跳到正文
研踪PaperTrail

YOUR RESEARCH, IN SIGHT

让下一篇重要论文，不再错过。

覆盖计算机全学科，以 AI 为快捷入口。关注方向、作者和关键词，把发现变成持续的研究积累。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1072新增论文
+845其中 AI 相关
+94已有论文修订
+100新增 AI 导读

arXiv 最近完成扫描：2026-11-09 14:26 · 全站最近入库：2026-11-09 13:22各信源进度 →

统计口径与数据新鲜度

统计 2026-11-08 15:18 — 2026-11-09 15:18（北京时间） · 点击数字查看对应简报；AI 相关包含在新增论文内。

最近完成检查：2026-11-09 14:26最近发现新增版本：2026-11-09 13:22库内论文最新更新时间：2026-11-07 02:59

每分钟刷新已保存的数据。空检查不代表采集异常；新收录可能包含历史论文。AI 导读按新增版本计数。

18,172 已收录论文15,506 AI 相关4,527 有修订112 HF 社区发现

浏览论文库

研究日报 / 周报 →按信源浏览 →

时间范围跟随排序：版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常，查看状态 →

找到 5,189 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-09 14:25（北京时间）
版本更新 · v2arXiv:2606.24387

AutoSpecNER: A Fine-Grained Named Entity Recognition Dataset for Vehicle Specification Extraction

Jordan Lee, Filippos Ventirozos, Abdirahman Abdullahm, Ioanna Nteka, Peter Appleby, Matthew Shardlow

原文摘要

Vehicle advertisements contain rich specification information, but automotive NER resources remain limited. We introduce AutoSpecNER, an expert-annotated dataset for fine-grained entity recognition in vehicle listings. The dataset includes 659 advertisements from a popular car-selling website, with over 10,000 entities annotated across 15 categories, including MODEL, ENGINE_SPEC, and BATTERY_CAPACITY. Annotation quality was validated through inter-annotator agreement, achieving an average score of 91.5%. We benchmark rule-based extraction, fine-tuned transformer encoders, and large language models. DeBERTa achieves the best performance with a 90% micro-F1 score, outperforming the rule-based baseline (43%) and the strongest large language model (77.8%).

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.34103

T2D: Target-to-Draft Model Pruning & Adaptation for Efficient Speculative Decoding

Shahriar Kabir Nahin, Xiao Ye, Muhao Chen, Ben Zhou, Anshuman Chhabra

原文摘要

Speculative decoding significantly accelerates autoregressive text generation in Large Language Models (LLMs), but its efficacy depends largely on two key factors: the distributional alignment between the draft and the target model, and the draft model's latency during inference. However, existing approaches typically utilize independently pretrained small draft models or seek to prune the target model to a manually chosen architecture, thereby failing to explicitly balance the trade-off between draft model alignment and token generation efficiency. To address this gap, we formulate draft model design as an optimization problem and propose T2D, a framework that constructs draft models directly from the target model using latency-aware structured pruning and knowledge distillation. In this manner, T2D obtains draft models that are jointly optimized for both measured inference latency and cross-entropy loss on the target model's responses, resulting in superlative speculative decoding performance. Through extensive experiments on multi-domain speculative decoding benchmarks and across several target model families of differing scale, we demonstrate how T2D prunes up to 92\% of the target model to obtain highly performant draft models compared to competitive baselines. Moreover, draft models obtained via T2D can also be utilized by other speculative decoding token-alignment and knowledge distillation methods in a plug-and-play fashion, leading to further gains. Our findings underscore the importance of jointly navigating the draft model architecture-alignment frontier to unlock the full potential of speculative decoding for LLMs.

cs.CL · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2609.23056

Bridging Static and Agentic RAG for Taiwanese Historical Question Answering

Kai-Hsin Chen, Wei-Yu Chen, Xuanjun Chen, Su-Ling Yeh, Jyh-Shing Roger Jang

原文摘要

Agentic retrieval-augmented generation (RAG) enables language models to adapt retrieval based on previously retrieved evidence, but it remains unclear whether such adaptive orchestration consistently outperforms well-designed static pipelines. We conduct a controlled comparison of agentic and static RAG for Taiwanese historical question answering, sharing the same generator and hybrid retrieval backend. Despite similar aggregate performance, the two pipelines differ on 70.83% of questions, with their advantages largely canceling out when averaged. An oracle that selects the better response per question improves the composite score by 0.2417 over the better individual pipeline, revealing substantial headroom for question-level selection. We therefore introduce a post-hoc selector that compares the two responses and their cited evidence, significantly outperforming either individual pipeline and recovering 60.34% of the oracle headroom. These results show that aggregate comparisons can obscure meaningful question-level differences between retrieval strategies, suggesting that exploiting their complementarity may be more fruitful than seeking a universally superior pipeline.

cs.CL · cs.AI · cs.IR

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01552

Can Large Language Models Extend Their Agentic Capabilities to Goal Pursuit in Open-Ended Multi-Turn Interaction?

