ArXiv Domain 2026-09-16

数据来源：ArXiv Domain

LLM Domain Papers

1. Token Merging for Multilingual Speech Recognition: A Systematic Study Across Model Scale and Fine-Tuning

Abstract:Leading multilingual speech recognition models like Whisper transcribe diverse, low-resource languages without language-specific training but are computationally expensive to deploy. Token merging mitigates this inefficiency by dynamically combining redundant features, shortening the sequence length during inference without requiring retraining. In this paper, we systematically evaluate token merging on the Whisper model family across sixteen diverse languages and three different model sizes. We also test how token merging interacts with fine-tuning (DoRA) on low-resource languages. Our findings show that merging tokens increases computational efficiency with almost no loss in transcription accuracy across most low-resource languages and model sizes, and it works even after the model has been fine-tuned. Our results demonstrate that token merging is a highly practical method for making multilingual speech recognition faster and cheaper to deploy.

中文摘要

摘要：像 Whisper 这样的领先多语言语音识别模型能够转录多种低资源语言，而无需特定语言的训练，但部署成本较高。令牌合并通过动态组合冗余特征缓解了这种低效率，在推理过程中缩短序列长度，无需重新训练。在本文中，我们系统评估了令牌合并在 Whisper 模型家族上的表现，涵盖十六种不同语言和三种不同模型规模。我们还测试了令牌合并与低资源语言微调（DoRA）的相互作用。我们的研究结果表明，在大多数低资源语言和模型规模中，令牌合并在几乎不损失转录准确率的情况下提高了计算效率，并且即便在模型微调之后依然有效。我们的结果显示，令牌合并是一种让多语言语音识别部署更快且成本更低的非常实用的方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13151 (timeout of 15000ms exceeded)

Authors: Dylan Luke Holyoak

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13151.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13151

Published: 2026-09-16T01:38:41.435Z

2. PhysMent: An Interactive Approach For LLM Reasoning In Physics Problems

Abstract:Large language models (LLMs) perform strongly on static science benchmarks, yet their ability to reason about the physical world through active experimentation remains poorly understood. We introduce PhysMent, a benchmark that evaluates LLM physical reasoning via iterative, toolmediated interaction with a MuJoCo physics simulator. Unlike static benchmarks that supply all quantities upfront, PhysMent requires models to discover information by applying forces, querying object states, advancing time, and modifying scene geometry before answering. The benchmark comprises 105 scenes of classical mechanics, organized across four difficulty regimes (Easy/Hard and Single/Multi), three scene modalities (standard, object creation, hidden objects), and a scene-manipulation category, evaluated with a six-dimensional scoring framework. Results show that current models perform reasonably well on qualitative single-concept tasks (up to 80% accuracy) but degrade substantially on quantitative tasks that demand precise, multi-step experimental procedures: most models fall below 30% on the hardest single-concept category, where the bottleneck is procedural (adaptive multi-step tool use) rather than conceptual load. Across the seven models, accuracy ranges from 25% to 67%, with failures due to premature answer submission, inefficient exploration, and inconsistent grounding in simulator feedback rather than conceptual gaps.

中文摘要

摘要：大型语言模型（LLMs）在静态科学基准测试中表现出色，但它们通过主动实验推理物理世界的能力仍然理解不足。我们引入了PhysMent，这是一个通过与MuJoCo物理模拟器进行迭代、工具辅助交互来评估LLM物理推理能力的基准。与预先提供所有量的静态基准不同，PhysMent要求模型通过施加力、查询物体状态、推进时间和修改场景几何来发现信息，然后再作答。该基准包含105个经典力学场景，按四个难度级别（简单/困难与单一/多重）、三种场景模式（标准、物体创建、隐藏物体）以及一个场景操作类别进行组织，并通过六维评分框架进行评估。结果显示，目前的模型在定性单概念任务上表现尚可（准确率高达80%），但在需要精确、多步骤实验程序的定量任务上性能明显下降：大多数模型在最难的单概念类别中准确率低于30%，其瓶颈是程序性的（自适应多步骤工具使用），而非概念负担。在七个模型中，准确率从25%到67%不等，失败原因在于过早提交答案、探索效率低下以及对模拟器反馈的不一致依赖，而非概念方面的缺陷。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13152 (HTTP 429)

