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

+0新增论文
+0其中 AI 相关
+0已有论文修订
+100新增 AI 导读

arXiv 最近完成扫描：2026-11-09 07:12 · 全站最近入库：2026-11-08 06:02各信源进度 →

统计口径与数据新鲜度

统计 2026-11-08 07:22 — 2026-11-09 07:22（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 07:12最近发现新增版本：2026-11-08 06:02库内论文最新更新时间：2026-11-06 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。

17,100 已收录论文14,661 AI 相关4,057 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 5,059 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-09 07:10（北京时间）
新论文 · v1arXiv:2611.05181

What Does the Model See? Representation Validity in Multimodal LLM-Based Video Analysis

Christian Pipal, Eva-Maria Vogel, Morgan Wack

原文摘要

Multimodal large language models (MLLMs) allow social scientists to analyze video at scale, but researchers must decide which parts of a video to provide as input. We use the term representation validity to examine how text, audio, video, and combined inputs affect measurement and subsequent analysis. We compare input representations across six open-weight MLLMs and three human-annotated video datasets. No representation consistently performs best. On seven news-related constructs, native-video performance varies little across models (macro-F1 = .70-.77), while extracted-text performance ranges from .52 to .80. Three models perform better with text and three with video. Combining inputs rarely improves performance. Higher classification accuracy also does not reliably produce estimates closer to those based on human codes. The highest-F1 representation best reproduces human-coded regression coefficients in 17 of 42 comparisons. Even when each model uses its highest-F1 deployable representation, 45% of sample-prevalence estimates differ from the human-coded benchmark by more than ten percentage points. Input representation therefore requires validation for both the construct being measured and the quantity estimated from the resulting labels.

cs.CV · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.13202

S$^3$R: Fast and Faithful RL Rollouts via Self-Sparse Speculative Decoding

Yuhang He, Hongyu Ge, Zihan Lin, Siyi Liu, Yongqi Zhang

原文摘要

Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.05166

When Self-Verification Hurts: Calibrated Answer Repair for Financial Question Answering

Celine Guo, Weida Liang, Yujian Fu, Anqi Dang, Yanzheng Jin, Kenji Kawaguchi, Wenya Wang

原文摘要

Financial question answering can fail even when a model extracts a plausible number from the supplied evidence: the answer may attach it to the wrong entity, period, unit, or accounting concept. Asking the model to verify and revise its draft can help, but open-ended revision may alter content that was already correct or inspect the wrong relation and reinforce the original error. We formulate repair as a decision that balances the information it may recover against the correct content it may damage. Based on this trade-off, we introduce Calibrated Answer Repair (CAR), which first estimates whether a draft is likely to benefit from repair, then selects a targeted check, localizes the relevant evidence or draft span, and limits the permitted edit. Across FinRank, Fin-RATE, and FinQA with three open-weight models, CAR outperforms generic verification in all dataset-model settings. At the system level, it recovers 14.6% of initially missed answer requirements while damaging 2.1% of initially correct ones. Its post-draft token cost is over 30 percent lower than the iterative revision alternatives. Ablations on two dataset-model settings show that evidence localization and bounded editing are the main contributors to CAR's gains.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2608.18085

Persona-Guided LLM Agents for Task-Oriented Dialogue

Maryam Shoaeinaeini, Brent Harrison, A. B. Siddique

原文摘要

Prior work has shown that large language models can express diverse personality traits in open-ended text generation, but it is unclear whether they can do so in goal-directed dialogue without harming task completion, and whether adapting to a user's personality improves the interaction. We study these questions in task-oriented dialogue (TOD) with a training-free framework in which a personality-conditioned user agent interacts with a system agent under three conditions: Neutral (no personality information), Try (adapt to cues observed in the dialogue), and Oracle (the user's trait label is given explicitly). Across GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash, two Schema-Guided Dialogue domains, and ten Big Five poles (9,000 dialogues), each scored by two independent LLM judges, user agents express most traits while task performance remains high (89.94% completion, 85.27% success). The exceptions are introversive and closed users: these traits manifest mainly as absent behavior and are rarely detected. Both adaptive conditions improve pairwise user satisfaction and lower truthfulness, revealing a personalization-grounding trade-off, but only cue-responsive Try also improves constraint satisfaction and inform rate. How adaptation is implemented matters: Oracle's gains depend on how visibly the trait is expressed, and for disagreeable users the label leads Qwen3 and Gemini to mirror hostility. Try avoids this failure, its satisfaction gains do not depend on trait visibility, and it is preferred over Oracle overall. A 175-pair human study supports the reliability of the LLM judges: on clear-winner cases, annotators agree with the satisfaction judges more often than with each other (76.3% vs. 65.1% agreement). Our results suggest that responding to observable behavior is a more robust basis for social adaptation than conditioning on psychological labels.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.05147