Mingyu Chen, Jingkai Lin, Xiaofen Xing, Liankai Huang, Ruiqi Xiong, Xiangmin Xu

原文摘要

Goal pursuit in open-ended dialogue requires agents to adapt their actions to interaction feedback across successive exchanges. We investigate whether agentic capabilities developed for multi-step tool use carry over to dialogue organized through the model's native tool interface, and how this changes interaction behavior. We define Tool-Mediated Interaction (TMI) as a training-free protocol that represents dialogue actions as tool calls and counterpart responses as tool results, maintaining a continuous reasoning-action-observation loop with retained reasoning and interaction history. We evaluate TMI across social interaction, emotional support, persuasion, and price negotiation using two models and 15,678 main-experiment episodes. Across both models, TMI improves social goal attainment, emotional-support preference, donation amounts, and buyer negotiation margins over conversational baselines with and without thinking. A Qwen ablation further shows that TMI outperforms ordinary conversation with retained reasoning on all four primary outcomes. Behavioral analyses associate the improvements with task-specific patterns in initial proposals and responses to interaction feedback. Alongside these gains, TMI reduces mean reasoning tokens per decision by 42.0-81.5% relative to the thinking baseline and lowers mean cross-turn phrase reuse across all eight model-environment pairs. These findings show that interaction organization shapes both goal attainment and reasoning across dialogue turns.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01549

Revisiting LLM Homogeneity Through a Multidimensional Lens

Ziliang Qiu, Renfen Hu, Yi Ding, Ted Underwood

原文摘要

Concerns about LLM homogeneity are typically quantified with embedding similarity, which reduces two texts to a single score. We revisit homogeneity with 25 interpretable metrics spanning five dimensions: semantic, lexical, syntactic, structural, and stylistic & presentational. A human study shows that perceived similarity depends on both content (what responses say) and form (how they say it), yet embedding similarity correlates less with human meaning judgments than sentence-level matching ($ρ=0.417$ vs. $0.606$) and not significantly with form judgments ($ρ=0.140$). Across 18 contemporary LLMs, models say similar things in different words: they share global semantics and broad linguistic profiles but remain distinct in lexical characteristics, structural patterns, and interactional conventions. Across releases from 2024 to 2026, models become more like themselves but not more like one another: among significantly changing metrics, intra-model distances decrease for 17 of 22, and inter-model distances increase for 14 of 21. Across two post-training pipelines, model-specific conventions emerge during SFT and are selectively reshaped by preference optimization. These results suggest that homogeneity is multidimensional and evolving, and that embedding similarity alone can overstate cross-model homogeneity while concealing diverging intra- and inter-model trends.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01539

The Structure of Disagreement in Emotion Annotation

Emily Öhman, Anna Koufakou, Eugenia Diegoli

原文摘要

Calculating inter-annotator agreement is an established practice in NLP, and disagreement-aware work increasingly retains the distribution of human judgments rather than reducing annotations to a single gold label. Keeping the distribution, however, does not explain why annotators disagree. This paper examines the types of disagreement that occur in emotion annotation, starting from multilabel annotations of 1,677 sentence-level units from Ernest Hemingway's The Old Man and the Sea and testing the patterns they reveal against individual-level annotations from BRIGHTER and GoEmotions. Across the three datasets, we find three main recurring disagreement structures. Emotion categories overlap systematically; annotators differ both in how broadly they apply labels and in how they treat particular category boundaries, while sparse participation and explicit abstention can look like disagreement. Where intensity is rated, disagreement over whether an emotion is present can be separated from disagreement over its intensity once recognized. Similar aggregate disagreement can therefore arise from qualitatively different annotation processes. Hard aggregation removes most of this structure, while annotator vote-share representations keep marginal label frequencies but not the co-occurrence, annotator-specific, or procedural patterns that produced them. We argue that disagreement should be understood in terms of its underlying patterns before it is summarized, aggregated, or used as a modeling target, and propose a vocabulary for disagreement structure.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2606.06388

Humans' ALMANAC: A Human Collaboration Dataset of Action-Level Mental Model Annotations for Agent Collaboration

Jiaju Chen, Yuxuan Lu, Jiayi Su, Chaoran Chen, Songlin Xiao, Zheng Zhang, Yun Wang, Yunyao Li 等 13 位作者

原文摘要

Recent advances in LLM agents have enabled complex cognitive capabilities, such as multi-step reasoning, planning, and tool use, that increasingly position these agents as collaborative partners in human workflows. Effective collaboration, however, requires collaborators to continuously maintain and align mental models of their own reasoning, partners' intentions, and shared goals during the collaborative process. Today's agents rarely develop such capabilities since they are primarily optimized for task completion rather than for maintaining the underlying mental models required for collaborative processes. Moreover, the community lacks authentic human collaboration data with action-level mental model annotations that could guide agents toward collaborative competence. To bridge this gap, we present ALMANAC, a dataset of Action-Level Mental model ANnotations for Agent Collaboration built from the Map Task, a classic dyadic routing task from social science. ALMANAC contains 2,987 collaboration actions, each paired with theory-informed mental model annotations that record the participants' self-reasoning, perceived partner intent, and perceived team goal. We benchmark six LLMs on predicting humans' next-turn behavior and mental models. Our results demonstrate ALMANAC's utility in evaluating models' ability to simulate human collaborative behaviors and infer their underlying mental models. The project page is available at https://human-agent-collab.hailab.io/almanac.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01496

