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

+1新增论文
+1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：2026-11-13 07:07 · 全站最近入库：2026-11-13 06:39各信源进度 →

统计口径与数据新鲜度

统计 2026-11-12 07:50 — 2026-11-13 07:50（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：2026-11-13 07:07最近发现新增版本：2026-11-13 06:39库内论文最新更新时间：2026-11-11 08:00

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。

21,676 已收录论文18,324 AI 相关6,056 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 3,743 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-13 07:05（北京时间）
版本更新 · v3arXiv:2602.10564

SplitCom: Communication-efficient Split Federated Fine-tuning of LLMs via Temporal Compression

Tao Li, Yulin Tang, Yiyang Song, Cong Wu, Xihui Liu, Pan Li, Xianhao Chen

原文摘要

Federated fine-tuning of on-device large language models (LLMs) mitigates privacy concerns by preventing raw data sharing. However, the intensive computational and memory demands pose significant challenges for resource-constrained edge devices. To overcome these limitations, split federated learning (SFL) emerges as a promising solution that partitions the model into lightweight client-side and compute-intensive server-side sub-models, thus offloading the primary training workload to a powerful server. Nevertheless, the high-dimensional activation exchanges in SFL lead to excessive communication overhead. To overcome this, we propose SplitCom, a communication-efficient SFL framework for LLMs that exploits temporal redundancy in activations across consecutive training epochs. Inspired by video compression, the core innovation of our framework lies in selective activation uploading only when a noticeable deviation from previous epochs occurs. To balance communication efficiency and learning performance, we introduce two adaptive threshold control schemes based on 1) bang-bang control or 2) deep deterministic policy gradient (DDPG)-based reinforcement learning. Moreover, we implement dimensionality reduction techniques to alleviate client-side memory requirements. Furthermore, we extend SplitCom to the U-shape architecture, ensuring the server never accesses clients' labels. Extensive simulations and laboratory experiments demonstrate that SplitCom reduces uplink communication costs by up to 98.8\,\% in its standard configuration and total communication costs by up to 95.8\,\% in its U-shape variant without noticeably compromising model performance.

cs.NI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13595

IFRS Compliance Verification: An Open Multi-LLM Benchmark and Content-Conditioned Ensemble Classifier

Aghoghomena Akasukpe, Li Yang, Xinyao Zhou, Chen Shen, Miguel Vargas Martin

原文摘要

Verifying corporate financial disclosures against the International Financial Reporting Standards (IFRS) is a core auditing task in which experts cross-reference disclosures with hundreds of fine-grained regulatory paragraphs: a laborious, costly, and error-prone process that motivates growing interest in large language models (LLMs) as compliance assistants. Yet single-model accuracy remains insufficient for high-stakes use, and naive ensembling is vulnerable to systematic per-model biases that can drag aggregate accuracy below that of the strongest individual annotator. This gap is addressed with a multi-stage framework, supported by a new benchmark of 19,998 IFRS regulation-disclosure pairs comprising 11,161 unanimous multi-LLM labels and 8,837 human annotations on items where the LLMs disagreed. The framework couples a three-stage domain-adapted sentence embedding pipeline (BERT masked language modeling on SEC EDGAR filings, unsupervised TSDAE and SimCSE adaptation, and supervised contrastive alignment) with a learned ensemble classifier that fuses frozen embeddings and LLM probability features from GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Pro via Tree-structured Parzen Estimator (TPE) optimized gradient-boosted trees. On the disagreement set, the ensemble achieves 81.7% weighted F1, exceeding the strongest single LLM by 10.4 points and majority voting by 14.6 points, and outperforms eleven unsupervised aggregation baselines; the retrieval pipeline attains Recall@15 of 45.3% on unseen regulations. The dataset, model checkpoints, annotation platform, and classifier implementation will be publicly released.

cs.CL · cs.IR

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13594

Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth

Xutao Ma, Somayeh Sojoudi

原文摘要

Large Language Models (LLMs) have achieved remarkable progress on complex reasoning tasks, yet the mechanisms by which they acquire reasoning capabilities and the efficiency of their reasoning remain poorly understood. In this paper, we investigate the path-finding problem over directed graphs, a symbolic abstraction of multi-step reasoning. We first characterize the architectural requirements for this task, proving that one-layer transformers require $Ω(N^2)$ size on dense graphs to implement the key DFS child-selection primitive. While a two-layer transformer implements full DFS with $O(N\log(N))$ size. Thus, a second layer is necessary and sufficient for near-linear DFS reasoning. We then extend this expressivity analysis to deeper models: for every $k\geq3$, a $k$-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon $1+2^{k-3}$, establishing an exponential increase in reachability horizon with depth. Finally, going beyond expressivity, given appropriately curated training data, we show that two-layer transformers can provably learn to execute DFS via gradient flow and generalize to unseen graphs. Together, our results provide a mechanistic account of how transformers implement graph search, how depth expands bounded look-ahead, and how such reasoning algorithms can emerge through training.

cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2609.34064

Learning Perturbation Robust Policies for LLM Agents with Stable Optimization

Pengxin Wang, Yuanzhe LI, Yuxin Ren, Huanrui Yang, Jingdi Chen

原文摘要

Reinforcement learning (RL) has become an effective post-training paradigm for long-horizon large language model (LLM) agents. However, we find that the resulting policies can be sensitive to various policy perturbations, such as hidden-state noise, pruning, and quantization. In this work, we study how to improve perturbation robustness during policy optimization. We first introduce the notion of a perturbation robust policy and analyze conditions under which perturbed policy updates preserve stable monotonic improvement. Based on this analysis, we introduce Stable Perturbation-Robust Policy Optimization (SPrPO), which applies adaptive and sensitivity-aware perturbations during RL training. We evaluate SPrPO on ALFWorld and WebShop and conduct systematic experiments across multiple perturbation types and scales, showing improved perturbation robustness while maintaining stable policy optimization.

cs.LG · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.07668

CACHEFORGE: LLM-Guided End-to-End Generative Cache Replacement Policy for Performance and Hardware Efficiency

Kaushal Mhapsekar, Bita Aslrousta, Brijesh Kumar Bhayana, Paula Contreras, Azam Ghanbari, Ethan Goodman, Anna Andriiko, Samira Mirbagher Ajorpaz

原文摘要

Modern cache replacement designs saturate because they operate within fixed representational structures, hand-crafted and heuristic based feature-engineered predictors, or offline imitation models that cannot generate new decision logic on their own. At the same time, replacement is shaped by the causal interaction of prefetching, thrashing, spatial locality, and access-type behavior, producing an enormous design space that is difficult to traverse manually. Prior approaches typically rely on heuristics, parameter tuning, or imitation of an offline optimal policy, capturing correlations rather than synthesizing new mechanisms. As a result, their performance gains often plateau and they overfit under dynamic workload conditions. CACHEFORGE is the first framework to evolve cache-replacement policies end-to-end by embedding a large language model inside a governed hardware-aware loop. In each iteration, the LLM proposes new C++ replacement logic, the policy is evaluated under a trace-based CRC-2 ChampSim simulator, and the framework enforces feasibility through reward shaping, structural checks, dynamic mutation, temperature scheduling, and cross-policy crossover. This closed-loop generation-evolution loop specifically designed for cache replacement policy enables the discovery of compact policies that satisfy hardware constraints while exploring algorithmic transformations beyond fixed predictor structures. Across SPEC CPU2006, CACHEFORGE outperforms all CRC-2 baselines. It improves the total hit rate by 27.36%, 19.69%, 13.72%, 13.15%, 11.83%, and 5.73% over MPPPB, ReD, Hawk-eye, SHiP++, LIME, and LRU, respectively. On memory-intensive workloads, it increases IPC by 10.15%, 7.89%, 6.34%, 3.64%, 3.12%, and 2.71% over LRU, MPPPB, LIME, ReD, SHiP++, and Hawkeye.

cs.AR · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v3arXiv:2608.23018

SplitLite: Low-Rank Residual Compression for Split Learning

Tao Li, Yulin Tang, Qi Guo, Xianhao Chen

原文摘要

Federated fine-tuning of on-device large language models (LLMs) faces a significant computing burden. To overcome this limitation, split learning (SL) has emerged as a promising solution, which offloads the primary training workload to a powerful server. However, SL requires exchanging high-dimensional activations and gradients between clients and the server, resulting in prohibitive communication costs. To overcome this challenge, we propose SplitLite, a communication-efficient split federated LoRA fine-tuning method that exploits the low effective rank structure of consecutive-epoch activation and gradient residuals. Our key finding is that, under low-rank LoRA updates with the pretrained backbone frozen, the activation and gradient residuals of the same data sample between adjacent epochs exhibit low effective ranks, with most of their energy concentrated in a small number of singular directions. By exploiting this property, SplitLite transmits only quantized truncated singular value decomposition (SVD) residual factors, thereby significantly reducing both activation uplink and gradient downlink traffic. Extensive experiments on the GLUE benchmark across a series of advanced on-device LLMs demonstrate that our method reduces activation uplink communication costs by up to 93.5\% and total communication costs by up to 88.6\%, without sacrificing model performance.

cs.LG · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13578

POEM: Output-Relative Block-Sparse Attention for Long-Context Prefilling

Haoyi Wu, Yunkai Dang, Wanyu Zhang, Lei Wang, Wenbin Li

原文摘要

Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.

cs.CR · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13577

TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation

Yunkai Dang, Wanyu Zhang, Meiyi Zhu, Qi Fan, Tianyu Ding, Wenbin Li

原文摘要

Large language models (LLMs) have shown strong capabilities across diverse applications, making both safety and utility essential for reliable deployment. Jailbreak attacks use diverse forms of adversarial prompting to bypass alignment safeguards and induce LLMs to generate harmful responses. Some representation-based defenses regulate internal activations using predefined steering directions or learned safety regions. However, adapting intervention to each input without a predefined representational target while limiting over-refusal remains a challenge. In this paper, we propose Target-free Entropy-guided Residual Modulation (TERM), an inference-time defense that uses prompt-calibrated attention entropy deviations to guide adaptive modulation of the model's own residual updates. Attention patterns vary across prompts and decoding steps, making raw entropy values difficult to interpret as a consistent signal for intervention. We therefore calibrate attention entropy against each prompt's own prefill baseline to establish an input-specific reference for reducing unnecessary intervention on benign requests. Adaptive scale normalization and temporal accumulation then account for differences in deviation scale and persistence to determine intervention strength. The resulting signal controls bounded scaling of the model's own residual updates, without relying on a predefined safety direction. We compare our method with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 models from 0.6B to 32B. Results show our method improves jailbreak robustness while maintaining low over-refusal on benign safety-sensitive requests with modest inference overhead.