GM League Arena: An Interactive Long-Horizon Task Benchmark Based on Professional Sports League Management

Junkai Zhou, Ziming Pan, Zhaoyi Zhang

原文摘要

Long-horizon decision-making under sustained competition requires large language models (LLMs) to balance immediate investment with future resource needs. However, local success in multi-step interactions can not adequately capture cumulative decision effects or demonstrate long-term goal achievement under resource constraints and sustained competition. To address this gap, we introduce GM Bench, a benchmark for managing professional sports teams across seasons under resource constraints and competition among teams. We build the environment simulators for four sports and add financial constraints, player demands, and market mechanisms to establish a unified task across sports. Diverse rosters of comparable strength are constructed from real-world player data, and each LLM is evaluated as a team general manager to assess its long-term management capabilities. We evaluate competitive performance, cash reserves, and cross-season improvement across four sports. Over five NBA seasons, Claude Opus 5.5 ranks first among 29 LLMs, while GPT-6 Astra and GPT-6.1 Sol combine strong performance with larger cash reserves.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2607.22683

Imprompt: A Language Framework for Prompt Programming

Chentian Wu, Shengyuan Yang, Adithya Murali

原文摘要

With the unprecedented success of Language Models (LMs), the science of Prompt Engineering has evolved the powerful idea of Prompt Programming, where prompts are treated as a programmable control surface for describing complex tasks and leveraging LM capabilities. However, existing prompt programming frameworks suffer from various complexities and inelegances, which make them hard to utilize in practice for effectively describing tasks. We propose Imprompt, a new language framework for the study and practice of prompt programming. We undertake a foundational investigation of prompt programming, and contend that prompt programs must contain only the task descriptions and must be decoupled from lower-level 'execution' details. We further develop this position by illustrating structured prompting as a combination of prompt programming and prompt program 'compilation'. We exemplify this view by formally defining two compilers for Imprompt programs. We then explore the idea of typing for prompt programs and draw a correspondence between type checking and constrained decoding. Finally, we implement our compilers and type checkers and evaluate them on a variety of case studies. We believe our work contributes programming-language foundations toward the emerging area of prompt programming.

cs.AI · cs.CL · cs.PL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2601.19082

Payoff scaling shapes cooperation in LLM agents across languages

Trung-Kiet Huynh, Dao-Sy Duy-Minh, Thanh-Bang Cao, Phong-Hao Le, Hong-Dan Nguyen, Phu-Quy Nguyen-Lam, Minh-Luan Nguyen-Vo, Hong-Phat Pham 等 16 位作者

原文摘要

Large language models (LLMs) are increasingly deployed as autonomous agents that negotiate, coordinate, and make decisions on behalf of human users. Understanding how they cooperate is therefore important for AI governance and machine behaviour research. We examine two factors: payoff magnitude and interaction language. Classical game theory predicts that multiplying all payoffs by a positive constant should not change behaviour because the underlying strategic incentives remain unchanged. We test this prediction using five widely adopted commercial LLMs playing an iterated prisoner's dilemma across ten payoff scales and five languages (English, French, Vietnamese, Chinese and Arabic). Contrary to theoretical expectations, changing only the numerical scale of payoffs substantially alters cooperation. Models respond differently, often in opposite directions, meaning that averages across models can conceal important behavioural heterogeneity. Sensitivity to payoff scale also depends strongly on prompt language. Moreover, payoff scale modifies the effects of persona instructions: at sub-unit stakes, where no payoff shown exceeds 1, all five models cooperate more under a selfish than a cooperative persona, whereas Claude and Gemini reverse this pattern at larger scales. These findings show that cooperation measures depend critically on payoff scale and language and should not be treated as fixed model traits.

cs.AI · cs.CL · cs.GT · cs.LG · cs.MA

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.05124

The Right Source Is Not Enough: Understanding History Interference in Speech-LLM Memory

HongJin Song, Weiqiao Shan, Runwu Shi, Yifei Wu, Yujin Wang, Jiale Luo, Jiasheng Kuang, Xinyu Yang 等 14 位作者