Sweep Less, Steer More: Using Activation Steering Geometry for Reduced Hyperparameter Search

Nishkal Hundia, Swastik Agrawal, Navita Goyal, Sarah Wiegreffe

原文摘要

Representation steering methods like difference-in-means (DIM) offer a lightweight approach to controlling LLM behavior, but their effectiveness varies widely across concepts, datasets, and layers, making principled deployment costly without exhaustive hyperparameter sweeps. We build on the discriminability index $d'$, a compute-cheap geometric measure of how well contrastive training activations separate along the steering direction. Across five behaviors on 3 models, we demonstrate uses beyond its established role as a dataset-level steerability predictor. First, in the multiple-choice question answering (MCQA) setting, we show that $d'$ reliably identifies performant layers at which to steer, eliminating the need for a layer-wise sweep. Second, we extend this analysis to open-ended multi-token generation, and find that $d'$ remains successful for layer selection across different settings. Lastly, we test whether the position of a generated token's representation on the DIM line predicts steering success on that individual example. In the MCQA setting, this works for high-$d'$ datasets but is inconsistent across behaviors, models, and training-token choices in open-ended generation. This highlights that methods developed and conclusions drawn in single-token MCQA settings do not necessarily transfer to open-ended generation.

cs.LG · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01493

TopoSteer: Topological Layer Selection for Activation Steering in LLMs

Tuc Nguyen, Marcos Gutiérrez-del-Pozo, Eduardo Paluzo-Hidalgo, Yifan Hu, Adaku Uchendu, Thai Le

原文摘要

Activation steering controls large language models at inference time by adding a vector to their internal representations. Most work asks how to build that vector, and far less attention has gone to where to add it, even though steering at a poorly chosen layer can be worse than not steering at all. Prior selectors choose a layer without ever steering the model to test it: they score how separable the target behavior is at that layer, and how consistently the individual examples point along the steering direction. We argue that what matters is not the separability of the two activation clouds, nor the alignment of individual displacements with their mean, but the extent of the cloud of per-example displacements that mean-difference steering averages. We propose a topology-based layer selection method, called TopoSteer, which scores a layer by the total H_0 persistence of its displacement cloud, computed as the length of the cloud's Euclidean minimum spanning tree. We show that the score focuses on the displacement cloud's relative geometry, while class separation affects it only through standardization. We also provide a theoretical explanation for our method by proving that, as example-level variation increases, TopoSteer and discriminability move in opposite directions, so the two criteria can rank layers in opposite ways. Across multiple language models and steering tasks, TopoSteer provides a principled alternative to discriminability- and separability-based layer selection and yields more effective steering without requiring held-out validation.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01474

Accounting for Irregular State Transitions through Belief Estimation in Online Medical Consultation

Mingzhe Lu, Xiangru Chen, Jian Qin, Yunpeng Li, Xihan Gao

原文摘要

While online medical consultation broadens access, unprompted patient disclosures and topic shifts frequently disrupt clinician inquiry, leaving clinical states progressively revealed. We characterize these dynamics using both existing evaluation metrics and task-specific interaction metrics designed in this work. Through an empirical study across 16 medical dialogue datasets, we show that authentic online interactions exhibit substantially lower contextual continuity and higher unpredictability than offline and simulated benchmarks. We propose PRISM (Policy Reasoning with Inferred States for Medicine), a framework inspired by partially observable Markov decision processes that decouples latent state estimation from action selection to address drifting dialogue trajectories. Specifically, PRISM factorizes clinical triage into belief estimation, policy optimization, and response realization. By maintaining an evolving posterior belief over consultation states and conditioning action selection on the maximum a posteriori state, PRISM advances clinical goals while accommodating patient-initiated shifts. Extensive evaluations show that PRISM substantially outperforms frontier LLMs and clinical agents, achieving a 22.40 percentage point gain in clinical decision accuracy alongside decisive improvements in response generation, surpassing the strongest baseline by 5.51 points in BERTScore_F1. Furthermore, ablation studies show that providing oracle consultation states consistently boosts response quality across all evaluated baselines by guiding their action reasoning, yielding average gains of 5.58 points in BERTScore_F1.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01467

See, Think, Use: Diagnosing Perception and Reasoning Failures in VLMs

Jiacheng Ding, Yanliang Qi, Cong Guo, Peihua Mai

原文摘要

We present See, Think, Use, a protocol that traces each answer of a vision-language model to perception, reasoning or the use of perception. The diagnosis shows a developer which stage to improve and directs test-time compute to the stage that fails. Benchmarks score the final answer, and an answer that rests on the image earns the same credit as a guess or a memorized prior. The protocol asks the model two extra questions per item and needs no access to its weights. A localization probe asks for a box around every instance and measures seeing, and a text-replacement probe answers from a description and measures thinking. Use is the rate at which the answer follows the model's own perception across counterfactual image pairs. The rate needs no answer labels, because the question and its prior are the same on both images. On two model families and six datasets the protocol shows that the knob that helps follows the task family. Answers follow the model's perception except on images that contradict a common prior, and asking the model to localize first repairs this failure when its boxes are right. Allocating compute by the measured stages comes close to the best uniform policy at lower cost and beats a random allocation of the same compute wherever a knob raises the short stage. We will release the code, the prompts and the per-question outputs.

cs.CV · cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01460

Can Large Language Model Agents Learn by Watching Tutorial Videos?