Authors: Joseph Chan, Utkarsh Jha, Xiyin Yang, Abhinav Jarajapu, Anik Sahai, Eddie Hu, Robin Jeshua Deepak, Stefano Saravalle, Aditya Shah

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13152.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13152

Published: 2026-09-16T01:38:41.435Z

3. Lexical Prompt Compression for Large Language Models: A Training-Free, Deterministic Pipeline with Empirical Pareto Analysis Across Eleven Task Categories

Abstract:Recent advances in large language models (LLMs) have made prompts increasingly large and complex. Techniques such as chain-of-thought reasoning (Wei et al., 2022) and in-context learning (Brown et al., 2020) frequently push real-world prompts past several thousand tokens, increasing inference cost and latency. Learned compression methods such as LLMLingua (Jiang et al., 2023) and Selective Context (Li et al., 2023) achieve high compression ratios but require auxiliary language models and are non-deterministic. We ask a complementary question: how far can a training-free, fully deterministic, CPU-only pipeline based on classical lexical NLP be pushed before output quality degrades significantly? Eleven toggleable lexical transformations - stopword removal, filler-phrase deletion, contraction and abbreviation substitution, part-of-speech-based pruning, lemmatization, WordNet-driven synonym shortening, and named-entity preservation - are assembled into a configurable pipeline. Fifteen configurations are evaluated on 1,242 English-only prompts from six sources (Dolly-15k, LMSYS-Chat-1M, WildChat-1M, MMLU, GSM8K, HellaSwag), spanning eleven automatically derived task categories, yielding 18,630 paired GPT-4o-mini completions. Output preservation is measured using BLEU, ROUGE-1/2/L, BERTScore-F1, and SentenceBERT cosine similarity. The most aggressive configuration achieves a mean token reduction of 40.3% (sigma = 9.2) at a BERTScore-F1 of 0.876 against the original-prompt output; a stopword-only configuration achieves 29.6% reduction at 0.913. The compression-versus-fidelity Pareto frontier is characterized per task category, with commonsense reasoning a systematic failure mode under aggressive compression. All code, prompts, and per-cell results are released for reproducibility.

中文摘要

摘要：大型语言模型（LLMs）的最新进展使提示变得越来越庞大和复杂。诸如思维链推理（Wei 等，2022）和上下文内学习（Brown 等，2020）等技术，常常使现实世界提示超过数千个令符，增加推理成本和延迟。学习到的压缩方法如LLMLingua（Jiang等，2023）和选择性上下文（Li 等，2023）虽具高压缩比，但需要辅助语言模型且非确定性。我们提出一个补充问题：基于经典词汇自然语言处理的无训练、完全确定性、仅CPU的流水线能被推到多远，输出质量才会显著下降？11种可切换词汇转换——塞字删除、填充短语删除、缩写和缩写替换、基于词性剪枝、词形化、WordNet驱动的同义词缩短和命名实体保存——被组装成一个可配置的流水线。通过来自六个来源（Dolly-15k、LMSYS-Chat-1M、WildChat-1M、MMLU、GSM8K、HellaSwag）的1242个纯英语提示评估15个配置，涵盖11个自动生成任务类别，生成18,630个配对的GPT-40-mini补全。输出保持通过BLEU、ROUGE-1/2/L、BERTScore-F1和SentenceBERT余弦相似度测量。最激进配置在BERTScore-F1为0.876时，平均令牌减少率为40.3%（sigma = 9.2），相较于原始提示输出;仅使用停止字配置则减少29.6%，为0.913。压缩与保真度的帕累托边界按任务类别进行描述，常识推理为激进压缩下的系统性失败模式。所有代码、提示和每单元结果均为可重复性而发布。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13154 (HTTP 429)

Authors: Shamin Chokshi

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13154.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13154

