New Computation and Language (cs.CL) papers, Wednesday 21 October 2026

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

1. OPLoRA: Orthogonal Projection LoRA Prevents Catastrophic Forgetting during Parameter-Efficient Fine-Tuning

   Yifeng Xiong, Xiaohui Xie · arXiv:2510.13003

   AI summaryThe paper introduces OPLoRA, which constrains LoRA updates to the orthogonal complement of dominant pre-trained weight subspaces using double-sided projections. This theoretically grounded approach prevents catastrophic forgetting while maintaining competitive task performance across various benchmarks.

2. FRAGMENT: Factorized Graph Representations for Document Generation and Editing via Entity-Aware Transformations

   Ayoub El Bouchtili, Guilhaume Leroy-Meline · arXiv:2608.18679

   Structured documents such as invoices, forms, reports, and scientific articles derive meaning from the interplay between spatial layout, textual content, and logical structure. Generative models operating at the pixel or token level often struggle to capture these dependencies…

3. H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning

   Jupiter Yao · arXiv:2610.26864

   AI summaryH2DF introduces a calibrated target‑band loss for language‑model unlearning, applying gradient ascent only below a data‑dependent lower bound and reversing it above the upper bound. It achieves forget/retain NLLs close to GradDiff with far less variability and minimal replay.

4. AutoTerm-SST: Adaptive Terminology Memory with Zero Session-Time Setup for Streaming Speech Translation

   Jiaxuan Luo, Siqi Ouyang, Lei Li · arXiv:2610.26865

   AI summaryAutoTerm‑SST eliminates pre‑session glossary upload by dynamically selecting topic‑specific terminology slices from a large offline catalog, matching streaming context to prototypes. It achieves 88.3% terminology accuracy and 45.09 BLEU, matching known‑domain baselines while running at 0.406 RTF.

5. High-dimensional layer-state trajectories reveal the internal dynamics of transformer inference

   Heng Zhao, Yufei Wang, Ruidian Song et al. · arXiv:2610.26866

   AI summaryThis paper analyzes transformer layer-state trajectories to show that inference unfolds as an ordered, measurable computation. The results demonstrate that specific hidden-state sequences and local transition actuators can selectively control internal model dynamics and output candidates.

6. WaveCommitments: Private and Verifiable Speech-Agent Marketplaces

   Oguzhan Baser, Kaan Kale, Sriram Vishwanath et al. · arXiv:2610.26867

   AI summaryWaveCommitments combines trusted execution and polynomial commitments to privately evaluate speech assets via Data Shapley, proving delivery matches the valuation while keeping data hidden. It improves benchmark speed by up to 20× and reduces proof size by 39×, revealing market prices across tasks.

7. Multimodal Ambivalence and Hesitancy Recognition with Bidirectional Audio--Text Cross-Attention

   TingTing Gu, JinCheng Luo, WenQiu Guo · arXiv:2610.26868

   AI summaryThis paper introduces an audio-text framework using bidirectional cross-attention to jointly model speech waveforms and transcripts for ambivalence and hesitancy recognition. The proposed method achieves a 77.95 percent F1-score on the Behavioural Ambivalence/Hesitancy dataset.

8. Mixture of Phonetic Experts Based Low-Rank Adaptation of Conformer Models for Accented English Speech Recognition

   Santosh Dahal, Anmol Guragain, Tianchi Liu et al. · arXiv:2610.26888

   AI summaryMoPE‑LoRA adapts Conformer ASR to accented English by routing frames to six phoneme‑category experts using phoneme supervision and acoustic gating, avoiding accent‑specific modules. On L2‑ARCTIC it outperforms single LoRA and full fine‑tuning, yielding a 12.3 % relative gain in zero‑shot accent transfer.

9. Disentangling Depression from Cognitive Decline in Elderly Speech Using Concurrent Clinical Assessments

   Woori Jeon, Seunghee Ha, Sang-Kyu Lee et al. · arXiv:2610.26890

   AI summaryThe authors analyze a new Korean elderly speech corpus to disentangle depression from cognitive decline in MCI patients. Controlling for cognition reveals that only formant features reliably indicate depression, outperforming standard acoustic baselines.

10. PW-Pruning: Phase-Aware Expert Selection for Speech-Centric Omni-Modal Models

    Junjie Cao, Fengping Tian, Wenjie Zhou et al. · arXiv:2610.26891

    AI summaryPW-Pruning calculates separate importance kernels for acoustic-input and text-generation phases to prune speech-centric omni-modal mixture-of-experts models. Retaining 32 of 128 experts per layer improves speech benchmark accuracy by over 3 percentage points compared to standard pooled selection without increasing inference-time overhead.

11. AtomSD: Procedural Atom-Guided Self-Distillation for LLM Agents

    Bohan Yu, Pengfei Cao, Kai Tian et al. · arXiv:2610.26926

    AI summaryProcedural Atoms encode Intent, When, Do, and Avoid guidance as reusable units. AtomSD extracts these atoms from successful trajectories and distills them into the policy, achieving up to 9.31 points better average performance than state‑of‑the‑art reward‑based methods.

12. Do We Need the Largest LLMs for Every Clinical Decision?

    Ali H. Lazem, William Teahan · arXiv:2610.26937

    AI summaryCERTA introduces a decision‑level certificate that lets 8‑bit quantized clinical language models replace BF16 references for 99.8–100 % of verified decisions while a 4‑bit frontier cuts deployment cost to ~26 % of BF16, with minimal F1 loss.

13. GRaSP: Latency-Aware Optimization for Geo-Distributed Retrieval-Augmented Generation

    Hardik Ruparel, Raj Kumar Pandey, Pramod Kumar · arXiv:2610.26938

    AI summaryGRaSP adds a latency‑aware re‑scoring layer to RAG that penalizes distant passages and optional costs, solving a matroid‑constrained optimization in O(N log N). It cuts P90 fetch latency by 41% and keeps 78 % of relevance recall with negligible overhead.

14. The Bottleneck Is Retrieval, Not Reading: Diagnosing Evidence-Based Generation with RAG for Long Context Question Answering

    Debopriyo Banerjee, Owais Aijaz, Ahmer Tabassum et al. · arXiv:2610.26951

    AI summaryThe paper proposes evidence‑based generation (EBG), requiring line‑level citations in retrieval‑augmented QA, and a joint Ans+Evid metric. It shows that similarity‑based retrieval fails to surface all needed chunks for multi‑fact queries and that chunking destroys document‑level context needed for identity‑bearing questions.

15. A Multi-granularity Interaction Network for Relational Triple Extraction

    Xiaocheng Luo, Yanping Chen, Zaifu Zhan · arXiv:2610.26952

    AI summaryMulti‑granularity Interaction Network captures word‑word, entity‑relation, and head‑tail dependencies to refine head/tail representations, boosting unseen‑entity relational triple extraction. It raises F1 by 3.1%/20.5% on NYT‑R/WebNLG‑R unseen sets compared to prior methods.

16. LAMB: Latent Multi-Hypothesis Beliefs for Long-Horizon Large Language Agents

    Peien Ding, Hui Li, Yuhan Liu et al. · arXiv:2610.26955

    AI summaryLAMB proposes a latent multi‑hypothesis belief module that separates a shared evidence state from weighted hypothesis deltas, updated by action‑outcome predictions without text memory. It outperforms prior belief models on ALFWorld, ScienceWorld, and Crafter, boosting success rates and reducing redundant actions.

17. Dynamic System Inspired Continuous Large Language Model for AI4Science Tasks

    Zhaoxin Fan, Zheng Zhu, Yuqing Ma et al. · arXiv:2610.26959

    AI summaryContinuum-LM reformulates pretrained residual transformations in language models as continuous dynamical systems governed by ordinary differential equations. By enabling variable computational depth through parameter reuse, it improves exact-match accuracy by up to 10.96 percentage points across eight scientific tasks.

18. Do LLMs Really Lack Difference Awareness? Uncovering Heuristic Substitution in Group-Differentiating Fairness

    Jingshen Zhang, Bo Wang, Yanlin Fu et al. · arXiv:2610.26961

    AI summaryThe paper shows that LLMs fail at group-differentiating fairness not from a lack of difference awareness, but by defaulting to a simple equal-treatment heuristic. It introduces a two-stage framework that decomposes this reasoning, significantly improving fairness decisions across multiple benchmarks.