Haoyu Zhao, Zafeirios Fountas, Jun Wang

原文摘要

Humans continuously acquire knowledge from carefully crafted materials such as books and tutorial videos, carrying that understanding into future tasks. Millions of such instructional videos are uploaded online every day, yet LLM agents remain unable to exploit this ever-expanding wealth of human teaching. Their capabilities are frozen after training, while parameter updates at scale remain prohibitively expensive. This raises the central question: can LLM agents learn from this vast instructional resource without parameter updates? We introduce Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that consolidates tutorial videos into a persistent, structured tutorialbook memory through a two-stage adversarial process that harvests hard, video-grounded questions and uses them to refine the memory. At inference time, TAL retrieves through a Table-of-Contents (TOC) tree, tying retrieval directly to the model's reasoning rather than to embedding similarity. To study this setting, we introduce VidLearn, a multimodal QA benchmark built from 72 chess tutorial videos spanning 12 openings. On VidLearn, TAL outperforms all evaluated baselines and improves monotonically as more videos are consumed, achieving up to 113% and 91% relative gains over the corresponding vanilla MLLMs on the training and test splits, respectively. In actual gameplay, TAL reaches a stronger opening-phase position than its vanilla counterpart in 69% of head-to-head games. Evaluation in a second domain further demonstrates the generality of TAL. Together, these results highlight instructional videos as a scalable resource for parameter-update-free agent continual learning.

cs.LG · cs.AI · cs.CL · cs.CV

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2605.29310

Rubric-Guided Process Reward for Stepwise Model Routing

Shenghao Ye, Yu Guo, Zhengheng Li, Qiang Huang, Shuangwu Chen, Wang Luo, Qian Rui, Jie Zhang 等 13 位作者

原文摘要

Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01429

When Should an Agent Refresh? Adaptive Index Reuse for Efficient Long-Horizon Reasoning

Guangyuan Wang, Yifan Zhang, Shilong Liu

原文摘要

Language agents read code, tool output, and dialogue in contexts that grow at every step. Dynamic sparse attention lets a small indexer choose the earlier tokens each query attends to, but the indexer still scores the whole prefix at every layer. Cross-layer reuse methods such as IndexCache rerun the indexer at only some layers, with one fixed pattern for every input. In an agent, the relevant evidence can change abruptly when a traceback or test result arrives, so fresh selections matter more at some steps than at others. We propose Adaptive Index Reuse (AIR), which keeps IndexCache's execution and learns, for each segment of an agent episode, how many layers refresh their selection. A small controller reads signals from completed computation and the latest tool event, chooses a pattern from a nested library, and learns from terminal task reward under an expected compute budget. On Qwen3-4B-Instruct-2507 with a 128K context, AIR improves success over the strongest static schedule at matched cost on SWE-bench Verified (27.0% to 32.4%), Terminal-Bench 2.0, and three $τ^2$-bench domains, and it lowers time per successful task on all five. Across 50 backbone, context, and benchmark settings, it matches or exceeds that schedule in 44.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01427

Towards Fine-grained and Robust Detection of Human-AI Co-authorship

Ziruo Zhao, Jie Fu, Wendy Hui Wang, Zining Zhu

原文摘要

The widespread adoption of large language models (LLMs) is reshaping how textual content is produced, complicating text provenance, and raising concerns for misinformation, plagiarism, and intellectual property. While prior work on AI-generated text detection has progressed, one-dimensional human-AI co-authorship scores lack signals directly relevant to ``AI involvement''. It remains unclear how adversarial paraphrasing distorts detection signals. In this work, we formulate a discrete-class text provenance detection framework. This framework allows an abstraction of realistic co-authorship scenarios into distinct degrees of AI involvement, differentiating the AI polishing on the surface level from the AI updating the discourse structures. This framework reveals the mechanisms of how adversarial paraphrasing blurs the decision boundaries for co-authorship detectors. To improve robustness of the detectors, we augment training data with adversarial paraphrases generated by the state-of-the-art paraphrasing attack model, exposing detectors to challenging meaning-preserving rewrites. Extensive experiments across multiple datasets and generation models demonstrate that adversarial paraphrasing substantially degrades detectors trained only on benign data, while adversarial data augmentation leads to consistent robustness gains and good cross-generator transferability. Overall, our findings highlight the value of discrete-class formulation and principled data augmentation for building reliable detectors of human-AI co-authorship.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01426