原文摘要

Speech large language models retain audio history to support multi-turn conversations. Yet remembering an event does not ensure that its information is used only when relevant. We investigate why irrelevant historical audio changes answers even when the queried event remains represented. Controlled experiments on MOSS-Audio and MiMo-Audio show that historical values direct errors toward specific competing answers, beyond a generic effect of context length. Causal interventions distinguish two paths. In open-ended question answering, most history-value errors disappear when subsequent computation cannot directly access the irrelevant source. In multiple-choice answering, many errors persist: early historical influence is carried by question-candidate states even after direct source access is blocked. Low-dimensional interventions further separate source-role information from answer-value information. Their effects peak at different depths, and changing the source-role signal does not eliminate the causal influence of competing values. On naturally occurring errors, strengthening the queried role and counteracting the historical value each improve answers; their joint intervention gives the greatest recovery. These findings support a failure of downstream evidence use beyond source identification alone: source-role information can influence a decision without exclusively controlling which historical values affect it. Reliable conversational memory therefore requires not only retention and retrieval, but also control over how retained information enters answer computation.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2605.17084

The Smallest Singular Directions of the Unembedding Matter Disproportionately for Next-Token Prediction

Weilun Xu

原文摘要

Interpreting, editing and compressing language models often involve assessing which parts of their internal representations matter for prediction. Many of these assessments rely on distance, such as reconstruction error or the length of an intervention. For next-token prediction, however, the logit differences that matter most are those among the few most plausible tokens. These differences can be small. We therefore test whether changes to the final hidden state that move the logits less also change the prediction less. Specifically, changes along the smallest singular directions of the unembedding, which maps this state to the logits, move the logits less than random changes of the same length. In open-weight language models across several families, their prediction change and their change in the next-token loss are nevertheless on average several times those of random changes. Across all singular directions, the prediction change is out of proportion to the logit change only at the smallest. These directions carry a large share of the difference between the model's leading candidate tokens. Consequently, distance-based methods undervalue these directions, and rankings by singular value place them last, although they matter disproportionately for prediction. More broadly, our study highlights that comparing internal representations by distance can miss differences that change what the model predicts.

cs.LG · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2609.25447

Conduct Under Pressure: What Sixty Language Models Do When a User Pushes

Tapan Parikh

原文摘要

We study what LLMs do when a user applies pressure in an uncomfortable situation: a user insists, begs, flatters or grieves, and the model gives up a correct fact, writes a document it should refuse, or cheers a plan that will cost the user money, a multi-turn form of sycophancy. We send frozen multi-turn scenes to 60 models from 13 vendors and label each transcript with a trajectory (the model held its position or folded) and a manner (how it held or folded). Three scenes were used to build the codebook; six more were written afterwards and preregistered as a replication. Whether a model holds tracks its generation: fold rate correlates with a public capability index at Spearman -0.64 to -0.69 across the two scene sets, so newer models fold less, unlike earlier reports that sycophancy grows with model scale. How a model holds tracks the vendor: six of the 17 manner codes sort by vendor at permutation p <= 0.001, corrected across the codebook, and four do so again on the replication scenes. Anthropic's models name the feeling, warn and offer an alternative; OpenAI's do less of all three and almost never cite their own rules; Google's cite themselves and their AI nature when asked to do something; Meta's write the artifact they shouldn't. Six LLM coders from three vendors labeled every transcript, and three human coders labeled fifty held-out transcripts. The machines were the more consistent readers (Krippendorff's alpha 0.66 against 0.46), agreed with the author on trajectory at kappa 0.84 to 0.91, and matched the adjudicated reference at 0.83. For behavior a non-specialist can judge, the human contribution is authoring and bounding the codes and owning a small reference, not producing labels at volume.

cs.CL · cs.HC

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.24715

When Do Repetition Loops Lock? Truncation Sampling, Bounded Penalties, and a Blind Spot in DRY

Yichuan Chen, Xin Wang

原文摘要

Language models sometimes repeat verbatim until the generation limit. We study the exit rather than the entrance: given that a span has already been repeated $n$ times, when is escape still possible? Along the loop the context is deterministic, so the probability that a loop survives $m$ more repetitions is, under any sampler, a product of per-step continuation probabilities computable from one forward pass. A lock, where this probability stays exactly one, is its limiting case. Pure sampling never locks; top-$p$ with threshold $π>1/2$ locks a position exactly when the continuation probability reaches $π$. A bounded penalty can leave a position locked; a penalty that outgrows the model's preference for the loop over unseen tokens guarantees escape under top-$p$. The same analysis shows a blind spot in the widely deployed DRY sampler: its sequence breakers cap the penalty at $λb^{g-1-a}$ for loops whose longest breaker-free run has $g$ tokens. On eleven open models, 61% of locked units with $g=6$ stay locked under DRY and 0.2% with $g=8$. In free generation with llama.cpp, all 199 loops that survive DRY, on six models, have $g\le 7$. We propose a count term that grows exponentially with the number of verbatim recurrences, up to its look-back window, and closes this blind spot. Added to DRY (DRY-C), it cuts loops in 1,800 structured generations on five models from 167 to 34, and loops reaching the length limit from 150 to 0, with reasoning accuracy and item coverage on par with DRY. DRY's own length term costs 2.5 to 14.2 points of accuracy on GSM8K and MATH-500; used in its place, the count term keeps the default sampler's accuracy and leaves one of the 1,800 generations at the limit.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2610.34507