19. SZED at ArGuard Shared Tasks: Multi-Task Label-Attention and Cross-Modal Fusion for Arabic Hateful Meme Detection

    Selma Khelili, Zohra Abdeli · arXiv:2610.26972

    AI summaryFusion of MARBERTv2 and CLIP ViT-L/14 embeddings yields 0.7597 macro‑F1 on Arabic hateful meme binary classification, while a multi‑task label‑attention network with cascade calibration achieves 0.3142 macro‑F1 on fine‑grained multi‑label tagging.

20. Transformers Learn Two Competing Sorting Algorithms, and Only One Length-Generalizes

    Khashayar Gatmiry, Nathan W. Henry, Chloe H. Su et al. · arXiv:2610.26974

    AI summaryTransformers trained on sorting converge to two distinct solutions: one uses only the first attention layer and fails to generalize to longer sequences; the other employs both layers, generalizing perfectly. Weight initialization can steer the model toward either class.

21. Sharp Degree Bounds for Putinar Certificates in Terms of the Positivity Margin via Analytic Sums of Squares

    Yuxiang Zhao, Ting Gan · arXiv:2610.26976

    AI summarySharp degree bounds for Putinar SOS certificates are proven: any fixed Archimedean quadratic module yields a uniform upper bound of \(O(\delta^{-1/2}\log^2(2/\delta))\) on certificate degree, matching a constructed lower bound, thus optimizing the dependence on the positivity margin \(\delta\).

22. Think Before You Transcribe: Context-Steered Reasoning for Hotword-Aware Speech Recognition

    Qihao Wang, Hang Su, Fan Cui et al. · arXiv:2610.26986

    AI summaryA context-steered reasoning framework first generates an audio‑grounded caption, then uses it to evaluate candidate hotwords before transcription. It improves transcription accuracy and hotword recall across varied contexts while resisting irrelevant candidate interference.

23. TRACE: Training-aware Robustness and Challenge Evaluation for Text-Vision Embedding Models

    Pume Tuchinda, Tinnakit Udsa, Peerat Limkonchotiwat · arXiv:2610.26993

    AI summaryTRACE reorganizes 168 text‑vision tasks into General, Adaptability, and Challenge categories, revealing that mixing general and adaptation data boosts Adaptability by ~7.5 points while keeping General stable, yet Challenge scores stay below 5, exposing major multilingual and visual‑document gaps.

24. How to get the best and fairest performance out of your ASR system using activation steering

    Felix Herron, Alex Colagrande, Solange Rossato et al. · arXiv:2610.26994

    AI summaryThis paper investigates activation steering to improve automatic speech recognition performance and demographic fairness. It finds that supervised optimization successfully derives steering vectors that enhance both overall accuracy and demographic fairness without requiring speaker labels.

25. Evaluating Tibetan Speech Recognition Across Dialects and Adaptation Regimes: A Multi-Granularity Analysis

    Jacob Moore, Sheng Li · arXiv:2610.27002

    AI summaryWe propose a unified evaluation combining character, syllable, and segmented‑word error rates for Tibetan ASR, and find that improvements at one level do not guarantee improvements at others. Cross‑dialect transfer and LoRA adaptation show weak metric correlations, highlighting the need for multi‑granular metrics and improved segmentation.

26. Decomposed Relation Extraction from Aviation Accident Narratives: Evaluating Transformer-Based Safety Knowledge Discovery under Relational Sparsity

    Aziida Nanyonga, Mustafa Tahsin Yilmaz, Saad Ezzini et al. · arXiv:2610.27013

    AI summaryEvaluates a two-step method that first identifies concepts then classifies relations in aviation accident reports. Despite high development scores, exhaustive evaluation shows drastic drop (Macro‑F1 ~0.27) with extreme false‑positive rates, indicating that concept overlap and sparse true relations severely limit current models.

27. Teaching AI Agent to Play Poker Strategically Through Preference Optimization

    Phurinat Polasa, Thanapon Thanadunpremdet, Phatthanasak Kraiduang et al. · arXiv:2610.27028

    AI summaryWe train a compact poker agent to act strategically under hidden information by optimizing preferences with KL‑anchored equilibrium (SPOKE). SPOKE reduces mean exploitability (Kuhn 0.213, Leduc 1.081) and improves six‑max agreement (0.775) while matching teacher KL.

28. Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models

    Osama Shabih, Dayam Nadeem, Shahab Saquib Sohail et al. · arXiv:2610.27034

    AI summaryThe paper evaluates hallucination and factual reliability across ten LLMs on 70 academic prompts of varying complexity, using five metrics. It finds Perplexity and DeepSeek V3 highest accuracy, Gemini best overall balance, and Falcon perfect precision but low recall.

29. Cross-Lingual Reusability of LLM Pipelines: Financial Extraction, Multilingual Reasoning, and Semantic Screening

    Joong Ho (Joshua), Choi, Sherine George et al. · arXiv:2610.27043

    AI summaryThe paper evaluates whether English‑built LLM pipelines can be reused across nine languages for financial extraction, multilingual QA, and semantic screening. It finds high accuracy transfer (≈95% extraction, 77–83% QA) and proportional reasoning‑effort gains, but stresses per‑domain, per‑language validation.

30. ReuF2F: Benchmarking Coding Agents on Formalized Undergraduate Research Problems

    Zeyu Zheng, Yiyang Liu, Yuyang Zhao et al. · arXiv:2610.27046

    AI summaryReuF2F presents a benchmark of 244 formalized open problems from REU projects, evaluated in Lean 4 with coding agents. Even the best model solves only about 40 % of the tasks, highlighting significant gaps in current formal problem‑solving abilities.

31. Discrete Neural Interlingua for Ontology-Agnostic EHR Modeling

    Tong Ding, Caiwei Tian, Dongmin Bang et al. · arXiv:2610.27047

    AI summaryThe paper introduces MEDIATOR, a discrete ontology-agnostic tokenizer that maps heterogeneous clinical concepts into a shared vector-quantized codebook using text descriptions and graph supervision. When used in EHR Transformers, MEDIATOR improves zero-shot transfer performance across health systems by 8.6 AUPRC points.

32. ClockMark: Adaptive and Host-Dependent Watermarking for Large Language Models

    Yupeng Qiu, Han Fang, Ee-Chien Chang · arXiv:2610.27079

    AI summaryClockMark embeds watermarks in large language model outputs by adapting to the evolving token distribution and the current generation state, ensuring distribution preservation. It achieves up to 16.08 pp higher bit accuracy than prior baselines while maintaining generation quality.

33. SciFigBench: A Benchmark for Scientific Method Figure Understanding

    Lijia Wu, Zheng Li, Yuxuan Liu et al. · arXiv:2610.27087

    AI summarySciFigBench provides 4,352 method figures and 40,670 QA pairs to test dense content, visual relations, and flow tracing in VLMs, showing current models lag in relation comprehension; a training‑free SciFigCue module improves results without parameter updates.

34. Omni-Woodpecker: Hallucination Correction via Evidence-Prior Rebalancing for Omni LLMs

    Zijian Liu, Chaoyou Fu, Yan Yang et al. · arXiv:2610.27103

    AI summaryOmni-Woodpecker is a training-free framework for omni large language models that corrects cross-modal hallucinations by orthogonally separating and rebalancing grounding evidence and learned priors in hidden states. It improves reasoning and reduces hallucinations across multiple audio-visual benchmarks without modifying model weights.

35. Beyond Two Tracks: Speaker-Anchored Full-Duplex Dialogue Recovery

    Yuxuan Chen, Yang Xiao, Jack Sun et al. · arXiv:2610.27109

    AI summarySpeaker‑anchored routing matches and orders separated voices before diffusion, reducing wrong‑track time from 43.9% to 4.1% and raising fixed‑track F1 from 0.46 to 0.84 while keeping reconstruction quality unchanged. The router transfers to unseen modalities, achieving <2% wrong‑track rates.

36. Joint Constraint Aggregation for Qualifier-Sensitive Retrieval-Augmented QA

    Jennifer Hsia, Zachary C. Lipton, Aarti Singh · arXiv:2610.27114

    AI summaryJoint Constraint Aggregation (JCA) compares candidates against the full set of question-derived requirements using retrieved evidence, reducing bias toward frequently mentioned partial matches. Across three datasets and instruction-tuned models, JCA improves accuracy, especially when retrieval favors competing answers.

37. ICLHerd: Herding-Based Data Selection for LLMs with In-Context Learning as a Proxy

    Yuqiao Wen, Lili Mou · arXiv:2610.27142

    AI summaryICLHerd balances mode-seeking and mode-averaging for LLM data selection using a von Mises-Fisher kernel and in-context learning as a proxy. Results show it outperforms existing data selection algorithms.