cs.CR · cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13562

An evidence-traceable benchmark of clinical reasoning in large language models

Zhiqun Zhao, Xiaodong Yang, Bruce Gu, Ming Li, Yi Fu, Song Li, Qi Mei, Zhi Liu 等 16 位作者

原文摘要

Benchmark scores for medical large language models (LLMs) are difficult to interpret when scoring criteria cannot be traced to clinical evidence. Here we developed Med Evidence Bench (MEB; https://www.medevidencebench.org.cn/), comprising 1,536 physician-constructed items across three levels: guideline knowledge, standardized clinical scenarios and real-world laboratory scenarios. Each criterion stating a medical claim is linked to a versioned source-text span. Across ten LLMs, the Evidence Trace Score (ETS) was 0.975-0.997 for guideline knowledge, 0.292-0.658 for standardized scenarios and 0.182-0.447 for real-world scenarios, and rankings varied across levels. For real-world scenarios, scores for identifying clinical information (mean, 0.770) exceeded those for matching evidence (0.165) or formulating recommendations (0.160). On 1,296 criterion-level judgments, automated grading agreed with the majority vote of three physicians (macro-F1, 0.930; 95% confidence interval, 0.915-0.945), although agreement was lower for evidence application. These findings from Chinese-language materials indicate that guideline recall does not ensure patient-specific evidence use and do not establish clinical effectiveness.

cs.CL

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13561

Discovering Structurally Plausible and Interpretable Cognitive Models with Large Language Models

Boyoung Kwon, Jiyoon Seo, Sang Ho Lee

原文摘要

Computational cognitive modeling has traditionally relied on manual model construction, which is a time-intensive process that requires substantial domain expertise. Although recent LLM-based approaches have begun to automate model generation, they often prioritize predictive accuracy at the expense of structural plausibility and parameter identifiability. We introduce DisCo, an LLM-guided framework that discovers interpretable symbolic cognitive models through evolutionary search with two structural components: a Validator that filters structurally invalid models, and a Cleaner that promotes parameter identifiability by removing redundant parameters. Across three behavioral tasks, DisCo consistently discovers models that achieve performance competitive with task-specific baselines. Ablation analyses show that the Validator is important for recovering psychologically meaningful structure. Without this component, models tend to violate behavioral constraints and show less consistent associations with external cognitive measures. These results suggest that LLM-driven model discovery, when coupled with appropriate structural constraints, can yield cognitively plausible models that go beyond predictive fit.

cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2606.32012

CoMet: Context and Multiplicity Decomposition for Multimodal Uncertainty Estimation

Sanghyuk Chun, William Yang, Amaya Dharmasiri, Olga Russakovsky

原文摘要

Uncertainty estimation has been a long-standing challenge in AI models; it amounts to "knowing what you don't know," and metacognition is notoriously difficult even for humans (cf. the Dunning-Kruger effect). Although it is still far from solved even in simpler classification systems, tackling it in multimodal large language models (MLLMs) is becoming increasingly important. Within MLLMs, uncertainty can stem from any of the diverse sources as well as from their relationships, and further can stem from the unbounded answers in the open-ended setting. To tackle the issues, we propose CoMet, an MLLM uncertainty estimation method by decomposing uncertainty into a context-specific term and a multiplicity-specific term. The former captures ambiguity induced by the given context (e.g., task or prompt), while the latter captures how many plausible answers determined by the context remain compatible with the given input. We train a lightweight post-hoc uncertainty module to estimate these quantities, which enables efficient uncertainty estimation without autoregressive answer generation or repeated sampling. Experiments on various open-ended multimodal benchmarks, hallucination detection, and multiple-choice visual question answering benchmarks show that CoMet consistently improves uncertainty estimation over existing baselines while remaining efficient in practice. Code is available at https://github.com/princetonvisualai/comet_uncertainty

cs.LG · cs.CV

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13550

Tabular Foundation Models in Financial Predictions

Ed deHaan, Allen H. Huang

原文摘要

This technical note introduces tabular foundation models (TFMs) to accounting research and evaluates their performance in one common benchmark task: predicting firms' next-year earnings changes. TFMs produce numeric predictions and learn in context, without post-release training or tuning, making them natural and easy to use for forecasting. The TFMs we study are pretrained on synthetic tabular data, so they avoid look-ahead bias from memorizing firms' outcomes. We compare Google Research's TabFM with a fine-tuned open-weight large language model developed in recent accounting research and with several tree-based models. TabFM is more accurate than all comparison models. It is also faster and less computationally intensive than the large language model. Several tree-based specifications produce predictions more quickly than TabFM, so choosing between them involves a trade-off between accuracy and speed. We conclude that TFMs have promise as a new tool for forecasting in accounting and finance. However, our evaluation is based on a single benchmark task, and no class of models is likely to dominate in all settings.