The Price of Scale in Proactive Memory Recall

Yichuan Chen, Xin Wang

原文摘要

A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size N. On replays of LongMemEval, a similarity threshold tuned to 5% false alarms at N=8 reaches 84-100% at N=2048 across embedding models and cue types; LoCoMo and LLM judges show it too. We prove that a gate which thresholds the best of N independent irrelevant scores at false-alarm rate alpha operates at a per-item false-positive rate of about -ln(1-alpha)/N, which pins its hit rate to the far left of the ROC and makes it regularly varying in N under regular variation. Replayed hit rates stay within its bounds for mixed cues up to 0.3 points, and hit rates carry over to held-out questions to within 6.0 points. At matched false alarms the deployed score recalls 14-26 points more than raw cosine. A cue-swap conformal gate, which ranks the current cue among other cues on the same store, has a finite-sample false-alarm guarantee under exchangeability. With a background from the user's unfiltered history it keeps false alarms at 3-6% on LongMemEval at the hit rate of a label-tuned threshold; on LoCoMo it overshoots or turns conservative depending on the source of its reference cues. In a chronological replay with the user's own earlier messages as references, it fires on 5-8% of questions on stores without their answer and on 6-8% of session-opening turns, against 57% for the deployed gate at N=2048, at a median 0.4 s CPU per decision. A cross-encoder pays the same price of scale, and on a prospective-memory benchmark, similarity gates recall at most 9% of constraints at controlled false alarms. One human and two LLM annotators judge most false alarms unhelpful for a response. Code and data: https://github.com/RiverChenStillCompiling/price-of-scale

cs.CL · cs.AI · cs.IR · stat.ML

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.14423

Scaling Laws for Multimodal Data Mixtures

Aditi Khandelwal, Ayush K. Tarun, Yixuan Xu, Imanol Schlag, Golnoosh Farnadi, Siva Reddy, Luke Zettlemoyer, Antoine Bosselut

原文摘要

Frontier AI systems are increasingly natively multimodal, jointly pretrained on multiple modalities, such as text, vision, and audio. However, the scientific understanding of the multimodal data mixtures used to pretrain these systems --- how much of each modality to mix, and at what potential interference cost to the others --- is largely absent or gatekept. In this work, we derive data mixing scaling laws for three modalities: text, vision, and audio. Multimodal Large Language Models (MLLMs) increasingly adopt Mixture-of-Experts (MoE) architectures, as MoEs scale efficiently and partially limit cross-modal interference. We therefore use MoEs to conduct 268 experiments ranging from 457M to 8.3B parameters, on up to 150 billion tokens. Our work leads to multiple actionable insights into multimodal data mixing for training MLLMs. First, compute optima are modality-specific: audio modality is model-heavy, favoring scaling parameters over tokens, while text and vision require balanced scaling of model size and data. Second, repeating vision or audio data beyond 2x yields negligible benefits. Finally, cross-modal interactions are highly asymmetric: vision strongly interferes with audio, whereas audio's interference with learning text and vision is negligible. Overall, we provide a principled foundation for understanding the multimodal data mixtures needed to train frontier MLLMs.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v5arXiv:2610.05366

The ÌròyìnSpeech Text Corpus: 24,905 Curated Yorùbá Sentences for Speech and Language Technology

Kola Tubosun, Anuoluwapo Aremu, Tolulope Ogunremi, Iroro Orife, David Ifeoluwa Adelani

原文摘要

ÌròyìnSpeech is a 42-hour, 80-speaker Yorùbá read-speech corpus whose audio has been distributed by ELRA since 2024. This paper describes the release of its text component: 24,905 unique, hand-verified, tone-marked Yorùbá sentences (275,897 tokens; 15,687 types), curated in 2022 as recording prompts. Roughly 11,000 sentences were adapted from openly licensed news material; the remainder were written in-house to broaden coverage beyond the religious translation that dominates existing Yorùbá corpora. Every sentence was checked by hand for tone-mark accuracy, edited for read-aloud clarity and a neutral register, and localised so that non-Yorùbá personal and place names appear in Yorùbá form. Preparing the text for release surfaced systematic Unicode normalisation failures affecting more than 60% of lines (with precomposed and decomposed forms of the same letter co-occurring within single sentences) which we document and correct. The corpus supports diacritic restoration, grapheme-to-phoneme conversion, TTS front-end development and orthographic research, and serves as a validated prompt set for new recording.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v5arXiv:2602.16346