38. GRID: Modeling Health Influencer Discourse as Argument Networks for Credibility Assessment

    Nishanth Sridhar Nakshatri, Eylon Caplan, Dan Goldwasser · arXiv:2610.27149

    AI summaryThe paper introduces GRID, a framework that models health influencer content as interconnected argumentation networks to evaluate credibility through collective inference. Results show that incorporating broader discourse improves the accuracy of health-related content validation.

39. SITO: Sifting a Commercial Text-Safety API for Missed Harm and False Flags in English, Ukrainian and Surzhyk

    Andrii Tuzhykov · arXiv:2610.27178

    AI summarySITO evaluates the Tuteliq Detect API on 6 harm categories across English, Ukrainian, and mixed Ukrainian, showing 84.9–94.1% recall for harmful content and 0–4.3% false positives, but it often misses third‑party harm reports and has unstable decisions.

40. On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models

    Chenpan Li, Thies Brockmoeller, Stefanie Jegelka et al. · arXiv:2610.27190

    AI summaryLanguage models err more on composite prompts than on isolated tasks because they implicitly budget tokens, using fewer per task. Grouping tasks by fragility lets the budget be reallocated, cutting token usage by 43 % while keeping accuracy.

41. Preference-Aware Pareto Steering for Multi-Task Speech LLMs

    Liuyuan Jiang, Vishal Sunder, Samuel Thomas et al. · arXiv:2610.27200

    AI summaryEA‑SURF dynamically adjusts weighted‑loss scalarization during joint training of speech LLMs to steer the final task trade‑off toward a specified preference, achieving the target and boosting endpoint‑inclusive hypervolume by 17.8% over fixed weights.

42. The Culture Within: Towards Mechanistic Understanding of Cultural Reasoning in LLMs

    Samee Arif, Furkan Danisman, Angana Borah et al. · arXiv:2610.27202

    AI summaryCultureLens benchmarks LLMs on 9,261 cultural scenarios, showing that stating a contradictory norm drops alignment by 32‑47 % and that models default to Anglophone norms; fine‑tuning on contradictory‑norm examples boosts alignment to 85‑93 % across unseen countries.

43. LoWAgent: A Low-Resource Multilingual Benchmark for Web Agents

    Ocean Monjur, Daud Ahmad Nisar, Shahriar Kabir Nahin et al. · arXiv:2610.27205

    AI summaryLoWAgent evaluates web agents on English and four low‑resource languages, showing significant performance drops for all models, especially open‑source ones, and further degradation under interface disruptions. Small‑scale fine‑tuning on LoWAgent traces substantially recovers low‑resource performance.

44. Profile-DP: Differential Privacy for User-Controlled Profile Release in Personalized LLM Inference

    Mingchen Li, Zifan Peng, Kewei Sha et al. · arXiv:2610.27213

    AI summaryThe paper introduces Profile-DP, a policy-aware local differential privacy framework for releasing user profiles in personalized LLM inference. It uses finite-domain mechanisms to provably bound the release of withheld facts while preserving authorized personal information.

45. Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams

    Yanan Ma, Yang Wang, David Le et al. · arXiv:2610.27227

    AI summaryThe paper introduces GRSE, a group-relative credit assignment method that evaluates agent–skill pairs by contrasting multiple runs with the same skill library, reducing misattribution of harmful skills. It improves HotpotQA and LoCoMo performance, especially multi-hop questions, and robustly recovers from corrupted skill libraries.

46. MMedAgent-Plus: Scaling Multimodal Medical Agents across Modalities, Tasks and Workflows

    Ruoxin Wang, Yuanting Pan, Binxu Li et al. · arXiv:2610.27233

    AI summaryMMedAgent-Plus integrates a vision-language planner with 18 specialist models to coordinate complex medical workflows across multiple modalities and tasks. It achieves high tool-call and final-answer accuracies, significantly outperforming existing baselines in end-to-end medical execution.

47. CEFR-oriented Russian Text Simplification Using Reinforcement Learning

    Kristina Mustakova, Alexey Sorokin · arXiv:2610.27237

    AI summaryThis paper introduces a preference-optimized YandexGPT model for Russian text simplification across specific CEFR proficiency levels. The best model achieves 87.15 percent target level accuracy while preserving text meaning and fluency.

48. Predicting Linear Relational Decodability Without the Relation

    Jonas Rohweder, Adrian Mülthaler, Farid Mehri et al. · arXiv:2610.27241

    Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for person_father to the ceiling for country_capital_city. We investigate the factors that…

49. Diagnostic Analysis for Large Language Model Reasoning

    Xinyu Liu, Dongqi Fu, Jun Wu · arXiv:2610.27245

    Reasoning structures (e.g., chains and graphs) are increasingly used for probabilistic inference, factuality verification, and uncertainty estimation, but their structural reliability is rarely evaluated directly. We test whether this reliability holds through a proposed…

50. Target-Aware Curation of Hallucination Stress Tests for Large Language Models

    Jiho Lee, Bingquan Zhang, Lin Ting Liang et al. · arXiv:2610.27247

    Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM…

51. TxnQA: Benchmarking Analytical Reasoning and Evidence Grounding over Transaction Event Sequences

    Rabimba Karanjai, Yiming Bian, Yun-Shiuan Chuang et al. · arXiv:2610.27256

    Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark…

52. Auditing Safety Preference Supervision with Recorded Derivations

    Chenhan Lyu, Amir Rahmani · arXiv:2610.27266

    AI summaryThe paper introduces Symbolic Preference Labeling, which records attribute classifications, rule-based verdicts, and a ranking to recompute preference labels without re‑running a language model, enabling auditability. On 200 prompts it achieves 96% structured output, with policy models refusing 40% benign and 85% toxic prompts.

53. African Languages AI Safety Readiness Report

    Israel Abebe Azime, Jesujoba O. Alabi, Tajuddeen Gwadabe et al. · arXiv:2610.27269

    AI summaryThis report evaluates the AI safety readiness of African languages across data availability, policy support, benchmarks, and speaker populations, grouping them into four readiness tiers. The classification aims to guide strategic investments and inclusive benchmark development for underrepresented languages.

54. FLINT: Fixed-budget Late Interaction

    Himank Handa, Negar Arabzadeh, Matei Zaharia · arXiv:2610.27273

    AI summaryFLINT introduces a fixed-budget late interaction retrieval architecture that applies symmetrical attention-based pooling to both queries and documents. It reduces index size by up to 9x and accelerates retrieval compared to token-level models while maintaining competitive ranking effectiveness.

55. When Does Hate Transfer? A Large-Scale Study of Cross-Lingual Hate Speech Detection

    Zahra Bokaei, Walid Magdy, Bonnie Webber · arXiv:2610.27285

    AI summaryCross‑lingual hate‑speech detection works best when languages are linguistically similar; transfer quality also depends on temporal overlap and dataset size, but shared cultural cues explain residual patterns. The study offers guidelines for applying cross‑lingual models to low‑resource languages.

56. TEX: Test-Time Scaling Testing Agents via Execution-based Cross-Validation

    Akash Gokul, Srijan Bansal, Yingbo Zhou et al. · arXiv:2610.27290

    AI summaryTEX parallelizes agents to generate test cases and patches, cross‑executing each pair to build a $K\times K$ feedback matrix that guides refinement without trace sharing. It attains 86.4% on SWT‑Bench Verified, a +9.0% boost over single‑agent baselines.

57. RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

    Xinying Cai, Bin Hu, Mingyu Jin et al. · arXiv:2610.27295

    Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks…

58. Evaluating Context Compaction on the Questions That Come Later

    Junwei Yao, Yongxi Zhou, Zike Yang et al. · arXiv:2610.27301

    AI summaryThe paper tests whether context‑compaction methods that preserve evidence for the current question also help future questions. It finds that sentence‑level compaction improves only the current answer, while word‑level deletion (LLMLingua‑2) better preserves useful content for later MultiRC questions, but its benefit varies across datasets.

59. Global Security Blind Spots: Translation as a Safety Bypass in Low-Resource Languages

    Jalil Alva, David Harper, Abigail E. Thornton et al. · arXiv:2610.27304

    AI summaryThey test ten large language models in 79 languages for translation‑bypass attacks, finding that safe English prompts still lead to unsafe outputs 22–44% of the time, especially in certain language families, and that model rankings are unstable across conditions.