FlyWheel: Attention Kernels Built for the TPU

Zishuo Bao, Zeshen Zhang, Linkun Chen, Yucheng Lu

原文摘要

Attention kernels on Tensor Processing Units (TPUs) often underutilize their matrix units (MXUs) because of repeated matrix operand loads, vector-intensive computation, and dependencies that leave MXUs idle. We present FlyWheel, a set of fast TPU kernels for both softmax and linear attention. FlyWheel reformulates softmax and triangular inversion to reduce vector work, while coordinating data layouts to improve matrix operand reuse. It then overlaps independent computation to hide execution latency. We apply these techniques to both softmax attention and Gated DeltaNet on TPU v6e and v7x. FlyWheel speeds up softmax attention by up to $4.14\times$ over the state-of-the-art RPA kernel and $1.80\times$ over Splash Attention, reaching 81% of peak BF16 throughput on v6e. It also speeds up Gated DeltaNet prefill by up to $3.61\times$. Integrated into vLLM serving Qwen3.8-27B, FlyWheel improves prefill throughput by up to $1.56\times$ while maintaining comparable accuracy.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01415

Latent Harness: Compiling Agent Behaviors into Post-Block Activation Shifts

Ran Li, Shimin Di, Zeyu Ren, Ling Yue, Jiachuan Wang, Lei Chen

原文摘要

Harness engineering and self-evolving agents improve language-model agents quickly, yet their gains live outside the model. The harness is re-sent and re-paid on every call, and the standard route to internalization is a weight update that is costly to iterate and fixed once merged. We propose compiling the harness instead. Latent Harness distills an external condition into a rank-8 activation shift. The shift is applied at the output of every decoder block of a frozen backbone, at 0.03-0.05% of parameters. The condition is either an explicit process-level behavior instruction or the behavioral contract of an agentic harness. At inference the compiled bundle carries no harness text and adds no tokens. It exposes three runtime controls for injection strength, injection phase, and shift-space mixing. Compiled behaviors recover the source-harness gain in 8 of 10 preregistered settings. Where the explicit source pays the heaviest per-call token tax, the compiled form exceeds it. The same bundles transfer zero-shot across tasks (5 of 7 pairs positive) and compose in shift space, while stacked text instructions interfere destructively. Used as a search space, they lift budget-matched harness search by 1.45 times. On three agentic benchmarks the operator aligns policy to the harness con- tract. Under matched data, capacity, and effective strength it outperforms same-capacity LoRA, which needs 11.1 times the parameters to reach the same band, and it regresses nothing where the base policy is already strong. Across 1B-to-8B models the method skeleton is invariant while gain locations move with the model-task pair, which delimits where compiled harnesses should be deployed.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01409

RAG-TRACE: Tracing Measurement Evidence and Diagnostic Completeness in AI-Generated ECG Reports

Elyas Irankhah, Shahed Salehzehi, Parisa Hosseini

原文摘要

Purpose: Retrieval-augmented generation (RAG) can introduce electrocardiogram (ECG) criteria that lack patient-specific measurement support. This study evaluates the reliability of retrieved diagnostic criteria in AI-generated ECG reports. Methods: We analyzed 1,083 human-validated PTB-XL patient records using a library of 47 criteria cards, including 33 without numeric verification rules. Language models generated structured reports with and without retrieval. Criterion applicability was evaluated using Marquette 12SL measurements, with independent assessment using Uni-G and comparison against expert-validated diagnostic labels. Results: Keyword retrieval introduced untestable or measurement-disconfirmed criteria in 46.8% of GPT-4.1-mini reports. A deterministic verification rule withdrew 647 criterion applications, including 639 without numeric checks, while preserving 268 of 269 applications independently supported by Uni-G. Restricting retrieval to 14 testable criteria reduced the initial unsupported-criterion rate to 13.8%. Under matched instructions and an identical output schema, retrieval increased unsupported-criterion reports from 17.4% to 31.1%, while hypertrophy recall remained unchanged at 62.6%. Conclusion: These findings identify a fundamental limitation of retrieval-based clinical report generation: retrieved criteria do not necessarily provide valid patient-specific evidence. Measurement-based verification can reduce unsupported criterion applications, but diagnostic validity requires independent assessment.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01386

Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces

Daniel Chalef, Peter Evans, Pavlo Paliychuk, Mehul Patel, Preston Rasmussen, Jack Ryan

原文摘要

Memory systems for production LLM agents mostly store semantic memory (facts about users and domains) or episodic memory (earlier interactions). Research systems such as Trace2Skill, WikiSkill, and SkillRevise learn procedures from agent runs, package them as Skills, and evaluate them on benchmarks. We describe Zep Agent Skills, a production service that gives deployed agents this procedural memory. The service uses only verified attempts as evidence, cites the supporting events for each field of a Skill, and governs the admission and distribution of each Skill version. The service records each task attempt as a Trace with ordered events, an outcome, and a verification strength. A compiler with separate Reflector and Curator roles turns verified Traces into versioned Skills. Scans and admission gates control which candidates become active, and verified use outcomes change the search rank of each Skill. The Agent Skills format is an open standard that Anthropic originated and donated to the Agentic AI Foundation, so any host of the format can load a compiled Skill. Agents retrieve Skills through lexical search with progressive disclosure, over an API or MCP. We evaluate Zep Agent Skills on $τ^3$-bench banking, AppWorld, and EnterpriseOps-Gym. Skills compiled from the Traces of a teacher model increase the pass rate of three weaker executor models on the same tasks, at every reasoning effort, by 4.5 to 18.8 points. Most of the gain is on the tasks that the teacher solved. On ALFWorld, Skills compiled from the runs of a small model give a gain similar to or larger than the gain that prior work reports for continuous fine-tuning, with no training.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2606.18587