Published: 2026-09-16T01:38:41.435Z

4. TestHallVQA: Exploring LVLMs’ Document-Level Reasoning under Redundant Contexts from Scientific Exams

Abstract:Large Vision—Language Models (LVLMs) are increasingly expected to perform visual question answering (VQA) over planar media. However, existing planar VQA benchmarks typically emphasize isolated challenges: some emphasize long-document understanding with limited reasoning depth, while others require complex visual reasoning but remain restricted to single-page, noise-free settings. Moreover, through theoretical analysis, we identify the impact of irrelevant visual tokens, which leads to measurable performance degradation but has received little attention with respect to systematic quantification. To address these limitations, we introduce TestHallVQA, a multi-image VQA benchmark that simultaneously embodies document-level scale and the difficulty of human examinations, while providing comprehensive task coverage. Leveraging TestHallVQA’s ability to controllably inject multi-level contextual redundancy, we further propose a novel metric, F1-R\textsuperscript{2}, which jointly quantifies LVLMs’ computational reasoning capability and their evidence retrieval robustness against document-level redundancy. Extensive experiments and analyses on mainstream LVLMs reveal their latent deficiencies across multiple dimensions, offering concrete insights and directions for future research. The associated datasets, code, and complete theoretical derivations are available at this https URL.

中文摘要

摘要：大型视觉-语言模型（LVLMs）越来越被期望在平面媒体上执行视觉问答（VQA）任务。然而，现有的平面 VQA 基准通常强调孤立的挑战：有些强调长文档理解但推理深度有限，而另一些则需要复杂的视觉推理，但仍局限于单页、无噪声的环境。此外，通过理论分析，我们识别了无关视觉标记的影响，这会导致可测量的性能下降，但在系统量化方面却很少被关注。为了解决这些限制，我们引入了 TestHallVQA，这是一个多图像 VQA 基准，它同时体现了文档级规模和类似人类考试的难度，并提供全面的任务覆盖。利用 TestHallVQA 可控地注入多层次上下文冗余的能力，我们进一步提出了一种新度量指标 F1-R extsuperscript{2}，它能够联合量化 LVLMs 的计算推理能力及其对文档级冗余的证据检索稳健性。在主流 LVLMs 上进行的大量实验和分析揭示了其在多个维度上的潜在不足，为未来研究提供了具体的见解和方向。相关数据集、代码及完整的理论推导可在此 https URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13158 (HTTP 429)

Authors: Yongqi Yu, Yu Zhang

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13158.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13158

Published: 2026-09-16T01:38:41.435Z

5. Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation

Abstract:Maxillofacial report generation from cone beam computed tomography is scored here by a composite objective placing 80% of its weight on a large language model judgement of factual entailment and 20% on lexical overlap, of which only the lexical fifth is visible during development. The grader’s BLEU-4 and METEOR routines are reproduced in pure Python and match the reference to machine precision, and an offline entailment surrogate, which tells a report written for one patient from one written for another at an area under the curve of 0.987, makes the composite objective cheap enough to optimise directly. Over the 622-case public release, a report selected against the visible lexical ranking scores 0.2909, whereas one selected against the composite objective scores 0.4122, because pursuing n-gram overlap drives entailment precision from 0.522 down to 0.266. A 29 million parameter encoder fine-tuned on the release reaches a prevalence-weighted out-of-fold area under the curve of 0.486 over 985 statements, indistinguishable from the corpus prior, while nine numbers read from the image header reach 0.945 for mandible coverage and 0.872 for condyle coverage, and acquisition centre alone predicts sentence choice at 0.718 against 0.663 for the image-derived model, identifying dictation convention rather than anatomy as the quantity the lexical metrics reward. The delivered system emits eight unconditional statements and five gated on header geometry under polarity, laterality and tooth-level consistency constraints, and reaches METEOR 0.3542 over 50 held-out cases from an unseen centre. The dataset and code are available at this https URL

中文摘要

摘要：来自锥形束计算机断层扫描的颌面报告生成在此通过一个综合目标进行评分，该目标80%的权重用于大型语言模型对事实蕴涵的判断，20%用于词汇重叠，其中开发过程中仅能看到词汇部分的五分之一。评分器的BLEU-4和METEOR程序在纯Python中重现，并与参考结果精确匹配，同时一个离线蕴涵代理可以区分为不同患者撰写的报告，其曲线下面积为0.987，使复合目标足够便宜，可以直接优化。在622例公开数据中，根据可见词汇排名选择的报告得分为0.2909，而根据复合目标选择的报告得分为0.4122，因为追求n-gram重叠会使蕴涵精度从0.522下降到0.266。一个经过公开数据微调的2900万参数编码器在985条陈述上的按流行率加权的交叉验证曲线下面积达到0.486，与语料库先验几乎无异，而从图像头信息读取的九个数字在下颌骨覆盖率和髁突覆盖率上分别达到0.945和0.872，而仅靠获取中心就能在句子选择上达到0.718，对比图像衍生模型的0.663，表明词汇指标奖励的是听写规范而非解剖结构。所交付的系统生成八条无条件陈述，五条在极性、侧别和牙齿一致性约束下的头信息几何条件下生成，并在来自一个未见过中心的50例保留病例上达到METEOR 0.3542。数据集和代码可通过该HTTPS URL获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13238 (HTTP 429)