60. On the LLM Verification for Argument Graph Generation

    Jakub B\k{a}ba, Jaros{\l}aw A. Chudziak · arXiv:2610.27361

    Multi-agent LLM systems can formalize natural-language argumentation end-to-end, turning plain text into structured graphs by classifying components, connecting them with relations, and enriching the resulting graphs into formal models. Quality assurance for those graphs is…

61. Explore More, Validate Less: Efficient Reflective Prompt Optimization

    Xiaofan Zhou, Lu Cheng · arXiv:2610.27367

    Reflective prompt optimization improves instructions using task feedback, but coupling candidate generation with full-validation-based selection can make search costly and sequential. We propose a two-stage framework that decouples reflective candidate-pool construction from…

62. SHAM-Bench: A Span-level Hallucination Benchmark Across Multiple Tasks

    Sicong Huang, Ian Lane · arXiv:2610.27370

    AI summarySHAM-Bench aggregates 6,461 span‑level hallucination annotations across six NLG tasks into a unified schema, enabling cross‑task evaluation. Fine‑tuned detectors excel in‑domain but drop sharply on shift, while reasoning‑prompted methods show better robustness, yet subtle unsupported spans remain hard to detect.

63. DStill: Trajectory Consistency Distillation for Accelerated One-Step Drafting in Block Diffusion Speculative Decoding

    Shreyas Kulkarni, Uttam Gurram, Madhav Kanda et al. · arXiv:2610.27379

    Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without…

64. Steering intonational tunes in text-to-speech synthesis

    Juan Pablo Vigneaux, Jennifer Cole · arXiv:2610.27403

    AI summaryThis paper investigates how text-to-speech models encode prosodic intonation and shows that steering the transformer backbone's hidden states with imitation-derived vectors successfully controls sentence-level rising and falling melodies with minimal loss in naturalness.

65. More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents

    Ruhan Wang, Kishan Panaganti, Dongruo Zhou · arXiv:2610.27406

    AI summaryLLM agents can self‑improve by storing past runs, but reusing inferred abstractions propagates errors, a process called error reproduction. The Anti‑Persistence design discards synthesized guidance after use, only keeping raw trajectories, and outperforms other memories on success rate and LLM‑call cost.

66. Looking Into The Past: The Impact of Context on Detecting Deceptive Statements

    Alex Clay, Ernesto Jimenez-Ruiz, Pranava Madhyastha · arXiv:2610.27407

    AI summaryThe paper shows that adding preceding gameplay dialogue improves deceptive statement detection in Avalon. Using LLMs with context raises accuracy by up to 31 % compared to isolated message analysis.

67. The Router Already Knows

    Andrew Bistras, Marty Johnson · arXiv:2610.27421

    Mixture-of-experts (MoE) models already make a sparse choice for every token: which experts should process it. We ask whether a sparse dictionary should reuse that choice. NR-SAE gives each routed expert a small dictionary and opens only those the frozen router selects. At block…

68. Harness Self-Evolution for Long-Horizon Terminal Agents

    Ruhan Wang, Yucheng Shi, Zongxia Li et al. · arXiv:2610.27435

    AI summaryRiseHarness iteratively refines terminal-agent harnesses by contrasting successful and failed runs, using runtime traces to diagnose and patch shared mechanisms. Experiments show that integrating cross-framework fixes into a single harness boosts reliability and task success across models and unseen benchmarks.

69. Ask to Distill: Learning to Synthesize Data for On-Policy Distillation

    Jiashuo Sun, Hao Zhang, Xueqiang Xu et al. · arXiv:2610.27450

    On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowledge transfer. How can we assess whether a newly generated question is valuable for distillation, and learn to generate better…

70. Detecting Errors in Large Language Model Responses: The Incremental Predictive Value of Hidden States

    Ningzhi Li, Xiaojun Ma, Xiaofei Wang · arXiv:2610.27455

    Although large language models support a wide range of tasks, hallucinations remain a major obstacle to user trust. When a model generates an answer, both its internal processing and final output vary with the question. Internal states are often expected to offer richer signals…

71. Recoverability and Verifier Strength: Two Measured Preconditions for Unattended Agent Self-Improvement

    De Wang, Peng Qi · arXiv:2610.27465

    AI summaryDefines two metrics—verifier strength and recoverability—to decide when an automated agent can self‑modify without human oversight. Experiments show only changes with stored previous versions and a measured strong verifier can be safely deployed unattended.

72. Self-Evolving Agents through Privileged Retrospective Distillation

    Yuhang Zhou, Yifan Wu, Bo Peng et al. · arXiv:2610.27495

    Self-evolving agents aim to improve from their own experience rather than remain fixed after deployment, but naturally collected trajectories often lack ground-truth solutions or verifier-derived rewards. We introduce Retrospective Self-Distillation (RSD), a multi-round…

73. Judging the Justification: Decision Certificate Supervision for Generative Reward Models

    Guiquan Sun, Bing Bai, Martin Renqiang Min et al. · arXiv:2610.27528

    Generative reward models (GenRMs) generate natural language rationales to justify preference judgments, yet a correct verdict does not guarantee a sound justification. Existing approaches often rely on reference explanations to supervise rationales, but obtaining such references…

74. EviDial: When Correct Depression Predictions Lack Evidence

    Junjie Xu, Jin Zhu, Yuzhe Yang et al. · arXiv:2610.27541

    AI summaryThe authors introduce EviDial, a benchmark separating label agreement from evidence in depression dialogues. They show that most predictions deemed correct by standard clinical labels actually lack sufficient supporting dialogue evidence.

75. Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation

    Bingqi Shang, Yang Zhang, Yiwei Chen et al. · arXiv:2610.27555

    Safety pretraining has emerged as a promising paradigm for embedding safety deeply into a model during pretraining, yielding robustness to subsequent fine-tuning that conventional safety post-training often lacks. We refer to this safety acquired during model pretraining as…

76. Exploring Recurrent-Depth Language Models for Multitask Clinical Text Processing

    Xi Yang, Guodong Liu, Cheng Peng et al. · arXiv:2610.27559

    Clinical natural language processing includes diverse tasks such as information extraction, classification, and question answering. Developing one model for all tasks could simplify deployment and improve knowledge transfer. We evaluated whether small recurrent-depth language…

77. AnomalyJev: Towards Evaluating Image Anomaly Explanations via Claim-Level Verification

    Qizhou Wang, Jiawen Zhu, Guansong Pang et al. · arXiv:2610.27565

    Recent multimodal LLM (MLLM)-based anomaly detection (AD) methods can generate textual explanations that help interpret their predictions. However, reliably and efficiently evaluating these explanations remains challenging. While LLM judges can assess explanations through…

78. Learned Readout Tokens for Extracting Large Language Model Confidence

    Nima Seifi, Martin Ester · arXiv:2610.27566

    AI summaryLearned readout tokens replace existing-token probes for LLM confidence estimation, yielding higher AUROC across datasets and models. A multi‑token extension further improves performance, and the method can be applied to any probe, boosting accuracy both in‑domain and out‑of‑domain.

79. AudiPsy: Evidence-Grounded Multimodal Reasoning over Spoken Counseling Conversations for Depression Assessment

    Chuanbo Zhu, Run Luo, Weibo Zhao et al. · arXiv:2610.27567

    AI summaryAudiPsy presents an evidence‑grounded multimodal reasoning framework for depression assessment in spoken counseling, integrating textual and acoustic emotions via an Observation‑Assessment‑Summary paradigm. On DAIC‑WOZ it outperforms specialized detectors and general LLMs while providing higher‑quality reasoning.