cs.LG · q-fin.CP · q-fin.ST

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13545

Closing the BF16-FP16 Convergence Speed Gap in RL Post-Training with Stochastic Rounding

Mahdi Nazemi, Oleg Rybakov, Tijmen Blankevoort, Asit Mishra, Carlo del Mundo

原文摘要

BF16 is widely used to pretrain large language models because it has the same dynamic range as FP32 and converges similarly to FP32 training. However, in reinforcement learning with verifiable rewards, where learning rates are typically low, we find that standard mixed-precision BF16 training converges more slowly than FP16 training at the same learning rate. We attribute this gap to weight sticking: when FP32 shadow weights are quantized to BF16 using round-to-nearest, small updates often leave the BF16 weights used for rollout generation and training unchanged. Consequently, several optimizer steps may be needed before updates to the shadow weights are reflected in the BF16 weights. Raising the BF16 learning rate reduces weight sticking and matches FP16's convergence speed, but it also increases the KL divergence from the reference model. We instead apply stochastic rounding when quantizing the FP32 shadow weights to BF16. This approach matches FP16's convergence speed across different models and datasets while reducing the KL divergence relative to BF16 training with a larger learning rate. Unlike FP16, it keeps BF16's wider dynamic range and needs no loss scaling.

cs.LG · stat.ML

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13544

DASWeave: Weaving Evidence across Speaking Conditions for Multimodal Depression, Anxiety, and Stress Screening

Binyan Chen, Tong Pan, Yadong Liu, Xin Yun, Shangfei Wang

原文摘要

Multimodal screening for depression, anxiety, and stress in children and adolescents uses verbal and nonverbal behaviour. Most existing multimodal methods learn from a single recording or interview. Yet a participant's behaviour changes with the speaking condition, and these changes hold evidence that is invisible in a single recording. We propose DASWeave to bring this evidence into a neural network that encodes each response and then combines all responses of a participant. We train supervised trees to find which behavioural statistics matter for screening, both within each response and in how they change across conditions. We feed the tree outputs into the network both before and after this combination. We fine-tune a large language model to relate what a participant says to how they say it across conditions, and feed its representation into the network after the combination. The language model reads behaviour as text descriptions. We therefore derive additional training targets for the network from a multimodal large language model that takes in the audio-visual features directly. On the official hidden test set of the ACM Multimedia 2026 AdoDAS Grand Challenge (Track 1), DASWeave ranks second among 95 eligible teams. Its mean F1 is 27.1% higher than that of the official audio-visual baseline. It also has the highest mean AUROC among the top five teams. Ablation studies on the validation set show that the supervised trees and the fine-tuned language model contribute most to mean F1.

cs.MM

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.11399

RoCoMark: Robust Code Watermarking Against LLM-Based Functionality-Preserving Rewriting Attacks

Hongjie Zhang, Xingcheng Li, Hanzhou Wu, Zhihua Xia, Xinpeng Zhang

原文摘要

Large Language Models (LLMs) present significant challenges to source tracing of generated code, as LLM rewriting can substantially alter the syntactic and structural characteristics of watermarked code. However, existing watermarking methods often suffer severe performance degradation under LLM-based functionality-preserving rewriting attacks. We attribute this vulnerability not merely to the intensity of rewriting, but to a fundamental distribution mismatch between the perturbations encountered during watermark training and the global, non-deterministic rewriting behaviors induced by LLMs. In order to deal with this challenge, we propose RoCoMark, a Robust Code waterMarking framework that explicitly learns and adapts to LLM rewriting behaviors. RoCoMark is a multi-round minimax game between the watermark embedder and LLM rewriter. In each round, the embedder first learns threat profiling from LLM-rewritten code by optimizing the watermark encoder through contrastive learning and knowledge distillation while mitigating catastrophic representation drift. Then the embedder improves strategies by optimizing the watermark encoder and extractor simultaneously. Finally, an adaptive attack-evolution mechanism based on Reflexion and short- and long-term memory pools is applied, enabling continuous interaction between the watermark embedder and evolving rewriting strategies. This iterative attacker-defender process progressively improves the robustness of the watermark against diverse and previously unseen LLM rewriting behaviors. Extensive experiments on four typical benchmarks, under attacks generated by five LLMs, show the effectiveness of RoCoMark. RoCoMark achieves an average detection AUROC of 74.2% and a Rewrite BitACC of 75.3%, while maintaining a code functionality pass rate above 99.7%.

cs.CR

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13531

Framing-Induced Vulnerability in Large Language Models: Evidence from Chinese News Headlines

Mo Zhou, Yiwei Zhang

原文摘要

Large language models increasingly evaluate, route, and moderate news, so a model's judgment of a headline has become a gatekeeping decision. We ask whether that judgment follows the fact being reported or the way the fact is framed. We combine an ecological audit of 63,926 Chinese news headlines with a controlled experiment. The experiment covers 30 real events, 30 synthetic events, and 29 natural headline pairs, tested on four Chinese models. Changing only the headline frame shifted both perceived misleadingness and event severity, and the direction and ordering of the shifts were the same for DeepSeek, MiMo, Dots, and LongCat. Emotionally intensified headlines raised severity ratings by 0.65 to 0.88 points, whereas positively reframed headlines lowered them by 0.34 to 0.52 points. Supplying the article body reduced this severity bias by 54 to 80 percent, and all four reductions were significant at p < .01. A small human reference study showed the same direction of effect. Headline-only LLM evaluation therefore carries a measurable robustness gap, and framing sensitivity belongs among the reliability properties that deployment and governance check.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.10552