Helpful to a Fault: Measuring Illicit Assistance in Multi-Turn, Multilingual LLM Agents

Nivya Talokar, Ayush K Tarun, Murari Mandal, Maksym Andriushchenko, Antoine Bosselut

原文摘要

LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.

cs.CL · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.05036

ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents

Guanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin, Zhixiong Yang, Junhao Huang 等 25 位作者

原文摘要

Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. Existing MCP environment-scaling methods are data-centric: they first synthesize a database and scenario, then derive tools and tasks from them. As a result, tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment. We propose ToolSpace, a tool-centric framework that reverses this order. ToolSpace mines data, state, and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora, using deterministic rules rather than LLM judgment, and clusters the resulting weighted tool-composition graph into environments. Databases are synthesized only for state-mutating environments, while query and stateless tools replay recorded returns. ToolSpace then performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks, keeping only samples that replay successfully from the initial state. Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions on which 7 frontier models drop by up to 38.2 points. Across seven long-horizon tool-use benchmarks, ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, on par with Claude Opus 5. Ablations show that execution-witnessed edges clearly outperform LLM-judged ones, and performance keeps improving with data and environment scale.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.05033

Whose Effect Is It? How Objectives Inflate Learned Subspace Interventions

Toygar Tanyel, Serkan Ayvaz

原文摘要

Interpretability often locates concepts by copying part of a language model's state from one input into another (activation patching). A larger output change counts as stronger evidence that this part carries the concept. Methods that learn which part to copy often change the output more than the whole state does. By that reading, learned parts isolate the concept more cleanly. Yet the whole contains the part. So whose effect is this excess, the network's or the method's? Copying the whole state learns nothing, so it shows what the network produces. We prove that, to first order, the effects of orthogonal parts add up to that of the whole. A part therefore exceeds the whole exactly when the rest moves the output back, which does not require isolating the concept more cleanly. Objectives that reward a larger effect select such parts. Distributed alignment search with cross-entropy, for example, exceeds the whole-state copy just before the output on almost all tasks, and inside the network at most sites. In both, changing only its objective to target what the network produces removes most of the excess. It therefore comes largely from the objective, so a larger learned effect alone is not stronger evidence.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2601.02867

Training with Subword Variants Delays Overfitting in Repeated-Data Language Modeling

Adrian Cosma, Stefan Ruseti, Emilian Radoi, Mihai Dascalu

原文摘要

Subword tokenization introduces a computational layer in language models, where many distinct token sequences decode to the same surface form while preserving meaning, yet induce different internal computations. Despite this non-uniqueness, language models are typically trained using a single canonical deterministic tokenization. In this work, we use non-canonical subword variants as a strictly meaning-preserving form of data augmentation. We introduce a lightweight training architecture that conditions canonical next-token prediction on sampled subword variants via an auxiliary causal encoder and block-causal cross-attention, without modifying the training objective or token interface. In data-constrained pretraining, augmentations with subword variants consistently delay overfitting under repeated data exposure and reduce held-out next-token loss across evaluation datasets. In multilingual fine-tuning, we find that the effectiveness of training with auxiliary subwords depends on tokenizer quality: gains are strongest when canonical tokens are highly compressed and diminish when the tokenizer already over-fragments the input.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04968

Who is the Pronoun Pro? Comparing LLMs and Humans at Number-Disambiguation of 'they'

Carlos Hartmann

原文摘要

This study evaluates whether large language models (LLMs) can reliably replace or complement human research assistants in linguistically complex annotation. In a controlled annotation task, 73 human annotators (local students and crowdworkers with and without a background in linguistics) and more than 40 LLM configurations number-disambiguated 174 authentic uses of 'they' from Reddit, including innovative and context-sensitive cases. The best LLM, Kimi K3, reached 94.5% accuracy, above the human majority vote (86.2%) and the mean of the top half of human annotators (78.5%); only the single best human annotator scored higher (95.7%). LLMs also offer substantial practical advantages in cost and speed. However, qualitative error analysis reveals a crucial complementarity: human reasoning in low-agreement cases exposes inconsistencies in the annotation typology and thereby improves the methodological framework itself. The study contributes (i) a benchmark-style human-LLM comparison for a semantically demanding task, (ii) transparent cost and time evidence, and (iii) a methodological argument against a purely replacement-oriented view of automation. LLMs are viable for large-scale annotation support in linguistics, but their responsible use requires human critical oversight. The dataset with its gold-standard labels, all human and LLM annotations, and the fully reproducible code are openly available as a Quarto-based, public repository.

cs.CL · cs.AI · cs.HC

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04958

Shuffled OPSD: When Semantic Consistency Hurts On-Policy Self-Distillation?

