. · arXiv:2610.23598

     AI summaryRL-guided Temporal Interval-driven Dual-channel Evidence Retrieval (TIDER) learns a lightweight RL locator to predict temporal windows for memory retrieval, then merges temporally grounded candidates with semantic ones. TIDER outperforms baselines by 8.82% on LifeBench and 3.66% on LoCoMo.

159. From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents

     Yewei Zhao, Derong Xu, Wei Chen et al. · arXiv:2610.23599

     AI summaryCognitive Differentiation for Self-Improving Agents (CDSE) splits a unified skill into Meta, Experience, Tool, and Script components, updates each with trajectory evidence, and coordinates them for adaptive use. CDSE outperforms baselines by 7.6 pp overall, with functional differentiation key.

160. SIT-Eval: Situation-Based Interactive Turn-by-Turn Evaluation of Conversational Agents

     Ravil Mussabayev, Saveliy Zavertyaev, Dmitriy Parpulov et al. · arXiv:2610.23612

     AI summarySIT‑Eval evaluates agents by sampling situational prompts turn‑by‑turn and scoring replies against rule‑based criteria via an LLM judge. Testing eight LLMs shows that aggregate scores mask rule‑specific failures, especially in ambiguity clarification and Socratic questioning.

161. Beyond Action Prediction: Learning Search-Free Planning in Latent Space

     Guanchen Du, Jinhao Du, Yanpeng Sun et al. · arXiv:2610.23615

     AI summaryHindsight Inverse JEPA Planning trains a goal‑conditioned inverse model and latent predictor jointly, relabelling offline trajectories as goals to shape visual representations and enforce action‑transition consistency. It achieves 86.8% success and 1.9–16.5× faster planning than search‑based methods.

162. GEMS: Molecular Structure Elucidation as Learned Search over Isomer Graphs

     Utku Umur Acikalin, Yingheng Wang, Goncalo J. Gouveia et al. · arXiv:2610.23617

     Identifying small-molecule structures from molecular formulas and tandem mass spectra remains a bottleneck in untargeted metabolomics and natural-product discovery. Even when database retrieval or de novo prediction misses the exact structure, the resulting candidates can…

163. PresenceBench: An Exploratory Evaluation of Large Language Model Adaptation to User State

     Peña Arroyo Gallego, Mauro Ruiz Moreno, Alejandro Acien et al. · arXiv:2610.23625

     AI summaryPresenceBench evaluates how LLMs adapt to a user’s functional state (readiness, flow, stability) by providing a synthetic state signal and measuring receptivity and specificity. Across four models, policy delivery improved human preference over baseline (≥0.5), but specificity could not be conclusively established.

164. TUPLE-Bench: Evaluating and Mitigating Privacy-Utility Trade-offs in Typed Decision Models

     Yiqu Yang, Qiaoyuan Zheng, Yang Tian et al. · arXiv:2610.23668

     AI summaryTUPLE‑Bench shows that typed decision models leak fine‑grained causes in 92–100 % of cases, even when causes are omitted, and full outputs reduce downstream cost but increase leakage. Workflow‑constrained privacy quantization can lower secret recovery while preserving workflow performance.

165. OwlSight: A Panoramic Monitor for Chain-of-Thought Reasoning

     Lewen Yan, Yanxu Zhu, Jie Zhang et al. · arXiv:2610.23670

     AI summaryOwlSight is a chain‑of‑thought monitor that uses a customizable “Factor” prompt to cover many tasks at fine and coarse granularity, outperforming general LLMs by 1.3–30.6 points and achieving state‑of‑the‑art results on the CoT monitorability benchmark.

166. On the possibilities and limitations of LLM-based multi-agent systems for multi-criteria decision-making in healthcare operations management

     Marco Boresta, Tommaso Giovannelli, Sepehr Namirad et al. · arXiv:2610.23673

     AI summaryThe paper develops a framework to audit large‑language‑model agents’ stakeholder preferences for multi‑criteria healthcare decisions, showing that explicit weights vary by role but converge with deliberation, while implicit pairwise choices are order‑dependent; robust optimization with preference uncertainty often selects different schedules than simple averaging.

167. GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search

     Hyunseo Kim, Suil Choe, Sukkyoung So et al. · arXiv:2610.23709

     Optimizing product listings for generative search requires language agents not only to rewrite effectively but also to continually self-improve by retaining successful strategies across diverse categories. However, existing approaches often treat structured listings as free-form…

168. 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…

169. Beyond Large Language Models: A Mental World Model Framework for Mental Modeling, Simulation, Evolution, and Interpretability

     Qingshi Wang · arXiv:2610.23722

     Understanding and simulating how subjective cognition and mental states emerge, evolve, and shape behavior is essential for advancing artificial intelligence toward deeper social cognition and general intelligence. Existing large language models (LLMs) and Theory of Mind (ToM)…

170. DrugHarness: An Evolvable Agent Harness for Computational Drug Discovery

     Yuxing Chen, Zhisheng Zhang, Cheng Tang et al. · arXiv:2610.23791

     Computational drug discovery requires agents to compose scientific operations and adapt to changing evidence and constraints. Here we introduce DrugHarness, an evolvable agent harness coupling composable scientific execution with conditional skill induction. Its environment…

171. FuTransformer: FP8 prefill and INT4 decode for Llama-style inference on Hopper

     Wei Wang · arXiv:2610.23796

     Transformer prefill reuses weights across tokens and is dominated by matrix multiplication and activation data movement; single-token decode has limited reuse and is dominated by weight and KV-cache reads. We present FUTRANSFORMER, a CUDA implementation of a Llama-style…

172. Beyond SAEs: Consistent Concept Discovery with Sparse Coding

     Jinho Choi, Hyesu Lim, Jaegul Choo et al. · arXiv:2610.23801

     Sparse autoencoders (SAEs) explain foundation model representations by decomposing activations into sparse combinations of learned dictionary atoms. However, the learned dictionaries can vary across training runs, limiting the reliability and reproducibility of SAE-based…

173. Self-Building Labs

     Jiaru Bai, Yuchi Zhao, Han Hao et al. · arXiv:2610.23826

     Self-driving laboratories (SDLs) accelerate scientific discovery by automating experimental execution and optimization. AI agents can increasingly propose new experiments, but existing SDLs can only run the procedures they were built for. Adapting them to a new experiment still…

174. AI-Driven Reproducibility and Analytical Robustness in Scientific Discovery

     Shicheng Liu, Krista Opsahl-Ong, Andrew Drozdov et al. · arXiv:2610.23845

     AI summaryWe create SEER‑Repro, a benchmark of 20 cancer‑registry studies, and an LLM agent that reproduces 77.4 % of 2,190 claims and automatically tests analytical robustness, flagging potentially consequential choices in 9.5 % of cases for expert review.

175. 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.…

176. Rethinking the design space of Vanilla Pixel Diffusion Transformers

     Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava et al. · arXiv:2610.23854

     Pixel-space diffusion models have recently re-emerged as a compelling alternative to latent diffusion, avoiding the lossy latent bottleneck and separately trained autoencoder. JiT shows that a vanilla diffusion transformer can generate directly in pixel space by replacing the…