Seeing Isn't Enough: Unlocking Visual Reasoning by Bridging the Reasoning-Access Gap with TIGER

Gregory Kang Ruey Lau, Minh Huynh Nguyen, Bryan Kian Hsiang Low

原文摘要

Multimodal Large Language Models (MLLMs) inherit strong textual reasoning capabilities from their language backbones and text pretraining, but often fail to apply the same level of competency to visual inputs. We show that this degradation cannot be attributed to only perception failures, by evaluating models on semantically-equivalent text and rendered-text image inputs: models can accurately read text in the images and recover near-text performance when provided their own transcriptions, but just cannot reason well on the images directly. We characterize this discrepancy as a modality reasoning-access gap, where models fail to route perceived visual content into existing reasoning machinery accessible with text. To bridge this, we propose TIGER (Text-to-Image Gap-targeted training for Enhanced Reasoning), which first mines synthetic modality counterfactuals where a model succeeds on a text problem but fails when provided the corresponding rendered image, and post-trains the model on the image input via Group Relative Policy Optimization (GRPO), along with SFT and DPO variants. Importantly, although trained exclusively on synthetic rendered-text images with no multimodal annotations, TIGER results in model performance improvements across diverse real-world visual benchmarks including geometric diagrams (MathVerse), multi-disciplinary sciences (EMMA, MMMU), complex charts (ChartQAPro), and embodied spatial reasoning (RoboSpatial-Home). TIGER also complements existing state-of-the-art RLVR visual reasoning methods, yielding further performance gains. Analysis involving reasoning-subspace activation and activation patching provide support indicating that TIGER helps to align visual-to-reasoning pathways toward text reasoning states. Our implementation is available at https://github.com/minhngh/Tiger.

cs.AI · cs.CV · cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.10044

CulturalCrosswords: A Multilingual Benchmark of Answer Constraints and Grid Completion

Isaac Saxton-Knight, Gabriel Saxton-Knight, Diego Oliveira, Tidor-Vlad Pricope, Aston Garratt, Jocelyn D'Arcy

原文摘要

When several answers fit a riddle, exact-match scoring can penalise a valid response. CulturalCrosswords tests how answer constraints affect large language models (LLMs) solving cultural riddles, whose solutions draw on knowledge specific to a community. We constructed 53 crosswords from crowdsourced riddles in 42 languages and evaluated 11 LLM configurations on five versions of each clue set: plain batches, batches with lengths, batches with lengths and partial-answer hints, and crosswords requesting an answer list or a completed grid. On 962 clues with alternative answers, lengths improve agreement with the reference by 22.8 percentage points, compared with 8.4 points when accepted alternatives also receive credit. Part of the gain therefore appears to come from selecting a required answer form. Crossword input increases the share of all crossings with compatible answers by 9.7 points; requiring a completed grid reveals large differences between models. The benchmark distinguishes agreement with a reference from the ability to produce answers that fit together and occupy the correct cells.

cs.CL · cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2607.26119

Probing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models

Antyabha Rahman, Akshaj Gurugubelli, Omar Ankit, Kevin Zhu, Aishwarya Balwani

原文摘要

Large language models trained with reinforcement learning (RL) to reason step by step have been increasingly shown to outperform their supervised fine-tuned (SFT) counterparts on mathematical reasoning tasks. Yet the mechanistic basis for this improvement remains unclear. We therefore ask, what internal representational differences enable RL-trained reasoning language models to reason better? Our work presents two primary lines of evidence: First, layer-wise linear probes tend to predict answer correctness more accurately in RL-trained reasoning language models than in instruction-tuned models on synthetic math problems, although output length alone matches the DeepSeek-Math-RL probe on two templates. Second, mean ablation studies suggest that accuracy drop rises with layer depth in DeepSeek-Math-RL but not in DeepSeek-Math-Instruct. Finally, we analyse token-count variability across accuracy bins for RL-trained reasoning language models and instruction-tuned models to measure whether token counts vary more on harder problems. Token counts tend to vary more on lower-accuracy problems, although in DeepSeek-Math-Instruct only with truncated responses included. At high accuracy, we observe lower variability in DeepSeek-Math-RL than in DeepSeek-Math-Instruct, and we observe higher variability in Olmo-3-Think than in Olmo-3-Instruct.

cs.AI · cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2611.10048

Veterans' Mental Health Crisis Severity Annotation: Do LLMs Agree More Than Human Experts?