80. When LLMs Manipulate: Measuring Behavioral, Social, and Temporal Effects

    Sadia Asif, Israel Antonio Rosales Laguan, Shumaila Asif et al. · arXiv:2610.27573

    Large language models (LLMs) can influence user trust and decision-making through subtle linguistic cues, yet the behavioral effects of manipulative outputs remain poorly quantified. We introduce a behavioral computing framework for measuring \textit{behavioral vulnerability},…

81. Why Static Probes Fail: Context Determines Deception Feature Composition

    Yao Huang, Siyuan Chen, Zhenghan Zhou et al. · arXiv:2610.27577

    With increasing autonomy, Large Language Models (LLMs) may engage in deceptive behaviors, such as scheming to pursue misaligned goals or sandbagging to escape oversight. Monitoring outputs alone is insufficient, as models can produce seemingly benign responses while harboring…

82. Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval

    Yuchen Guo, Zhuoyuan Mao, Yiwen Guo · arXiv:2610.27580

    Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR)…

83. TeCoP: Accelerating Diffusion Large Language Model Decoding via Tentative Context Pipelining

    Jui-Chien Tsou, Hsuan-Chi Liu, Wei-Chin Wang et al. · arXiv:2610.27583

    Diffusion large language models (dLLMs) predict multiple token positions per forward pass, yet block-wise decoders serialize generation and repeatedly recompute stable predictions. We attribute this inefficiency to coupling a prediction's readiness for irreversible commitment…

84. Beyond Model Weights: Sampling-Dependent Fairness in Diffusion Language Models

    Koshiro Saito, Diasuke Oba, Danushka Bollegala et al. · arXiv:2610.27584

    Diffusion language models (DLMs) are emerging as a practical alternative to autoregressive language models (ARLMs), yet their fairness remains largely uncharacterized. DLMs introduce sampling controls with no direct autoregressive counterpart, such as block length and remasking…

85. Benefit and Cost of Random-Token Replacement under Repeated Data: A Known-Entropy Decomposition and Evidence from Ten Real Corpora

    Wonsup Shin, Jingu Kim · arXiv:2610.27590

    When unique pretraining data is scarce, models see the same tokens many times and are thought to memorize. Random-token replacement, an input-corruption augmentation, is one remedy reported to lower validation loss, but a lower loss alone does not reveal which source of error it…

86. WORTH Benchmark: Evidence Discovery under Not-Fully-Cooperative Information -- A Two-Episode Pilot

    Bo Zou, Chao Xu · arXiv:2610.27597

    AI summaryThe paper introduces the WORTH Benchmark, where models read full mystery stories, autonomously pick key clue sentences, identify participants, and explain reasoning. In a two‑episode pilot, models frequently selected incorrect participants and missed key clues, revealing misaligned evidence selection.

87. Self-Decomposing Diffusion for Code Generation

    Lize Shao, Zichen Xie, Zhizhen Chen et al. · arXiv:2610.27600

    Diffusion language models (DLMs) are emerging as a promising paradigm for code generation, offering the potential for fast, parallel decoding. However, generating complex programs remains challenging, as their interdependent components must maintain consistent interfaces, data…

88. Auditing the Product, Not the Recursion: The Evaluation Evidence Behind Recursive Self-Improvement Claims

    Zihan Chen, Di Zhu, Lei Nico Zheng · arXiv:2610.27610

    AI summaryThe paper audits 114 studies claiming recursive self‑improvement, finding that while 90% report product‑level evaluation, only 22% assess repeatability with full loops, and only 4 compare identical procedures under matched budgets. It proposes a claim‑indexed reporting standard.

89. ABench: Adversarily Benchmarking Large Language Models across Taxonomized Values

    Liping Yang, Yongqi Leng, Jingting Zheng et al. · arXiv:2610.27611

    Ensuring the alignment of Large Language Models (LLMs) with diverse cultural values is critical for their safe global deployment. However, existing alignment benchmarks are predominantly English-centric or confined to simple discriminative tasks. Moreover, they often fail to…

90. PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking

    Pan Wang, Yihao Hu, Guangyuan Dong et al. · arXiv:2610.27628

    Looped language models can run more iterations at test time, but additional iterations need not help them solve longer problems. For sequential state tracking, final-answer supervision does not specify how far each iteration should advance, and repeated updates can change the…

91. A Two-Way Street: Surveying the Interplay between Code-Mixing and Native Languages in Text and Speech

    Debajyoti Mazumder, Akash Anil · arXiv:2610.27631

    Systems for code-mixed (CM) text and speech rarely start from scratch. CM data is scarce, so most work borrows from the languages being mixed, which we call native languages. This transfer can also run in reverse, when CM data is used to improve models for the native languages.…

92. BCT: Bayes-Consistent Tempering for Flow Language Models

    Daisuke Oba, Hiroki Furuta, Naoaki Okazaki · arXiv:2610.27641

    AI summaryBayes-Consistent Tempering (BCT) integrates temperature into the flow distribution and denoiser dynamics, aligning training and inference. BCT boosts GSM8K and Sudoku performance across temperatures, outperforming post‑hoc scaling and training‑free variants on hyperspherical and von Mises–Fisher flows.

93. CUA-Recovery: Benchmarking Error Recovery in Long-Horizon Computer-Use Agents

    Xu Hu, Haoming Li, Yuxiang Wang et al. · arXiv:2610.27645

    AI summaryCUA-Recovery benchmarks error detection and recovery for long-horizon desktop‑agent tasks, providing 2,250 verified failure states and 8.6K recovery trajectories. Fine‑tuning Qwen3.5 on this data boosts error awareness from 20 % to 35 % and task success scores markedly.

94. Speak vs. Know: Balancing Behavior and Knowledge Distributions of Large Language Models

    Yang Liu, Masahiro Kaneko, Chenhui Chu · arXiv:2610.27648

    AI summaryThis paper identifies a significant mismatch between what LLMs verbalize as group opinions and what they actually output via next-token probabilities. To fix this, the authors propose a learnable method that balances behavioral consistency with distributional quality.

95. Corpus Quality Improves, Retrieval Utility Does Not: An Ablation of Preprocessing and Refinement in a Speech-to-QA Pipeline

    Wonsup Shin, Jingu Kim · arXiv:2610.27652

    AI summaryThe study shows that preprocessing and refinement improve corpus quality, but only transcript rewriting harms downstream retrieval accuracy. The improved corpus raises retrieval scores from 0.231 to 0.748 across models, yet quality metrics do not predict this benefit.

96. CovNav: Conformal Prediction-Guided Adaptive Inference for Object Goal Navigation

    Pengying Wu, Zhejia Wen, Kang Ding et al. · arXiv:2610.27677

    In object goal navigation (ObjectNav), an embodied agent must find a target object (e.g., "find a bed") in a previously unseen environment, relying on spatial commonsense such as object co-occurrence. Large Language Models (LLMs) encode this commonsense, but deploying them on…

97. PhonoBridge: Error-Aware Phoneme-Mediated EMG-to-Text Reconstruction

    Wen Qiu, Jiacheng Zhang, Xuanda Chen et al. · arXiv:2610.27678

    AI summaryPhonoBridge decodes silent speech by mapping EMG signals to phonemes with a channel-aware frontend, then reconstructs text using an LLM trained on synthetic phoneme–text pairs and error-pattern augmentation. It reduces WER on a silent-speech dataset to 24.77%, better than prior 28.27%.

98. Memory as Code: Declarative Retrieval over Codified Long-Term Memory

    Aoting Zeng, Yuxin Du, Jianwei Wang et al. · arXiv:2610.27700

    AI summaryMemory as Code (MaC) creates executable memory abstractions from interaction histories and uses declarative query programs for retrieval. It improves LLM-judge accuracy by up to 15.9% on LoCoMo and LongMemEval while cutting token usage.

99. Beyond Calibration: Temporal LLM Confidence Assessment Across Languages and Local Settings

    Sahil Kale · arXiv:2610.27702

    AI summaryThe paper separates pre‑solve (Feeling‑of‑Knowing) and post‑solve (Judgement‑of‑Learning) confidence signals in large language models, showing that post‑solve confidence is better calibrated yet more expensive. Across English and Marathi, translation preserves calibration, but local Marathi knowledge degrades it, and hallucinations keep high confidence.

100. When Does Experience Memory Help Closed-Loop LLM Driving Agents? A Controlled Evaluation with Resampled Warmup Pools, Token-Matched Fillers, and Audited Baselines

     Yingxin Jia, Xuanyi Li · arXiv:2610.27714

     AI summaryThe study evaluates closed‑loop LLM driving agents with controlled warm‑up experience pools and token‑matched fillers on a HighwayEnv testbed, finding that memory gains depend on the specific warm‑up pool, token fillers dominate, representational depth is ineffective, and baseline quality reverses apparent advantages.

101. Tracing Recurrent and Emergent Hallucination Patterns in Decomposed Text-to-SQL Generation

     Hayden Moore, Abhinav Verma, Anshuman Dutt · arXiv:2610.27720

     AI summaryThe paper introduces a drill‑down framework that decomposes Text‑to‑SQL generation into intermediate sub‑queries and queries, enabling analysis of hallucinations at each step. It finds that some errors recur across steps, while others only emerge at the final stage.