Authors: Ajo Babu George, Govind Arun, Sidharth N Krishna, Uma Ranjan

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13238.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13238

Published: 2026-09-16T01:38:41.435Z

6. RFCLLM: Evaluating LLMs’ Reasoning Ability of Network Protocol State Machines

Abstract:Mapping textual specifications into formal representations is essential for ensuring the correctness of protocol designs and implementations. LLM-generated mappings, used for networking security or testing, are assumed to capture a perfect understanding of the specification, which may not hold in practice. The goal of this paper is to assess the extent to which LLMs can interpret the specification correctly. We examine the degree to which an LLM’s implicit representation of a finite-state transition system-defined via natural language descriptions-aligns with a manually generated ground-truth model. We designed 4 tasks and 1482 task queries for 16 protocols. We evaluated different judge biases, observed the inherent difficulty gaps between tasks, looked into the effect of 4 context types, and the influence of protocol characteristics. Our work contributes to a step toward verifying whether LLMs can really be trusted in FSM (Finite State Machine) reasoning of protocol specifications.

中文摘要

摘要：将文本规范映射为形式化表示对于确保协议设计和实现的正确性至关重要。用于网络安全或测试的LLM生成的映射，通常假设可以完美理解规范，但实际上可能并非如此。本文的目标是评估LLM在多大程度上可以正确解读规范。我们研究了LLM对通过自然语言描述定义的有限状态转换系统的隐式表示，与人工生成的真实模型的一致程度。我们设计了4个任务和1482个任务查询，涵盖16个协议。我们评估了不同的评判偏差，观察任务之间的固有难度差异，研究了4种上下文类型的影响，以及协议特性的影响。我们的工作为验证LLM在协议规范的有限状态机推理中是否真的值得信赖迈出了一步。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13389 (HTTP 429)

Authors: Anqi Chen, Dan Goldwasser, Cristina Nita-Rotaru

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13389.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13389

Published: 2026-09-16T01:38:41.435Z

7. CVSS-X: A Multilingual Speech-to-Speech Translation Corpus for 28 Languages

Abstract:We introduce CVSS-X, a large-scale synthetic speech-to-speech translation corpus that extends CVSS by reversing the translation direction. While CVSS translates from 21 languages into English, CVSS-X enables translation from English into 28 target languages spanning 12 language families. The corpus comprises approximately 240,000 parallel speech pairs per language, totaling over 16,000 hours, eight times larger than CVSS. We provide two variants: CVSS-X-C with two canonical voices per language, and CVSS-X-T with cross-lingual voice cloning, both fully generated. Evaluation shows comparable translation quality to CVSS with consistent performance across typologically diverse languages. Combined with CVSS, this enables research on bidirectional and multilingual speech-to-speech translation. The code is available at this https URL and the dataset under CC-BY-NC 4.0 license at this https URL.

中文摘要

摘要：我们介绍了 CVSS-X，这是一个大规模的合成语音到语音翻译语料库，通过反向翻译方向扩展了 CVSS。CVSS 将 21 种语言翻译成英文，而 CVSS-X 实现了从英文翻译到跨越 12 个语系的 28 种目标语言。该语料库每种语言包含约 240,000 对平行语音，总计超过 16,000 小时，是 CVSS 的八倍。我们提供了两种版本：CVSS-X-C，每种语言有两种标准音，和 CVSS-X-T，支持跨语言语音克隆，两者均为完全生成。评估表明其翻译质量与 CVSS 相当，并在类型学上多样的语言中表现一致。与 CVSS 结合，可以开展双向和多语言语音到语音翻译的研究。代码可在此 HTTPS URL 获取，数据集在 CC-BY-NC 4.0 许可下的 HTTPS URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13413 (HTTP 429)

Authors: Lucas Rafael Stefanel Gris, Alef Iury Siqueira Ferreira, Frederico Santos de Oliveira, Augusto Seben da Rosa, Alexandre Costa Ferro Filho, Arlindo Rodrigues Galvão Filho, Anderson da Silva Soares