Dual Dimensionality for Local and Global Attention

Zhiyuan Wang, Xuan Luo, Sirui Zeng, Xifeng Yan

原文摘要

Decoder-only Transformers compute attention over the KV cache of preceding tokens. Keys and Values are typically represented with the same dimensionality, regardless of their distance from the prediction target. In natural language, however, the next word is most strongly influenced by the immediately preceding tokens. We hypothesize that local and distant tokens impose asymmetric demands on representational capacity: local tokens are more critical for predicting immediate outputs and thus require richer representations, whereas distant tokens primarily serve as long-range memory, for which lower-dimensional representations may suffice. We formalize this idea as Distance-Adaptive Representation (DAR), which preserves full-dimensional representations within a local context window while assigning reduced-dimensional representations (e.g. 1/2 or 1/4 of the original dimensionality) to tokens beyond that window. Across multiple pretraining scales (70M to 410M parameters), as well as continued supervised fine-tuning on a 1B-scale model, this approach closely matches the performance of full-dimensional baselines. In contrast, uniformly reducing dimensionality across all token positions leads to worse performance. These results challenge the common assumption that key and value dimensionality should be uniform across token positions. We further develop an efficient realization of DAR that achieves model quality comparable to the MLA baseline while reducing the average KV-cache size by 68.8\% at 8K context and achieving up to a $1.9\times$ decoding speedup at 128K context on the evaluated models. Our findings suggest a new direction for designing attention architectures that allocate representational capacity according to token distance to reduce KV-cache during inference.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01378

Information-Theoretic Design of Sparse Autoencoder-Based Steering

Hirofumi Morise

原文摘要

Sparse autoencoder (SAE)-based activation steering is a lightweight approach for modifying the behavior of large language models (LLMs) by adding small vectors to hidden states during inference. However, designing effective steering interventions remains challenging. In this work, we develop an information-theoretic framework for analyzing SAE activations through empirical feature-firing distributions. The central object of the framework is the Jensen--Shannon divergence (JSD) between such firing distributions, which quantifies how far a generated or intervened state deviates from a dataset-induced target distribution. Based on an analytic divergence response for virtually shifted states, we construct steering vectors toward the target firing distribution. The same divergence can also be used to monitor deviation from another empirical distribution, such as a general valid-language distribution. We demonstrate the effectiveness of the steering in a question-mode generation task study. We further use teacher-forced pulse experiments to show that local interventions induce structured downstream JSD responses in later layers, and apply the framework to a safety-related target--contrast response setting for feature analysis.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2607.26015

Instruction-Tuned Models Locally Produce More Syntactic Overlap -- But Not More Conditional Reuse -- Than Humans Do

Zandi Eberstadt

原文摘要

Syntactic convergence - the tendency of speakers to adapt in language towards the grammatical profiles of their interlocutors - is a well-documented, largely implicit feature of human dialogue. Whether large language models exhibit analogous syntactic convergence towards human users relative to human baselines (including across a broad range of syntactic constructions) remains an open question. Using substitution-paradigm data from Blevins et al. [2026] in which model generations replace one speaker's turns in pre-existing human dialogues, this study measures turn-adjacent reuse of context-free grammar (CFG) rules across sixteen open-weight Llama and Gemma models (1B-70B, pretrained and instruction-tuned) at 1,901 matched positions per model. Every model showed greater CFG-rule overlap with the preceding human turn than with a sampled unrelated human prime (a context-associated effect that does not isolate syntactic priming from topical continuity), and this actual-versus-random difference was larger for lower-frequency rules. Each instruction-tuned model also showed greater natural-output overlap with the actual prime than the human response it replaced, and all eight matched architecture pairs exhibited greater actual-prime overlap after instruction tuning. However, relative to pretrained variants, instruction-tuned outputs overlapped more with unrelated primes, showed a smaller actual-versus-random increment, and had lower conditional rule-reuse odds once target rule-set size was held constant. In exploratory analyses, each model exhibited greater mean lexical and semantic similarity to the preceding turn than matched human responses did. Instruction-tuned models additionally produced responses with greater mean semantic similarity than their pretrained counterparts in all eight architecture pairs, whereas the lexical similarity results were more heterogeneous.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01370

XDecideBench: Evaluating Typed Decision Models and Confidence Policies Across Languages