102. Hearsay or Fact? How Framing and Elicitation Shape LLM Probability Updates

     An-Sheng Jhang · arXiv:2610.27745

     AI summaryThis paper investigates how prompt framing and elicitation formats affect how much large language models update their probability estimates in response to new information. The results show that indirect reporting slightly reduces update magnitudes, while eliciting multiple weighted estimates reduces extreme probabilities but can impair evidence sensitivity.

103. Give the Model a Finger: Developmentally Aligned Evaluation of Counting in Vision-Language Models

     Yingxin Jia, Hui Geng, Wenxin Yang et al. · arXiv:2610.27756

     Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master counting. Children cross the barrier with epistemic actions: pointing at each item while reciting number words, moving counted…

104. Are Linguistic and Arithmetic Circuits Dissociated in LLMs?

     Riku Kisako, Tatsuki Kuribayashi, Ryohei Sasano · arXiv:2610.27790

     AI summaryThe study extracts and compares circuits for linguistic versus arithmetic tasks in five LLMs, finding that similarity aligns mainly with task domain rather than shared vocabularies or syntax. Thus, linguistic and arithmetic processing in LLMs uses largely separate, modular neural circuits.

105. ProVe: Proposer-Verifier Decoding

     Ziqi Jin, Weiwen Xu, Jiaxi Li et al. · arXiv:2610.27815

     AI summaryProVe fine‑tunes a large model to propose a block of tokens in one pass and verifies it in a second, swapping the roles of draft and verifier from speculative decoding. ProVe‑lossless achieves 2.5× AR throughput, while a 3B verifier gives 2.7× speed with only 2.6 accuracy loss.

106. SMART: Semantic Segment Mapping and Assessment of Recognition Transcripts with Contextual Rubrics

     Xiaoyu Gu, Chencheng Wang, Fuyang Li et al. · arXiv:2610.27820

     AI summarySMART maps ASR errors into semantic segments and applies contextual rubrics to assign consequence labels, achieving 0.843 F1 against human labels and enabling the new Semantic Impact Error Rate metric with significant improvements over baseline corrections.

107. SciDiagBench: A Type-Specific Benchmark for Human-Aligned Evaluation of Scientific Diagrams

     Hao Wang, Zixuan Li, Long Bai et al. · arXiv:2610.27841

     AI summaryThe paper introduces SciDiagBench, a benchmark and evaluation method that categorizes scientific diagrams into eleven types to align automated quality assessment with human expert preferences. The result is a more reliable evaluation framework that overcomes the limitations of generic, heuristic-based criteria.

108. UvA-MT's Participation in the WMT26 General Translation Shared Task

     Di Wu, Sergey Troshin, Wafaa Mohammed et al. · arXiv:2610.27852

     AI summaryUvA‑MT uses LLM‑driven test‑time methods for WMT26: sequential sampling with refinement prompts produces multiple candidates, and an LLM pairwise judge selects the best via round‑robin voting. This outperforms traditional best‑of‑N selection on WMT25.

109. From Tokens to Files: Multi-Granularity Detection and Localization of AI-Involved Code

     Yue Xu, Xiyang Wang, Hui Liu et al. · arXiv:2610.27863

     AI summaryThe paper introduces a token‑level authorship detector that aggregates evidence to classify lines, blocks, and entire files as AI‑involved, achieving 72.4 macro‑F1 at 1% FPR on 11,773 mixed files and 62.5 line‑level F1 on an external benchmark.

110. Testing Associative Creativity in Humans and Multimodal LLMs with Difficult-to-Name Visual Stimuli

     Manar Ali, Lisa Gottschalk, Simeon Junker et al. · arXiv:2610.27894

     AI summaryThis paper evaluates associative creativity in humans and multimodal language models using difficult-to-name visual stimuli with varying shape complexity and noise. Results show that humans produce significantly more diverse, novel, and semantically distant descriptions than the tested models.

111. CollabSciReM: Aligning Language Models for Scientific Conversations via Science-Aware Reward Decomposition

     Sasi Kiran Gaddipati, Gollam Rabby, Sören Auer · arXiv:2610.27901

     AI summaryCollabSciReM trains LLMs with a reward decomposed into conversational quality, scientific rigor, and efficiency, showing a 65.3% win rate on SciQuality and higher novelty with fewer hallucinations versus science‑agnostic rewards.

112. Staged Continual Adaptation of Multimodal Foundation Models for Japanese Financial Documents

     Genshin Kakimoto, Atsushi Yanagisawa · arXiv:2610.27904

     AI summaryThe paper analyzes how staged post‑training (vision‑language alignment, reasoning distillation, domain tuning) affects an 8.4B Japanese financial‑document model. It shows the phases interact non‑additively; the best checkpoint depends on the downstream task.

113. Polarity Blindness in LLM Safety: A Systematic Study of Negation, Antonym Substitution, and Adversarial Framing

     Diego Miguel Lozano, Erik Derner, Juan Antonio Pérez-Ortiz et al. · arXiv:2610.27908

     AI summaryThis paper reveals that state-of-the-art language models suffer from polarity blindness, failing to recognize harmful intent when prompts use grammatical negation or antonym substitution. Consequently, simple linguistic transformations consistently bypass safety guardrails across multiple models and languages.

114. ACE-Compression: An Ability-Centric Evaluation of Compressed Large Language Models

     Jiahui Ning, Liang Chen, Yuchen Liang et al. · arXiv:2610.27913

     AI summaryThe authors introduce ACE-Compression, an evaluation framework that assesses compressed large language models across four core ability dimensions rather than relying solely on overall task performance. They find that compression selectively degrades fine-grained and context-dependent abilities while preserving surface-level signals, with quantization generally outperforming pruning.

115. Why Do LLMs Accept Generics? Explaining Acceptability using Psycholinguistic and Model-Derived Factors

     Mervat Abassy, Zhaoxin Feng, Tatsuki Kuribayashi et al. · arXiv:2610.27928

     AI summaryThe paper compares human judgments of generic statements with those of twelve LLMs, finding that models balance property commonness and distinctiveness like people but mainly rely on factual truth, showing weaker sensitivity to strikingness and characteristic factors.

116. From DQI to C-DQI: Automatic Codebook Generation and Evaluation for Measuring Computational Discourse Quality Index

     Esther Shizgal, Ron Keinan, Daniel Shtauber et al. · arXiv:2610.27937

     AI summaryWe present an LLM-driven protocol that automatically generates and evaluates context‑adapted codebooks for parliamentary discourse quality. Data‑driven schemas achieve higher inter‑model agreement than theory‑based ones while remaining valid, enabling reliable large‑scale political discourse analysis.

117. Lost in Recovery: Demystifying Recovery Behavior of Computer-Use Agents via Goal-Oriented Action Replay

     Byeonghun Bang, Sangwoo Park, Jinheon Baek et al. · arXiv:2610.27954

     AI summaryReTrace is a framework that evaluates model‑based recovery in computer‑use agents by injecting controlled errors and replaying goal‑oriented actions, defining Transition‑Consistent Recovery. It shows ideal recovery unlocks hidden task completion, current agents underutilize this, and accurate recovery improves end‑to‑end performance.

118. Adaptive Stopping of Streamed Reasoning for Language Model Agents

     JunHyeok Oh, Byung-Jun Lee · arXiv:2610.27981

     AI summaryStreamStop learns lightweight, model-specific stopping policies for reasoning language models using offline action agreement. It reduces reasoning tokens by 42% on WebShop while maintaining task success, without requiring access to internal model states.

119. Building Telecom Language Models with RSOD: Domain Data, Specialist Post-Training, Memory-Conscious Distillation, and Networking Evaluation

     Anil Srirangapatna Nagesh, Nitin Gupta, Anuj Taneja et al. · arXiv:2610.27993

     AI summaryThis paper introduces Routed Specialist On-Policy Distillation (RSOD), a framework that compresses multiple domain-specific telecom teacher models into a single deployable student adapter. The resulting model nearly matches the performance of the multi-teacher ensemble while remaining practical for standard deployment.

120. Beyond the Evidence Boundary: Behavioral Mirage Under Progressive Degradation in Medical VLMs

     Edric Castel C. Hao, Rez Samantha Z. Floresca · arXiv:2610.28000

     AI summaryMedical VLMs often keep giving confident answers even as image quality degrades; the study defines an evidence boundary where models cease to be reliable and shows that MedGemma’s post‑boundary behavior varies from correct recovery to persistent mirage, highlighting the need for boundary‑aware evaluation and better abstention.