K M Sajjadul Islam, Pranav Yaramalla, Nadiyah Johnson, Mark Flower, Praveen Madiraju, Zeno Franco

原文摘要

Large language models (LLMs) offer a scalable approach for annotating mental health and crisis-related social media content, but their reliability in safety-critical tasks remains unclear. This study evaluates LLM-based crisis-severity annotation of veteran-related social media posts against annotations from eight Veteran Subject Matter Experts (SMEs). Using 149 tweets and five severity categories, we evaluate 15 open-source and 10 proprietary LLMs across eight prompt variants, examining annotation performance, inter-model agreement, prompt sensitivity, and high-severity crisis detection. Human reviewers showed higher ordinal than nominal agreement (α = 0.494), indicating that disagreements often occurred between neighboring severity levels. In contrast, selected LLMs demonstrated substantially higher inter-model consistency, with ordinal agreement (α) reaching 0.8336 under the baseline prompt. Proprietary models generally showed better agreement with expert annotations and greater stability than open-source models. GPT-5.6-Terra achieved the best overall performance, while the open-source Qwen2.5-7B showed the strongest performance in identifying Acute Crisis with Suicidal Ideation cases. However, higher model consistency did not necessarily correspond to agreement with human judgments or reliable detection of rare high-severity cases. Label-flipped demonstration experiments also revealed model-specific sensitivity to misleading in-context examples, while explanatory demonstrations did not consistently improve performance. Overall, LLMs show promise as scalable complements to expert annotation, but require careful evaluation before use in high-stakes mental health applications.

cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13485

Differentially Private Statistical Inference for Rankings from Pairwise Comparisons

Xintao Xia, Jia Gu

原文摘要

Reliable inference for rankings of large language models based on human pairwise comparisons must preserve the privacy of individual preference records. To meet this need, we develop methods for score and rank inference under the Bradley-Terry model with edge differential privacy. For an objective-perturbed estimator, a new leave-one-out analysis establishes finite-sample bounds on Euclidean and entrywise estimation errors. For sufficiently dense comparison graphs, the error bounds extend to smaller privacy budgets, including regimes in which its privacy term dominates its statistical term. A uniform asymptotic expansion yields asymptotically valid confidence intervals for individual scores and pairwise score differences. To improve finite-sample performance, we account for both regularization bias and privacy-induced randomness by combining a plug-in ridge-bias correction with Gaussian-Laplace calibration. We further establish a bootstrap procedure for simultaneous rank inference, providing asymptotically valid confidence sets for individual ranks and tests of top-$K$ membership. When the comparison probability is public and known, all proposed inferential procedures require no additional privacy budget. We evaluate finite-sample performance through simulations and apply the proposed methods to rank large language models for creative writing.

stat.ML · cs.LG · stat.ME

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13470

An Empirical Study on Improving LLM-Based Issue Classification through Domain-Informed Analysis

Anubhav Gupta, Tanner Wright, Gema Rodríguez-Pérez

原文摘要

Issue reports that describe Intrinsic Bugs, Extrinsic Bugs, and Not a Bug cases often use similar language, but they might point to different causes and require different maintenance actions. This makes three-way issue classification difficult for Large Language Models (LLMs), particularly when prompt-based refinements begin to plateau. We study whether domain-informed analysis of LLM outputs can improve classification beyond prompt modification alone. Using 428 manually labeled issue reports, we evaluate six LLMs across 18 prompt configurations and analyze 7,704 model-generated rationales to identify recurring failure patterns in how models justify predictions, use evidence, and handle conflicting information. We translate these findings into a multi-stage pipeline that separates bug validity from bug origin. The pipeline improves macro F1 and Extrinsic Bug recall for five of the six models. The strongest pipeline configuration reaches 0.827 accuracy and 0.757 Extrinsic Bug recall, improving over the best-accuracy prompt-only baseline by +1.2 and +7.5 pp, respectively. The decomposition further reveals that bug origin classification is more reliable than validity, indicating that the bug-validity decision accounts for most of the remaining error. These results show that expert failure analysis can support more reliable LLM-based software engineering classifiers.

cs.SE

查看论文
arXiv ↗PDF ↗
版本更新 · v4arXiv:2602.02711

Dynamic Mixed-Precision Routing for Efficient Multi-step LLM Interaction

Yuanzhe Li, Jianing Deng, Jingtong Hu, Tianlong Chen, Song Wang, Huanrui Yang

原文摘要

Large language models (LLMs) achieve strong performance in long-horizon agentic tasks through multi-step interaction and reasoning at test time. However, repeatedly invoking these models incurs substantial inference cost over an episode. Post-training quantization (PTQ) reduces this cost, but using a fixed low precision throughout an episode can sharply lower task success. We observe that interaction steps differ sharply in their sensitivity to quantization: given the same context, decoding a step at low precision turns a success into a failure at only about 1% of the steps, which we call critical steps. Based on this observation, we propose Dynamic Mixed-Precision Routing (DMR), a framework that adaptively selects between high-precision and low-precision versions of the same LLM at each decision step. Leveraging an entire-space factorization to jointly predict a step's divergence and its effect on the outcome, DMR learns to recognize these critical steps by training only a lightweight router on the model's hidden state. By reusing a shared high-precision KV cache and the hidden states, DMR avoids redundant context prefill and a separate text encoder, enabling precision switching with less than 1 ms of overhead per step and adding less than 0.11% to step latency for routing. This mixed-precision design also enables nested quantization, reducing the memory footprint by 39% on average compared with FP16. Across three long-horizon agentic benchmarks with nested quantization, DMR achieves an average 2.44x decode speedup over FP16 at comparable task success rates, without fine-tuning the underlying model. On ALFWorld with GPTQ, DMR achieves a 1.71x decode speedup with 86.9% success, compared with 87.2% for FP16.