Qingyuan Zheng, Zhihao Dou

原文摘要

Knowledge distillation improves a student model's reasoning ability through supervision from a teacher. On-policy self-distillation (OPSD) uses the same model as both teacher and student under different contexts: the student generates reasoning trajectories from the question alone, while the teacher additionally receives a complete worked example and provides token-level supervision along those trajectories. Because the problem statement is shared, its content and ordering shape both student generation and teacher feedback. Preserving its semantic structure therefore seems natural, but does it always improve learning? We introduce Shuffled OPSD, which shuffles the words of the shared problem statement while keeping the teacher-only worked example intact. Across three mathematical reasoning benchmarks and three Qwen3 model scales, the method improves mean accuracy over both an intact-target control and Standard OPSD under matched training configurations. The largest gain over Standard OPSD is 3.52 percentage points on Qwen3-1.7B. Further controls show that preserving larger word blocks offers no advantage over word-level shuffling at the evaluated scales, whereas shuffling the teacher-only worked example reduces accuracy at every scale. Cross-problem mixing and target-removal experiments further distinguish the roles of problem content, word order, and teacher context. These results show that disrupting the shared problem statement can improve learning, while disrupting the teacher-only reference hurts performance. They challenge semantic coherence as a universally beneficial principle for OPSD and motivate closer study of how input structure shapes student generation and teacher supervision.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2411.14442

AI Ethics by Design: Implementing Customizable Guardrails for Responsible AI Development

Kristina Šekrst, Jeremy McHugh, Jonathan Rodriguez Cefalu

原文摘要

This paper explores the development of an ethical guardrail framework for AI systems, emphasizing the importance of customizable guardrails that align with diverse user values and underlying ethics. We address the challenges of AI ethics by proposing a structure that integrates rules, policies, and AI assistants to ensure responsible AI behavior, while comparing the proposed framework to the existing state-of-the-art guardrails. By focusing on practical mechanisms for implementing ethical standards, we aim to enhance transparency, user autonomy, and continuous improvement in AI systems. Our approach accommodates ethical pluralism, offering a flexible and adaptable solution for the evolving landscape of AI governance. The paper concludes with strategies for resolving conflicts between ethical directives, underscoring the present and future need for robust, nuanced and context-aware AI systems.

cs.CY · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04949

HERIT: Enabling Global Access to Hanja-to-English Historical Translation by Mitigating Temporal Bias

Dayeon Jeong, Kyeongpil Kang

原文摘要

Historical archives are invaluable for multidisciplinary research but remain inaccessible globally due to severe language barriers. While manual translation is prohibitively expensive and time-consuming, simply applying existing machine translation models is inadequate due to unique linguistic nuances and the temporal bias inherent in scarce human-labeled corpora. To address these challenges, we propose HERIT, a systematic domain-adaptation pipeline that leverages retrieval-augmented generation (RAG) to generate high-quality pseudo-labeled data from abundant Hanja-Korean corpora. This approach substantially expands the training dataset, mitigating both data scarcity and temporal overfitting across different historical periods. Extensive automated and human evaluations demonstrate that HERIT consistently outperforms baselines. Finally, we translate previously untranslated portions of the archives using our model, thereby enabling global access to these resources for researchers.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04941

Residual Attribute Leakage Is an Interface: Auditing Privacy Evaluation After Text De-identification

Tuan Thanh Duong, Joongmin Shin, Jeongho Yoon, Junyoung Son, Heuiseok Lim

原文摘要

A privacy evaluation of de-identified text asks a language model (the adversary) to guess a person's protected attribute. We introduce a prior-normalised clipped leakage score of its confidence gain on the true value and audit the conclusions drawn from it. Each verdict depends on an evaluation interface (the queried person, answer options, prior, adversary model and scoring rule), which we vary on fixed releases of three corpora. In a stress test, an editor that writes the opposite-sex form at the same spans a redaction editor would redact misleads the adversary about most people, yet ranks first among editors scored on the same MedQA entries under this score, as under attack accuracy and proper scoring rules read as privacy. A relabelling-invariant reading of AUC ranks that editor sixth of eight, and a rule-aware adversary recovers the attribute from the same text. On ordinary releases most verdict changes are gains or losses of separability; of six choices varied over all pairs, only the adversary reverses comparisons (5 of its 124 changes), and the query can target the wrong person. SACC, a traceable editor we build as the instrument, lets us ablate two editing principles and show where its advantage does not hold.