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13413.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13413

Published: 2026-09-16T01:38:41.435Z

8. Causal Analysis and Mitigation of Spurious Onsets in Full-Duplex Speech LLMs

Abstract:Speech-to-speech LLMs like Moshi, and its derivative PersonaPlex, can listen and speak concurrently through full-duplex generation. However, they can begin speaking inappropriately during prolonged user silence: under digital-zero input, Moshi and PersonaPlex initiate speech in 12/40 and 11/40 five-minute continuations, respectively. What causes this spurious speech? We investigate two hypotheses: either repeated sampling selects speech despite persistently low onset probabilities, or conditioning on the model’s nonspeech outputs causes an abrupt spike in onset probability. We find that, at every observed onset, speech probability spikes by over nine orders of magnitude in one 80-ms frame, supporting the latter hypothesis. Then, to suppress these onsets without blocking genuine responses, we ask a causal counterfactual question: is the model responding to user speech, or would its next-token distribution remain similar if the preceding user input were muted? Accordingly, we suppress onsets whose distributions change little under this intervention. Across 40 held-out trials per model with realistic microphone noise, our method suppresses 13/13 Moshi and 9/9 PersonaPlex spurious onsets, while preserving 40/40 genuine responses per model. Our inference-time method requires no retraining and runs in real-time, with 95th-percentile decision time below 61 ms, within the 80-ms frame budget. Our code is available at this https URL.

中文摘要

摘要：像Moshi及其派生模型PersonaPlex这样的语音到语音大语言模型（LLM）能够通过全双工生成同时听和说。然而，在用户长时间沉默时，它们可能会不适当地开始说话：在数字零输入情况下，Moshi和PersonaPlex分别在40个五分钟连续片段中启动了12次和11次讲话。这种虚假讲话的原因是什么？我们研究了两种假设：要么是重复采样在持续低启动概率的情况下仍选择了讲话，要么是对模型非语音输出的条件作用导致启动概率突然飙升。我们发现，在每次观察到的启动中，语音概率在一个80毫秒帧内飙升了九个数量级以上，支持后一种假设。然后，为了在不阻碍真实响应的情况下抑制这些启动，我们提出了一个因果反事实问题：模型是在响应用户语音，还是如果前面的用户输入被静音，它的下一个标记分布仍然相似？因此，我们抑制在这种干预下分布变化很小的启动。在每个模型进行的40个带有真实麦克风噪音的保留试验中，我们的方法抑制了Moshi的13/13次和PersonaPlex的9/9次虚假启动，同时保留了每个模型的40/40次真实响应。我们在推理时的方法无需重新训练并能实时运行，第95百分位的决策时间低于61毫秒，位于80毫秒帧预算范围内。我们的代码可在此https网址获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13445 (HTTP 429)

Authors: Kento Nishi

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13445.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13445

Published: 2026-09-16T01:38:41.435Z

9. Hindsight Bias in Clinical Temporal Reasoning: How Future Data Exposure Affects Large Language Model Judgment

Abstract:Clinical decisions are prospective, but clinical language models are often evaluated on retrospective records that reveal the final diagnosis, treatment response, and outcome. Such evaluations may reward the use of future information rather than reasoning under the uncertainty present at the decision point. We introduce a paired benchmark for measuring outcome-conditioned shifts consistent with hindsight bias in clinical temporal reasoning. It contains 171 case reports from the PubMed Central Open Access Subset—-40 sepsis and 131 GLP-1/diabetes cases—-represented as both textual narratives and human-annotated and LLM-generated textual time series (TTS). For each case, questions are tied to a clinically meaningful cutoff and paired with a prospective reference answer and an outcome-consistent \emph{hindsight trap}. Models answer each question using either a TTS truncated at the cutoff or the complete timeline; additional conditions vary the narrative source (original or synthetic) and TTS annotation source (human or LLM). We evaluate accuracy (Acc), hindsight trap rate (HTR), answer instability rate (AIR), and hindsight bias rate (HBR), each of which captures different signals of hindsight bias. Across GPT 5.6 Sol, Gemma 4, GLM 5.2, and Opus 5, full timeline exposure produces consistent hindsight-sensitive shifts, while temporal masking reduces bias without lowering accuracy.