cs.AI

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.14804

AgentWeave: Efficient Distributed Agent Serving via Flow Decomposition

Kaibin Guo, Pengtu Li, Zicong Hong, Zhiyuan Fang, Wuhui Chen, Rachid Guerraoui, Anne-Marie Kermarrec

原文摘要

Agent applications are evolving into workflows composed of large language model (LLM) calls and tool invocations, termed modules. Existing distributed agent systems typically adopt a disaggregated architecture, deploying different modules on separate devices. While this enables pipeline parallelism, imbalanced module latencies introduce severe pipeline bubbles that limit throughput. A natural alternative is an aggregated architecture that co-deploys all modules across all devices, eliminating bubbles by keeping every device continuously active. However, this introduces request blocking: short requests are forced to progress in lockstep with longer ones in the same batch, and newly arriving requests must wait until the entire ongoing batch completes the full workflow. To address this, we propose AgentWeave, an efficient distributed serving system for agent applications built on the aggregated architecture. Its core mechanism, flow decomposition, decomposes long agent requests into fine-grained execution units while preserving agent semantics, allowing short and newly arriving requests to proceed without being stalled by long-running ones. Complementing this, a flow-adaptive scheduler dynamically balances the throughput gains of decomposition against its scheduling overhead. Experiments across four representative agent applications demonstrate that AgentWeave improves throughput by 1.28-1.81x, reduces average latency by 2.08-2.45x, and reduces P90 latency by 1.56-2.04x over state-of-the-art baselines.

cs.DC

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13452

SLIC: Reinforcement Fine-Tuning Small LMs for Multi-Turn Analog Circuit Optimization

Yan Xu, Zhengqi Gao, Ching-Yun Ko, Zhang-Wei Hong, Wenjie Lu, Tao Yu, Xin Zhang, Duane S. Boning 等 9 位作者

原文摘要

Analog integrated circuit (IC) design is a multi-turn, simulator-grounded engineering task that demands iterative reasoning over coupled performance tradeoffs, yet proprietary process design kits (PDKs) make hosted large language model workflows difficult to deploy in practical design settings. We present SLIC (Small LM Iterating on Circuits), a reinforcement fine-tuning framework that trains a 4B-scale, locally deployable language model for closed-loop analog circuit optimization. SLIC pairs structured circuit representations with discretized relative-update actions, domain- and success-gated multiobjective rewards, and Reward-Guided Rollout Steering (RGRS), which turns per-step simulator feedback into signals for both policy learning and online rollout construction in this sparse, highly constrained domain. We further release an open, goal-conditioned benchmark of 41 amplifier topologies with standardized protocols for in-distribution validation, held-out topology evaluation, and cross-technology-node transfer. Trained solely with open source 65 nm CMOS predictive technology models (PTM), SLIC-Random achieves 87.1% in-distribution four-specification Pass@1 within eight turns, surpassing the 763B DeepSeek-V4.1-Flash Think by 48.5 percentage points. Beyond the training distribution, SLIC-Softmax reaches 72.6% Pass@1 on held-out topologies, and the trained policies consistently transfer across five unseen technology nodes without additional fine-tuning. This transfer extends to a commercial TSMC 65 nm CMOS PDK, where SLIC-Random achieves 49.9% held-out-topology Pass@1, demonstrating practical deployment through local inference without transmitting proprietary data to external LLM services. Performance further scales with test-time simulator interactions, reaching up to 83.2% held-out-topology Pass@1 at a 32-turn budget.

cs.LG

查看论文
arXiv ↗PDF ↗
版本更新 · v2arXiv:2610.02418

Mitigating Private Data Leakage in LLMs with Whiteout

Anna Yoo Jeong Ha, Ronik Bhaskar, Haitao Zheng, Ben Y. Zhao

原文摘要

Modern large language models (LLMs) are trained on massive, largely unfiltered datasets, including content scraped from nearly every accessible website and user inputs. As a result, LLMs often memorize and reproduce personally sensitive information (PSI) such as birth dates, phone numbers, and home addresses. This leads to significant privacy risks, particularly for high-profile individuals such as executives, politicians, and judges. Existing mitigations largely rely on machine unlearning. However, these methods often remove more information than needed, degrade model utility and safety, and are highly vulnerable to attacks. This paper presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine PSIs, by overwriting them using precise and carefully designed obfuscation samples. We evaluate Whiteout on modern LLMs of varying sizes and makers, including a widely-used OpenAI model. Results show that Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, and outperforms existing alternatives. We also test Whiteout against a wide range of countermeasures, from black-box attacks like jailbreaking to white-box adaptive attacks like relearning and quantization. Finally, we conclude with a discussion on the security and ethical implications of Whiteout.