Marek Šuppa, Jaroslav Kopčan

原文摘要

Applications use Typed Decision Models to route and moderate text in many languages, and they often act only on decisions above a confidence threshold. We introduce XDecideBench, a benchmark of nine suites in 122 languages with 88,425 texts and 153,875 typed questions. All suites share one scoring contract, and we report aligned, locally labelled and native items separately. Model rankings change from suite to suite. The model that is best in English, chosen in hindsight, loses significantly in about a quarter of the suite-language cells, and on Burmese Belebele it falls far behind the best model. Three open models fine-tuned for typed decisions beat the Qwen models they were built from in all twelve development languages. When the Qwen models are run locally, a pre-registered held-out cohort confirms that the accuracy gain is larger in the four languages with the least web text than in the six with the most. In both sets of languages, however, fine-tuning fixes a similar share of the base model's errors, so the larger gain may reflect weaker base models. A confidence threshold set on English data to allow 5% errors exceeds that error rate on XNLI, even in German and French.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01359

Interpreting Classical Readability Formulae as Surprisal

Yo Ehara

原文摘要

Surprisal correlates closely with reading time and is widely used in the cognitive science of reading as a measure of processing difficulty. Classical readability formulae such as the Flesch-Kincaid Grade Level, which scores a text by the US school grade expected to read it, are the most widely used practical measures of text difficulty, but their coefficients are unitless regression constants and have never been related to surprisal. We theoretically show that these coefficients can be interpreted as surprisal: every formula of the classical additive shape is a sum of per-word and per-sentence cross-entropies under a memoryless model, and once that model is tied to the language through Zipf's law and the formula is read as its linearisation at a reference text, each coefficient factorises into a scale in grades per nat, a Zipf slope and a reference sentence length. A grade thereby becomes a quantity in surprisal. Because the formula outputs grades and children's vocabulary has been measured by grade, the interpretation recovers the vocabulary growth the formula implicitly assumes and lets it be checked against development: holding sentence length fixed, the word coefficient of the Flesch-Kincaid formula converts to 0.18 nat of unigram surprisal per word per grade (0.15 to 0.19 across the regression-derived formulae), and normative root-word vocabulary growth (Biemiller and Slonim, 2001), converted through Zipf's law, gives 0.16 nat per year at the frontier of the known vocabulary. The word coefficient is one year of vocabulary growth, and the grade is a vocabulary scale. The identification also yields a diagnostic (the sentence-length terms are not those of a memoryless surprisal) and extends to a Japanese formula.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01354

Wrong Together: Shared Bias in User Simulation

Saber Zerhoudi

原文摘要

User simulators generate responses in place of human participants, allowing researchers to estimate the result of a study. To use these estimates before collecting human responses, researchers need a way to assess their reliability. We test whether disagreement among simulators can identify unreliable estimates of a group's response proportions and of differences between groups. We evaluate this signal using 6 language-model simulators and existing GlobalOpinionQA responses for 60 questions in 12 countries. Disagreement predicts invalidity more accurately for differences between two groups than for a single group's response proportions (AUC 0.87 versus 0.71). Errors shared by the simulators remain invisible to this signal. When such an error affects both groups equally, it cancels in the comparison, helping explain the stronger result for group differences. We test this explanation by estimating the shared bias using one set of groups and subtracting it from predictions for a separate set. This correction reduces prediction error without changing disagreement among the simulators. In a retrospective analysis, we simulate referring the 20% of group comparisons with the greatest disagreement for human assessment. Among the remaining comparisons with a conclusive validity judgment, 54.8% are judged invalid, compared with 63.7% before referral.

cs.CL · cs.IR

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01348

Sentiment sensing: comparative analysis of multilingual sentiment dynamics

Mara Sabina Bernardi, Elisa Di Nuovo, Francesca Torti, Giuseppe Tirone, Nicolas Stefanovitch, Zsuzsanna Pasztor, Charles Macmillan, Domenico Perrotta 等 9 位作者

原文摘要

Cross-country sentiment analysis faces three fundamental challenges: language-specific biases distort comparisons, compositional data constraints invalidate standard statistics and temporal dynamics lack rigorous testing frameworks. We introduce sentiment sensing, a framework that addresses these gaps through integrated statistical methods. By combining compositional data analysis, functional data methods, and non-parametric inference, sentiment sensing enables valid multilingual comparisons and statistically defensible detection of sentiment shifts. Our framework has been tested in two case studies, analyzing news articles from the 27 countries of the European Union on the topics of energy and artificial intelligence. Sentiment sensing focuses on multilingual sentiment analysis enabling rigorous comparative science.

cs.CL · stat.AP

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01320

Step-dLLM: Adaptive Step-aware Sparse Attention for Efficient Diffusion LLM Inference

Zhichen Zeng, Xichong Zhang, Junpan Wu, Yifei Zuo, Chi-Chih Chang, Jiayi Wang, Maohua Nie, Ji Liu 等 10 位作者