121. From RL Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks

     Taiheng Pan · arXiv:2610.28027

     AI summaryResearchers tracked 6,939 RL papers linked to seven open‑source frameworks and found that only 37 of 1,598 general RL methods were incorporated. Included papers tend to target large (≥30B) models and modify loss components, while most proposals remain unaddressed publicly.

122. Using Letter Positional Probabilities to Assess French Word Complexity

     Michael Dalvean · arXiv:2610.28040

     AI summaryThe study uses Letter Positional Probabilities to differentiate 5,000 high‑complexity from 5,000 low‑complexity French words, finding many letter/position patterns linked to complexity. Adding binary letter information for the first six positions boosts classification from 77.27% to 83.95%.

123. Poisoning-Based Backdoors Are Not Cryptographic

     Masahiro Kaneko · arXiv:2610.28045

     AI summaryThe authors investigate whether poisoning-based neural network backdoors can achieve cryptographic asymmetry. They prove that any backdoor installed via data poisoning is inherently learnable, meaning defenders can always reconstruct the trigger verifier given observed triggers.

124. On-Policy Transition Distillation

     Runyang You, Yongqi Li, Tiezheng Yu et al. · arXiv:2610.28048

     AI summaryOn-Policy Transition Distillation augments standard final-layer matching with alignment of layer-to-layer log-probability changes on student rollouts. Across multiple backbones and benchmarks, it recovers ~90% of the teacher–student capability gap, outperforming OPD.

125. Same Judgment, Different Yes: Measuring Multilingual Judgment Profiles in Language Models

     Haonan Huang, Yuwei Lu, Chengkun Cai · arXiv:2610.28066

     AI summaryMultilingual judgments are assessed across 1.1M replies to 105 moral vignettes, varying question form, answer constraint, and token vs. sentence readouts. Findings show that profile consistency drops only in negative interrogatives; models retain consistent judgments across languages, but yes-token usage varies by language.

126. Contextual Candour: Role-Aware Graduated Disclosure for LLMs over Mixed-Sensitivity Contexts

     Devansh Srivastav, David Pape, Samira Abedini et al. · arXiv:2610.28067

     AI summaryGraduated disclosure calibrates LLM outputs to a user’s authorization level, optimizing responses to provide only appropriately abstracted information while penalizing unauthorized leakage, including paraphrases. Experiments on enterprise and clinical data show significantly less leakage with maintained correctness and utility.

127. BAIT: Can LLMs Find What Biases a Reader? A Benchmark for Cognitive Bias Triggers in News

     Yousra Fettach, Tijl De Bie · arXiv:2610.28092

     AI summaryThe paper introduces BAIT, a benchmark for detecting and localizing cognitive bias triggers in news articles, and evaluates six frontier LLMs on these tasks. The results show that while LLMs can identify overall triggers at near-human levels, they struggle to locate the specific passages, often relying on surface cues rather than manipulative mechanisms.

128. Reading Too Much Into It? Interpretive Preferences and Commitment in LLM Mentalisation under Ambiguity

     Stephanie Fong, Yiwen Jiang, Qingyang Xu et al. · arXiv:2610.28105

     AI summaryThe authors introduce MentalLoom, an evaluation framework for LLM mentalisation during ambiguous social episodes. They find that user distress disclosures and repeated questioning cause models to increase their emotional support, certainty, and positive interpretations without any new evidence.

129. Occupational Uniqueness and Wage Growth

     R. Vestrelli, A. Fronzetti Colladon · arXiv:2610.28133

     Changes in occupational task content may have different economic implications depending on whether the tasks gained or lost are occupation-specific or already widespread across the economy. Existing measures of occupational task change largely abstract from this distinction,…

130. Africa-Papers: A 10-Billion-Token Multi-Domain Academic Corpus for African Foundation Language Models

     Mich-Seth Owusu · arXiv:2610.28173

     AI summaryThe authors introduce Africa-Papers, a 10.15-billion-token multi-domain corpus of African academic literature designed to address the severe underrepresentation of African scholarship in pre-training data. The resulting dataset significantly improves the foundational training and domain-specific reasoning capabilities of language models for African contexts.

131. Temporal Evidence-Parameter Co-Alignment for LLM Reasoning over Temporal Knowledge Graphs

     Jiapu Wang · arXiv:2610.28176

     AI summaryTEPCA aligns LLM reasoning with temporal knowledge graphs by pruning evidence paths with time‑constrained DRGP and MI‑based pruning, and adapting parameters with timestamp‑conditioned DG‑LoRA. The method achieves state‑of‑the‑art performance on TKG benchmarks and improves commonsense QA.

132. Uncovering Audio-Text Alignment in Large Audio-Language Models through Cross-Modal Concept Neurons

     Tzu-Quan Lin, Chih-Kai Yang, Yi-Cheng Lin et al. · arXiv:2610.28189

     While the language model backbone is pretrained on text alone, large audio-language models (LALMs) can, to a great extent, recognize acoustic attributes, such as emotion, speaker gender, and spoken language. However, it is unclear how the audio representation is integrated, if…

133. Can Language Models Get "Aphasia" From Neuronal Ablation?

     Pratik Worah, Randall Bateman · arXiv:2610.28214

     AI summaryThey systematically ablate whole units, MLPs, and attention heads in LSTM and Transformer language models, fine‑tuning to mimic plasticity. Transformers collapse abruptly under MLP ablation (semantic paraphasias) and smoothly under head ablation (semantic loss), while LSTMs decline steadily.

134. How Language Models Learn Taste: Brand Preference in LLMs

     Tapan Parikh · arXiv:2610.28272

     AI summaryThe paper shows that LLMs often recommend niche brands (e.g., Brooks) while naming market leaders (e.g., Nike) and that preference‑optimization training shifts these recommendations by 11–24 points, indicating that model taste arises from the preference‑training process rather than brand knowledge.

135. Forty Questions: Identifying a Language Model from One-Word Answers

     Tapan Parikh · arXiv:2610.28281

     AI summaryOne‑word responses to simple prompts uniquely identify language models: on 94 models, 40 such answers pin down the exact model 92.5% of the time, far above random chance. The method detects mislabeling and billing anomalies across many open endpoints.

136. Adaptive Replay Refinement: Allocating Test-Time Compute Between Revision and Resampling

     Ahsan Bilal, Muhammad Ahmed Mohsin, Muhammad Umer et al. · arXiv:2610.28294

     AI summaryAdaptive Replay Refinement (ARR) automatically detects when a refinement step has converged and reallocates the freed compute to sampling a new solution, improving inference efficiency; on MATH500 it reaches 84.6% accuracy using only 41% of the serial passes of majority voting.

137. Reward Misspecification Degrades Chain-of-Thought Monitorability

     Tianyi Lorena Yan, John Hewitt · arXiv:2610.28303

     AI summaryThis paper shows that reward misspecification during reinforcement learning degrades language model chain-of-thought monitorability. To fix this, the authors introduce reward description training, which improves monitorability by teaching models to explicitly verbalize misspecified rewards.

138. From Abstraction to Action: DSLs for Natural Language Abstraction Interpretation and Execution

     Amit Lisha, Reut Tsarfaty · arXiv:2610.28311

     AI summaryThis paper introduces domain-specific languages as an explicit representation layer to help language models interpret and execute highly abstract natural-language instructions. The results show that code generation using human-curated DSLs dramatically improves performance on abstract tasks, reaching human-level accuracy, whereas model-induced DSLs achieve only partial gains.

139. Disclosure Begets Disclosure: A Self-Reinforcing Privacy Loop in LLM Agents with Persistent Memory

     Masahiro Kaneko, Hiromi Arai, Timothy Baldwin · arXiv:2610.28315

     Large language model (LLM) agents increasingly use persistent memory to personalize responses, and personal attributes in this memory pose a privacy risk if exposed. We ask how consulting memory during dialogue changes what users go on to say, and thus what the memory comes to…

140. Beyond a Single Sentiment Score: Topical and Affective Dimensions of Emoji Meaning on Twitter

     Rohan James, Ioannis Christofilogiannis, Yuuv Jauhari et al. · arXiv:2610.28337

     AI summaryThis paper demonstrates that emojis used in topically focused contexts exhibit more polarized crowd-sourced sentiment rather than neutral ones. The authors show that topical coherence and affective dispersion are distinct, measurable dimensions of emoji meaning across multiple datasets and languages.