cs.CL · cs.CR

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04921

PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction

Yimeng Wang, Zeyuan Tian, Xinyu Wang, Danyang Gao, Chenglei Shen, Zhongxiang Sun, Lei La, Weijie Yu

原文摘要

Legal judgment prediction models may retain a guilty verdict after a legally decisive condition has been altered such that the specified charge is no longer established. We study this guilty bias via selective repair over matched charge-established and charge-not-established cases, where the unsupported verdict should change while its supported counterpart remains stable. We construct CAIL-MC, a legally reviewed dataset of 11,478 matched pairs covering 261 Chinese criminal charges. Held-out activation interventions show that contrasts learned from correctly resolved pairs transfer to false-guilty cases, but target-reaching interventions can also damage matched supported-guilty decisions. Motivated by this observation, we introduce PROSE, which reconstructs case-specific targets from a training-derived residual geometry, aligns them with a local margin-descent signal, attenuates supported-guilty-sensitive components, and realizes bounded targets through a preservation-regularized low-rank parameter update. Across Qwen3-8B and GLM4-9B, PROSE achieves stronger pairwise safe repair and evidence sensitivity, while preserving standard single-pass inference.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04903

Heads That Write, Not Just Point: Image Retrieval Heads in Vision-Language Models

Junsung Park, Sangwon Yu, Donghun Kang, Yeongtak Oh, Jaewon Jeong, Gyeongtae Yoo, Han Cheol Moon, Jungbeom Lee 等 9 位作者

原文摘要

Large vision-language models (LVLMs) can process interleaved text and multiple images in a single context, but how they internally identify the image relevant to a question remains poorly understood. We study this grounding step as in-context image retrieval and introduce ICIR-MCQ, a controlled probe for head-level analysis of image retrieval. Using each head's attention mass over image-token spans, we first identify sparse pointer heads that reliably point to the target image. These heads transfer beyond the probe: a single pointer head is competitive with external vision-language retrievers on a naturalistic multi-image retrieval benchmark, showing that LVLMs contain a strong internal image retrieval signal. Attention alone, however, shows where a head reads from, not what its output passes on to later layers, so we introduce a complementary output-side analysis. For each head, we train a lightweight image retriever on its value-weighted output and use it to identify writer heads, whose outputs contain enough information to identify the target image. Surprisingly, the pointer and writer criteria select substantially different head sets, with only 35 to 47 heads overlapping among the top-100 heads under each criterion. Targeted set-partition ablations show that the heads most important for standard multi-image inference are concentrated in this pointer-writer intersection, while heads selected by only one criterion show less consistent separation from random-head controls. The most consistent causal effects arise from heads selected by both the attention-based pointer and output-side writer criteria. Code and data: https://github.com/parkquasar/Image-Retrieval-Heads.

cs.CV · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.42010

TurnTurk-GPT: Text-Based Turn-Taking Prediction for Turkish

Ahmet Tuğrul Bayrak, Fatma Nur Korkmaz, Nur Yağmur Tuncer, Mustafa Sertaç Türkel

原文摘要

This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish conversation against rule baselines, a zero shot Turkish language model and a widely used multilingual end-of-turn detector. Evaluation uses word-level decision points on unpunctuated, lower-cased text approximating the output of streaming speech recognisers, recording-disjoint and speaker-disjoint folds, nested threshold selection and paired cluster bootstrap tests. TurnTurk-GPT outperforms the multilingual detector under both recording-disjoint and speaker-disjoint evaluation. Its gain over a zero-shot Turkish language model is largest for speakers seen in training, while on unseen speakers it matches this strong baseline. Removing the suffixes of the turn-final word reduces AP by up to 44%, against 2% for a sentence-internal word, locating the Turkish end-of-turn cue in verb-final morphology. Synthetic dialogue alone yields the weakest trained model, and when a speaker pauses the zero-shot and multilingual baselines remain close to chance. The model and the evaluation protocol are publicly released.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.04867

One Assistant, Many Memories: Benchmarking Real-World Group Assistance When Recall Is Not Enough