中文摘要

摘要：临床决策具有前瞻性，但临床语言模型通常基于回顾性记录评估，揭示最终诊断、治疗反应和结局。此类评估可能奖励未来信息的使用，而非在决策点存在的不确定性下进行推理。我们引入了一个配对基准，用于衡量与临床时间推理中事后视角偏差一致的结果条件变化。该基准包含来自PubMed中央开放获取子集—-40例败血症病例和131例GLP-1/糖尿病病例—-既以文本叙述形式，也以人类注释和大型语言模型生成的文本时间序列（TTS）表示。每个案例的问题都与临床有意义的截止线相关联，并与前瞻性参考答案和结果一致的\emph{事后诸葛亮陷阱}进行匹配。模型通过截断TTS或完整时间线回答每个问题;其他条件会改变叙述来源（原始或合成）和TTS注释源（人类或LLM）。我们评估准确率（Acc）、后见明陷阱率（HTR）、答案不稳定性率（AIR）和后见明偏见率（HBR），每种条件捕捉不同的事后视角偏差信号。在GPT 5.6 Sol、Gemma 4、GLM 5.2和Opus 5中，完整时间线暴露产生持续的事后敏感性变化，而时间掩蔽则减少偏见而不降低准确性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13454 (HTTP 429)

Authors: Misaki Matsuura, Sayantan Kumar, Ojas Kadam, Jeremy C. Weiss

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13454.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13454

Published: 2026-09-16T01:38:41.435Z

10. A Hybrid Hierarchical 1D-CNN-BiLSTM Framework for Extractive Summarization of Biomedical and Clinical Text

Abstract:Large language models have made abstractive summarization remarkably fluent, but generated summaries can hallucinate facts, posing serious risks in biomedical and clinical domains. We address this by removing generation from the pipeline and framing summarization as extractive sentence selection. Our Hybrid Hierarchical CNN-LSTM Summarizer uses stacked multi-kernel convolutions to compose sentence-level embeddings into richer inter-sentence representations, followed by a bidirectional LSTM to model long-range dependencies across the document. A lightweight scoring head assigns per-sentence importance scores and is trained end-to-end with binary cross-entropy against oracle extractive labels. At inference, a dynamic mean-plus-standard-deviation threshold with a top-3 fallback selects sentences directly from the source and chronologically reorders them into the final summary. Since every output sentence is copied from the input, the model avoids generation-induced factual drift. On PubMed, our architecture outperforms isolated CNN and LSTM baselines, while ablations show that wider convolutional receptive fields improve sentence scoring. On MIMIC-CXR and MIMIC-IV BHC, the model performs well on unstructured narratives but defaults toward positional baselines on highly templated reports. These results suggest that structural constraints can provide a reliable path toward factually grounded summarization systems that are trustworthy by design rather than by correction.

中文摘要

摘要：大型语言模型使抽象摘要生成变得异常流畅，但生成的摘要可能会产生虚构事实，在生物医学和临床领域带来严重风险。我们通过从流程中移除生成环节，将摘要任务定义为提取式句子选择来应对这一问题。我们的混合层次CNN-LSTM摘要器使用堆叠的多核卷积将句子级嵌入组合成更丰富的句间表示，然后通过双向LSTM建模文档中跨句长程依赖关系。一个轻量级评分头为每个句子分配重要性分数，并通过与黄金提取标签的二元交叉熵进行端到端训练。在推理阶段，使用动态均值加标准差阈值，并在必要时采用前3句回退，从源文本中直接选择句子，并按时间顺序重新排列为最终摘要。由于每个输出句子都直接复制自输入，模型避免了生成引起的事实偏差。在PubMed上，我们的架构优于独立的CNN和LSTM基线，而消融实验显示，增加卷积感受野宽度可以提升句子评分。在MIMIC-CXR和MIMIC-IV BHC数据集上，模型在非结构化叙述文本中表现良好，但在高度模板化的报告中则倾向于位置基线。这些结果表明，结构性约束可以为构建事实依据可靠、通过设计而非修正达到可信的摘要系统提供可靠途径。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13481 (HTTP 429)

Authors: Saad Bin Ather, Muhammad Saif, Ali Hassan Khan, Manzer Abbas, Hajra Waheed

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.13481.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13481