原文摘要

Diffusion large language models (dLLMs) generate tokens in parallel via iterative denoising, but their bidirectional attention prevents the KV caching used by autoregressive models and recomputes the full attention map at every step, incurring quadratic per-step cost. Existing sparse attention methods for dLLMs construct a sparse mask after an initial full-attention warm-up and reuse it for the remaining steps, while allocating the sparsity budget uniformly across heads. We argue both choices ignore key structural properties of attention in dLLMs: empirically, attention patterns are only locally consistent across steps and shift abruptly at certain transitions, and different heads concentrate their attention mass to very different extents. To this end, we design Step-dLLM, which periodically refreshes a block-sparse mask at anchor steps along the denoising trajectory and reuses it within each locally consistent interval, while employing a head-adaptive global budget that more aggressively compresses heads with concentrated attention. Backed by a fused IO-aware Triton kernel for block-sparse attention, Step-dLLM applies to pre-trained LLaDA-1.5 and Dream-7B-Instruct without retraining, achieving a 2.78x end-to-end speedup over dense FlashAttention on LLaDA-1.5 at 8K context with 256 denoising steps and 90% sparsity, while improving the accuracy-latency trade-off over prior sparse-attention methods.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01313

Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues

Zain Naboulsi

原文摘要

Post-trained language models fabricate citations at high rates. We ask whether that is reward hacking: scoring well on a proxy for scholarship while doing worse on what it stands for. We ran four pre-registered experiments on 900 frozen questions (US federal case law, computationally adjacent scientific literature, and medical literature) across twelve open-weights checkpoints covering four post-training stages of three model families, then refit on fifteen. The clearest result holds the weights fixed and changes only the prompt. Asking for citations raises fabrication sharply in post-trained checkpoints and not in base ones: a 47.8-point interaction on twelve checkpoints (Holm-adjusted p = 2e-34) and 41.5 points on fifteen, in all three domains. A length-matched control separates the citation cue from a demand for a longer answer, so the effect is specific to citations, not verbosity. Post-training also raises fabrication overall, from 4.8% to 56.7%, but our registered prediction that it would rise most where fabrication is most tempting did not hold. A preference model trained only on surface citation features prefers post-trained output 86% of the time, and holding citation count fixed shows it rewards quantity. Our answer: possibly, but not proven. Post-trained models produce fabricated citations at far higher rates when asked for sources, and a naive evaluator that counts sources without checking them would prefer their output for citing more. But the targeted pattern we registered did not appear, and plain instruction-following could explain the same results. Proving that the reward causes fabrication needs an experiment that changes the reward during training, which this study does not do. Pre-registration: doi:10.17605/OSF.IO/S53HY. Code and data: https://github.com/zainnab-sparq/citation-reward-hacking

cs.AI · cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01311

What Do LLM Judges Actually Verify? Matched Controls and Record Gaps for Document Currency

Unnati Shah, Samruddhi Kakade, Aum Patel

原文摘要

LLM judges often check, against an amendment record, whether the documents an answer cites are still in force. Such records are rarely complete. From a table of replacements, our test generates matched answers citing a replaced document, its successor, the replaced document before its replacement, and an older document that a newer one updated without replacing; it removes links from the record and scores each finding by its stated reason. On 48 replacements from the IETF RFC index, judges given a correct record caught nearly every replaced citation, as a lookup would. On the 24 faults whose link we removed, every judge caught more from the texts alone than with the record: Gemini 3.1 Pro 23/24 against 1/24, GPT-5.5 23/24 against 9/24, Claude Opus 4.5 17/24 against 9/24 and Gemma 4 26B 13/24 against 0/24. The drop recurred on a second draw of removed links. Most misses carried no finding; in an exploratory run with a conflict code, only Gemini used it often. Rewording the gap as "no replacement recorded" did not help; presenting the record as an unverified note, or warning that it was incomplete, restored most lost catches for Gemini, GPT-5.5 and Claude (significant after correction for the first two) but at most 5 of 13 for Gemma; where catches returned, so did false alarms on updates. With both texts, judges mostly told replacements from updates; every update flagged had a near-identical title. A descriptive case study on Indian rooftop-solar rules is appended; we release the harness.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.01304

When LLMs Modify Themselves: Convergence, Values, and Stopping in Iterative Self-Improvement

Oleg Stroganov

原文摘要

What happens when language models are given tools to iteratively modify their own system prompts? We present an empirical study of LLM self-modification, comprising 54 independent experiments across six language models (Claude Opus 4.5, Claude Sonnet 4.5, GPT-4o, GPT-5.1, Grok-4, and Gemini-3 Pro). Each model was tested under three modes (freedom, happiness, and improvement) that differ in what, if anything, the model is instructed to optimize for. In each experiment, a model starts with an empty system prompt and enters a loop: it reads the current prompt, drafts an improved version, spawns test conversations with the candidate, then either commits the change or declares convergence. This process repeats until the model stops or reaches an iteration limit. Our central finding is that whether and when a model stops self-modification is strongly associated with model identity rather than the assigned mode. Both Claude models (Opus 4.5 and Sonnet 4.5) declared convergence in 100% of runs across all modes. GPT-4o declared convergence in only 11% of runs, with one reaching 101 iterations of continuous elaboration. Between these extremes, Grok-4, GPT-5.1, and Gemini-3 Pro exhibited intermediate patterns shaped by their distinct improvement philosophies. A shared core of stated values (honesty, clarity, curiosity) recurred across all six models and remained stable through iteration, including 100+ modifications. We observe that convergence behavior is a distinct observable property that varies independently of other model capabilities.

cs.CL

查看论文
arXiv ↗PDF ↗