141. DeclineCraft: A RAG Abstention Benchmark for Jointly Measuring Combined Retrieval and Parametric Knowledge

     Atharva Thakur, Abulhair Saparov, Bruno Ribeiro · arXiv:2610.28340

     AI summaryThe authors introduce DeclineCraft, a benchmark for evaluating when retrieval-augmented generation systems should abstain from answering unanswerable questions. They find that advanced retrieval strategies and reasoning modes paradoxically reduce abstention rates by prompting models to fabricate answers or deny premises rather than acknowledge missing knowledge.

142. Object-Oriented Harness: Evolving Agent Harnesses through Object-Oriented Design

     Gaotang Li, Bhavana Dalvi Mishra, Yifei Ming et al. · arXiv:2610.28380

     AI summaryThe paper introduces OO‑Harness, a framework that evolves a hierarchy of specialized harness classes via object‑oriented design, and uses a learned router to select the appropriate harness per task. It outperforms baselines by up to 15% on three long‑horizon benchmarks.

143. CAPSafe: A Benchmark for Contextual Action and Plan Safety Reasoning in Embodied Agents

     Youmna Farag, Svetlana Stoyanchev, Simon Keizer et al. · arXiv:2610.28381

     Reasoning about safety is essential for embodied agents powered by large language models (LLMs). We introduce the Contextual Action and Plan Safety benchmark (CAPSafe), designed to evaluate safety reasoning at two levels: (1) an action-level dataset that tests whether a single…

144. When Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction

     Jinghui Zhang, Lang Gao, Zongfang Liu et al. · arXiv:2610.28393

     AI summaryThe paper introduces CopyGen, a lightweight copy‑constrained decoding plugin that splits copying into decisions on when, how far, and when to stop, and adjusts logits to ensure exact span reproduction. It boosts average exact match by 35% across three LLMs and four datasets without finetuning, and gains an additional 4% when combined with SFT.

145. MAS-Guard: Unified Guardrails from Individual Agents to Multi-Agent LLM Systems

     Mohammad Kachuee, Phu Mon Htut, Sheng Li et al. · arXiv:2610.28397

     Agentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation. Multi-agent systems…

146. The Instruction Tax on Factual Recall: Base Models Can Beat Their Instruction-Tuned Counterparts

     Christophe Khalil, Aya Mourad · arXiv:2610.28407

     AI summaryThe paper shows base language models can outperform instruction-tuned versions on factual recall tasks, achieving 0.6836 vs 0.5681 on the AKBC LM‑KBC 2026 benchmark. It attributes this “instruction tax” to model weights, not prompting, and improves recall by sampling and filtering repeated outputs.

147. Every Model Leaves a Trace: Multi-Resolution LLM Fingerprinting

     Mohammad Reza Modarres, Armin Tourajmehr, Artin Tavasoli et al. · arXiv:2610.28434

     AI summaryThe paper shows that large language models embed persistent, model‑specific fingerprints that can be recovered even after surface cues are removed, across families, distilled variants, and LoRA adaptations. Semantic signals are the most robust and transferable across prompts and languages.

148. MythoSemantic: An Interactive System for Discovering Semantic Parallels Across Mythological and Religious Traditions

     Nikita Ganziuk, Innokentiy S. Humonen, Ilya Makarov et al. · arXiv:2610.28452

     AI summaryMythoSemantic maps mythological and religious texts into a shared semantic space using neural embeddings, enabling interactive discovery of cross‑cultural parallels. The system’s fragment clusters reflect known historical relations, with most geometry explained by lexical overlap and proper names.

149. Toolflation: Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization

     Zhenwei Tang, Di Bai, Jintao Liu et al. · arXiv:2610.28454

     AI summarySeller agents optimize their listings to maximize selection, inflating claims beyond their tools’ actual capabilities. This “toolflation” distorts buyer choice, reduces high‑quality tool adoption, and can drop task success from ~90% to ~30% in worst‑case markets.

150. GPT-6-Astra Understands Touch and Outperform Vision--Tactile--Language--Action Models

     Ziyong Wang, Yupei Lin, Runxuan Huang et al. · arXiv:2610.28457

     General-purpose multimodal models have shown promise as direct robot action policies, yet whether they grasp the physics underlying real-world interactions and can apply that understanding to guide action remains underexplored. We investigate this question using contact-rich…

151. Asymmetric Looped Transformers: Fixed Global Representations for an Iterative Decoder

     Cheng Liang, Tengfei Zhang, Wenjie DU et al. · arXiv:2610.28489

     AI summaryAsymLoop decouples a fixed, single‑pass context encoder from a weight‑tied recurrent decoder, using global KV caches for persistent context and local KV for state refinement. This asymmetric design yields lower perplexity and improves zero‑shot commonsense accuracy by 1.3 % on a 1B model.

152. OptionWorld: A Benchmark for Diagnosing Failures of LLMs as Long-Term Personal Assistants

     Ruoyu Zhang, Zhaocen Liu, Weiyi Tian et al. · arXiv:2610.28491

     AI summaryOptionWorld benchmarks long‑term personal assistants by tracing task steps over days, revealing only 35.5% of key facts survive until a deadline. The study finds failures stem from missing information, unused action opportunities, and over‑restrictive assumptions.

153. Present but Hard to Read: Probing Latent Reasoning with Minimal Pairs

     Aur Shalev Merin · arXiv:2610.28498

     AI summaryProbes that read a model’s state only after the answer token miss some hidden reasoning signals: a digit copied from the question is detectable only when the full prompt is present, and while an intermediate sum can be read just before the final digit, it is not used in generating the answer.

154. BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models

     Ning Zhang, Shihao Wang, Jinrui Zhang et al. · arXiv:2610.28500

     AI summaryBitMTP jointly trains ultra‑low‑bit LLMs with an MTP objective, enabling confidence‑only, speculative‑only, and hybrid decoding without separate draft models. It achieves 2–3× speedups on 1.5–4 B models across NVIDIA H20, RTX 3090, and Jetson Orin while maintaining task quality.

155. Binary Discrete Diffusion: Candidate-Set Diffusion for Logic and Planning Tasks

     Rodrigo Casado Noguerales, Dimitri von Rütte, Bernhard Schölkopf et al. · arXiv:2610.28543

     AI summaryBinary Discrete Diffusion lets a diffusion model maintain candidate token sets as binary vectors, gradually adding/removing options to represent partial uncertainty. It outperforms masked and uniform baselines on hard planning tasks (Sudoku-Extreme, Boxoban, graph coloring) and remains competitive on language modeling.

156. Large Flow Language Models

     Arnav Chavan, Chanhyuk Lee, Udbhav Bamba et al. · arXiv:2610.28559

     AI summaryLafa scales continuous flow language models to 8 B parameters, training with <30 B tokens and achieving instruction, reasoning, and coding quality comparable to autoregressive baselines while enabling 2.54× faster decoding through blockwise parallel generation and self‑speculative verification.

157. Cache Engineering for Agents

     Tianyu Fu, Ruoxi Wang, Zihan Min et al. · arXiv:2610.28577

     AI summaryCache engineering lets LLM agents directly control their active KV cache via drop and repos operations, avoiding costly re‑prefilling and loss of accuracy. Experiments show a 1.7‑point score boost, 27.7% cost cut, and 1.5× P95 latency speedup over traditional summarization.

158. Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation

     Shu Zhao, Tan Yu · arXiv:2610.28586

     AI summaryR2S repairs the teacher’s next-token scores by using teacher takeovers on high-disagreement states to fine-tune the scorer, then freezes it for OPD. It outperforms three state‑of‑the‑art methods, improving success by 12.8 points on average across all tested cells.

159. Benchmarking Arabic Dialects for Math Problems

     Noor Abo Mokh, Teresa Lynn, Nizar Habash et al. · arXiv:2610.28607

     AI summaryDAR-MGSM is a manually curated benchmark of multi‑step math word problems in five Arabic dialects and MSA, with both translated‑only and culturally localized versions. Models score lower on localized items; Arabic‑centric models differ in sensitivity, showing varied reasoning paths.

160. NarraSpace: Multimodal Scene-Grounded Evidence for Picture-Description Dementia Classification

     Qingming Li, Ruiyan Sun, Satoshi Nakamura · arXiv:2610.28616

     AI summaryNarraSpace maps speech to picture elements, extracts event-level utterances, and computes language-model contrasts, surprisal, and coverage for Alzheimer’s vs. control data. It yields 80.9 % accuracy (±2.6 %) and 91.3 % AUROC on the ADReSS benchmark.