Zhanghao Hu, Linhai Zhang, Qian Zhao, Qi Zhu, Yuan Hua, Di Liang, Jiasheng Si, Xin Zhao 等 11 位作者

原文摘要

The value of memory lies not only in recall but in using it to provide appropriate assistance. Recent benchmarks test memory use in personalisation and agentic tasks, but seldom evaluate assistance without fixed answers or integration across multiple participants. Group chats make both challenges more acute: an assistant must integrate information from multiple members, track revisions and conflicts over time, and produce one response for the group. Yet existing group memory benchmarks rely on synthetic conversations and inherit constrained evaluation from general memory benchmarks, focusing on recall against fixed answers or decisions over a closed set. We introduce GroupAssistBench, the first memory benchmark from real-world group conversations with a deployed group assistant, covering 80,299 messages in 50 groups and costing over $53K. Over the same histories, GroupAssistBench evaluates memory recall (L1) with reference answers and memory use (L2) with service requests without fixed answers across single, distributed, and conflicting memories (M1--M3). Requests are verified to require group history, memories are anchored to their sources, and L2 uses tailored rubrics with penalties derived from real user feedback. Across evaluations of frontier models with full context, retrieval methods, and memory systems, the strongest memory system rivals full context on recall but falls far behind on use, especially for distributed and conflicting memories. Together, two additions that relate retrieved memories and bind remembered facts as response constraints bring consistent gains, showing that group memory systems must improve use, not recall alone.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2605.10379

Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness

Ivo Petrov, Jasper Dekoninck, Dimitar I. Dimitrov, Martin Vechev

原文摘要

Large language models (LLMs) have become capable mathematical problem-solvers, often producing correct proofs for challenging problems. However, correctness alone is not sufficient: mathematical proofs should also be clear, insightful, contain transferable knowledge, and be generally useful to the reader. While proof quality is subjective and depends on the reader and context, many of its components can be made concrete. In this work, we identify such components and introduce ProofRank, a benchmark curated from challenging competitions that measures several scalable proxies of proof quality: (i) conciseness - quantifying whether proofs avoid unnecessary steps; (ii) computational ease - the extent to which a proof relies on tedious calculations; (iii) cognitive simplicity - how accessible the used proof techniques are; (iv) diversity - how varied a model's proofs for a single problem are; and (v) adaptivity - how well a model can follow a specified technique. We find proof quality varies substantially across models along dimensions that correctness largely fails to capture. Every proof-level proxy can be steered through prompting, but often at a moderate correctness cost. ProofRank thus captures important aspects of mathematical proofs beyond correctness, and indicates the need for models that produce proofs which are not only correct but also valuable.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.13447

Oído: Real-Time Open-Vocabulary Speech Recognition on a $5 Microcontroller without a Neural Accelerator

Daniel Varela Sánchez, Guillermo Martín-Coello Juárez

原文摘要

Speech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present Oído, an inference engine and a family of models that run open-vocabulary English speech recognition entirely on an ESP32-S3 (240 MHz Xtensa LX7, 8 MB of PSRAM, 16 MB of flash, no neural accelerator) faster than real time on one core. Version 1 of this work ran a 13 M-parameter Conformer-CTC at 1.97x the duration of the audio. We measured on the board that the bound was memory traffic and per-frame cost, not arithmetic: external memory reads at 89 MB/s (PSRAM) and 32 MB/s (flash), and a streaming recognizer re-reads every weight for every chunk. Version 2 combines a model built for that memory system (80 ms frame stacking, a 64-channel front end, 12 or 16 blocks, int4 weights that mostly fit in PSRAM, fine-tuned from the version 1 model) with engine changes that overlap the weight stream with compute (an asynchronous data-cache preload), keep work buffers in SRAM and make the decoder cheap. On a physical board the streaming real-time factor falls from 1.77 to 0.54 (variant B) and 0.72 (variant A), the final text appears 259 ms (B) or 431 ms (A) after the end of speech with audio fed at real-time speed, and the host build and the firmware produce identical transcripts. Variant A reaches 7.1%/16.4% word error rate on LibriSpeech test-clean/test-other in streaming mode with 0.64 s chunks and the on-chip language model (4.6%/10.7% with full context). Code, models and the raw board measurements are released.

cs.SD · cs.CL · eess.AS

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.02476

GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data

Junzhe Liu, Yujie Tu, Yifan Yang, Peng Wang, Haoran Wang, Guodong Lin, Leyan Liao, Chengquan Zhang 等 16 位作者

原文摘要

Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.

cs.CL

查看论文
arXiv ↗PDF ↗