cs.CR · cs.AI

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13426

BioMem: Agentic Memory for Biomedical LLM Agents

Mingchen Li, Yongkang Xiao, Rumeng Li, Rui Zhang, Hong Yu

原文摘要

Large language models (LLMs) have demonstrated considerable potential in biomedical applications, while agentic memory systems have emerged as a promising approach to preserving and utilizing information across tasks and interactions. However, a fundamental question remains: how effectively and reliably do these systems retain, retrieve, update, and utilize biomedical information? In this study, we systematically investigate agentic memory in the biomedical domain, focusing on task effectiveness, retrieval quality, information fidelity, and long-term robustness. We evaluate five memory approaches across four LLMs in three representative settings: biomedical knowledge reasoning, clinical prediction, and long-term conversational memory. Our analyses reveal how memory construction, evidence relevance, and conversational history shape downstream biomedical performance, highlighting the distinct challenges of retrieving relevant information, preserving its fidelity, and effectively utilizing it for reasoning. Building on these findings, we introduce Hybrid Memory Retrieval and Utilization (HMRU), a training-free framework that jointly enhances evidence accessibility and task-aware utilization to improve biomedical reasoning and long-term conversational memory. HMRU achieves an average relative improvement of 8.6% across biomedical question-answering and conversational memory benchmarks, with additional gains in selected clinical prediction tasks. It further reduces conversational answer errors by 24.9% and closes 48.0% of the performance gap to full-context inference. Our findings provide a systematic characterization of biomedical agentic memory and practical insights into developing more effective and reliable memory-augmented LLM systems.

cs.CL

查看论文
arXiv ↗PDF ↗
版本更新 · v5arXiv:2601.21343

Self-Improving Training & Thinking

Ellen Xiaoqing Tan, Jack Lanchantin, Zichun Yu, Polina Kirichenko, Shehzaad Dhuliawala, Danwei Li, Thao Nguyen, Jing Xu 等 14 位作者

原文摘要

Large language models are classically trained in stages: pretraining on raw text followed by post-training for instruction following and reasoning. However, this separation creates a fundamental limitation: many desirable behaviors such as safety, factuality, overall generation quality, and reasoning ability are only added at a late stage, even though the patterns learned earlier strongly shape a model's capabilities. To tackle this issue, we introduce a new way to pretrain and mid-train models that incorporates these behaviors earlier. We utilize an existing strong, post-trained model to both rewrite data and to judge policy model rollouts, thus introducing thinking and reinforcement earlier in training. We further improve reasoning at all stages by introducing the concept of self-improving post-thinking: improving the model's reasoning and self-judging ability by reasoning after generation. In our experiments, we show these approaches can give strong gains across multiple axes.

cs.CL · cs.AI · cs.LG

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13401

SPARK: Semantic-preserving Prompt Mutation for Code Generation using Consensus-based Verification Framework

Alireza Daghigh, Chung-Yu Wang, Luca Filippelli, Hamed Taherkhani, Hadi Hemmati, Hung Viet Pham

原文摘要

Automated code generation with large language models (LLMs) has improved in recent years. With the lack of test cases in real-world usages, generated code is not automatically verifiable and often requires manual validation. Previous studies have leveraged consensus verification, where code solutions and corresponding test cases are generated and evaluated based on functional consistency. Building on this, SPARK incorporates semantic-preserving mutations to problem descriptions such as paraphrasing, grammatical adjustments, function renaming, and other alterations to enhance the robustness and correctness of LLM-generated code. For each mutated description, SPARK generates a set of code solutions and test cases. These candidate solutions are evaluated using consensus-based verification, where the solutions' quality is measured by how consistently they pass generated tests. We assess the effectiveness of SPARK on three benchmarks (LiveCodeBench, BigCodeBench, and DLBench) across three LLMs (o4-mini, DeepSeek-R1-Basic, and Gemini-2.5-Pro). Our evaluation demonstrates that SPARK outperforms state-of-the-art consensus-based verification approaches by up to 4.1%. Our generalization experiment highlights the practical potential of SPARK as an effective strategy for enhancing the reliability and correctness of LLM-based code generation systems in scenarios where traditional validation resources are scarce, while our qualitative analysis shows insight into how SPARK's mutation helps LLMs generate better code solutions.

cs.SE

查看论文
arXiv ↗PDF ↗
新论文 · v1arXiv:2611.13393

Plumbline: When the Loss Lies, Validating LLM Training Correctly

Ruochen Wen, Hongyang Wang, Siwei Li, Wenbo Chen, Xian Li, Jingguo Liu, Leo Fan, Yiming Yan 等 11 位作者

原文摘要

Correctness is a prerequisite for training large language models (LLMs) at scale, yet it is hard to ensure: a fault can sit in any part of a run, and how it shows differs across frameworks and the libraries they depend on. Some failures are loud and some are silent, and both take time to detect and trace. A single fault can waste a long run: the run crashes late, or the model scores lower only at the final evaluation. We present Plumbline, a correctness validation system for LLM training, used in production at Amazon. Plumbline checks each part of the training lifecycle, from the data a run trains on to the model it serves, down to the new layers each model brings and up to whole runs. Its diagnosis reuses these checks to trace failures to their root causes. Both rest on a systematic correctness methodology designed to carry across models and frameworks. Plumbline supports two agentic use cases: Day-0 onboarding, which validates the training of a newly released model before teams rely on it, and root-cause tracing of training failures. Across 3 frameworks and 5 models, it found 61 problems, 32 of them silent. Of these, 42 are fixed or reported upstream, 8 of them first raised by us. We further group the 61 into seven failure patterns that show where training breaks, and report how problems recur across frameworks and models.

cs.DC · cs.LG

查看论文
arXiv ↗PDF ↗