Published: 2026-09-16T01:38:41.435Z

Agent Domain Papers

1. ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

Abstract:In this work, we present ZGCM-1, a fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency. ZGCM-1 is founded on a core premise: compact models cannot passively memorize the open web, but can overcome parametric capacity limits by coupling deliberate internal thinking with active external tool use. To support this paradigm across a 256K context, we develop an end-to-end, high-efficiency open training recipe: Architecture & System Co-design: interleaved gated sliding-window and full attention, and a stable FP8 Muon optimizer; Progressive Curriculum & MDP Mid-Training: context scaling across 16K, 64K, and 256K, and the reformulation of interaction traces into Markov Decision Processes. Furthermore, we establish an AI-native R&D workflow where agent swarms autonomously manage cluster operations, data curation, and rapid diagnostic evaluation. Extensive evaluations show that ZGCM-1-7B is competitive across 7B model family on general benchmarks. On several challenging mathematical reasoning and agentic search suites, it remains competitive with frontier models orders of magnitude larger, such as Qwen3-235B-A22B and GLM-5.1. We also show that our pre-training design offers a ~4.2x efficiency improvement in 16K pre-training time-to-loss. Across the full development lifecycle, we distill eight actionable empirical findings-spanning architectural scaling, SFT quality pruning, long-context generalization, and agentic co-training dynamics. To facilitate community research, we open-source model weights from the pre-training, mid-training, and post-training stages, intermediate checkpoints, training code, per-stage data and data recipes, and W&B logs.

中文摘要

摘要：在本工作中，我们提出了 ZGCM-1，这是一种完全开源的 7B 密集基础模型，从零开始训练，具备极高的数据、系统和算法效率。ZGCM-1 基于一个核心前提：紧凑型模型无法被动地记忆开放网络，但可以通过将有意识的内部思考与主动的外部工具使用相结合，克服参数容量的限制。为了在 256K 上下文中支持这一范式，我们开发了一套端到端、高效的开放训练方案：架构与系统协同设计：交替门控滑动窗口与全注意力机制，以及稳定的 FP8 Muon 优化器；渐进课程与中期 MDP 训练：上下文扩展涵盖 16K、64K 和 256K，并将交互轨迹重新表述为马尔可夫决策过程。此外，我们建立了一个 AI 原生的研发工作流，其中智能体群体自主管理集群操作、数据整理和快速诊断评估。大量评估表明，ZGCM-1-7B 在 7B 模型家族中在通用基准上具有竞争力。在若干具有挑战性的数学推理和智能体搜索套件中，它仍与规模大几个数量级的前沿模型（如 Qwen3-235B-A22B 和 GLM-5.1）保持竞争力。我们还展示了我们的预训练设计在 16K 预训练时间到损失上提供了约 4.2 倍的效率提升。在整个开发生命周期中，我们提炼了八个可操作的经验结论——涵盖架构扩展、SFT 质量剪枝、长上下文泛化以及智能体联合训练动态。为了促进社区研究，我们开源了预训练、中期训练和后期训练阶段的模型权重、中间检查点、训练代码、各阶段数据及数据方案，以及 W&B 日志。

LLM Analysis

Q: 这篇论文试图解决什么问题？该论文试图解决基础模型领域中的两个核心瓶颈： \*\*1. 规模壁垒（The Scale Barrier）\*\* 前沿的数学推理与深度智能体搜索能力长期被视为数百亿乃至千亿参数模型的专属领地，极高的计算需求将资源受限的学术社区排斥在训练与探索之外。论文质疑了“先进智能必须依赖巨大参数量”这一普遍假设，提出并验证了一个核心命题：紧凑型模型（7B规模）虽受限于静态参数容量，但能够通过\*\*审慎的内部长链思考（deliberate internal thinking）\*\*与\*\*主动的外部工具使用（active external tool use）\*\*的结合，突破参数边界，在极具挑战的推理与智能体任务上达到甚至超越数量级更大的前沿模型表现。 \*\*2. 透明度壁垒（The Opacity Barrier）\*\* 当前多数具有竞争力的模型仅以“开放权重”形式发布，而非完全开源。上游的数据过滤配方、中期训练课程、长上下文扩展策略以及多轮智能体交互轨迹等关键训练细节仍属专有黑箱。这阻碍了对训练动态、能力涌现机制以及模型容量极限的系统性学术研究。为消除这一障碍，论文提供了ZGCM-1的端到端完全开放方案，包括： - 预训练、中期训练和后训练各阶段的模型权重 - 中间检查点与训练代码 - 每阶段的数据配方与处理流水线 - 详细的W&B训练日志与评估工具简言之，该工作旨在\*\*在学术可负担的计算预算下，通过系统级的效率优化与“思考+工具”的范式耦合，构建一个完全透明、高性能的7B基础模型\*\*，使广泛的研究社区能够参与并复现前沿级的数学推理与智能体搜索研究。 Q2: 有哪些相关研究？根据论文的文献综述与引用，相关研究可归纳为以下几个方向： ### 1. 大语言模型与推理前沿 - \*\*长程推理模型\*\*：DeepSeek-AI（2026）、Kimi Team（2026）等通过扩展推理链推动了长程思维链（Chain-of-Thought）的边界。 - \*\*工具增强的智能体\*\*：GLM-5-Team（2026）、Qwen Team（2026）、OpenAI（2026）、Anthropic（2026）等探索了在真实环境中利用工具进行深度代理搜索（agentic search）。 - \*\*开放权重模型\*\*：Qwen3系列、MiniCPM4.1、Olmo-3、MiMo等7B–8B规模模型构成了该文直接对比的基线。 ### 2. 高效架构与注意力机制 - \*\*混合与稀疏注意力\*\*：Longformer（Beltagy et al., 2020）、Sparse Transformers（Child et al., 2019）以及 Gemma 3（Gemma Team, 2025）采用的交错局部-全局注意力策略，为本文的 \*\*5:1 门控滑动窗口注意力（Gated SWA）\*\* 提供了基础。 - \*\*注意力变体\*\*：MLA（DeepSeek-AI, 2024）、FlashAttention-2（Dao, 2023）、GQA（Ainslie et al., 2023）以及门控注意力机制（Qiu et al., 2025）。 - \*\*归一化与位置编码\*\*：RMSNorm（Zhang and Sennrich, 2019）、SwiGLU（Shazeer, 2020）、RoPE（Su et al., 2024）及 QK-Norm（Dehghani et al., 2023）。 ### 3. 训练效率与优化 - \*\*低精度训练\*\*：FP8 格式与混合精度训练（Micikevicius et al., 2022）。 - \*\*优化器\*\*：Muon 优化器（Jordan et al., 2024; Liu et al., 2025）在隐藏层参数上的可扩展应用。 - \*\*激活正则化\*\*：TWEO（Liang et al., 2025）用于抑制极端异常值，支撑稳定的 FP8 预训练。 - \*\*开放训练基线\*\*：OLMo 3（Team Olmo, 2025）作为 BF16/AdamW 的效率对比基准。 ### 4. 数据筛选、课程学习与预训练 - \*\*数据混合优化\*\*：RegMix（Liu et al., 2024）、SmolLM2 的数学数据配方（Allal et al., 2025）以及代理模型驱动的混合搜索。 - \*\*课程学习\*\*：Zhang et al.（2026）展示了按难度排序训练样本可提升预训练效率；本文进一步发现\*\*词汇复杂度\*\*对通用语言有效，但对代码与数学不可靠。 - \*\*数学与科学语料\*\*：Llemma / Proof-Pile-2（Azerbayev et al., 2023）、RedPajama（Together Computer, 2023）、olmOCR（Poznanski et al., 2025）、FinePDFs（Kydlíček et al., 2025）及 Nemotron 专用推理数据（NVIDIA, 2025c）。 ### 5. 长上下文建模 - \*\*渐进式上下文扩展\*\*：Qwen2.5-1M（Yang et al., 2025b）采用多阶段长度扩展；本文在此基础上推进至 \*\*16K → 64K → 256K\*\* 的渐进课程。 - \*\*单阶段扩展\*\*：其他技术报告（如 Xiaomi MiMo、Qwen3）选择在单一最终预训练阶段完成扩展。 ### 6. 智能体数据与交互建模 - \*\*智能体持续预训练\*\*：Su et al.（2025）将交互轨迹用于持续预训练；本文进一步将其重构为 \*\*马尔可夫决策过程（MDP）状态-动作监督\*\*。 - \*\*交互范式\*\*：ReAct（Yao et al., 2023）的推理-行动循环被用于深度研究评估。 - \*\*智能体数据治理\*\*：WebDancer（Wu et al., 2025a）等专注于网络信息检索代理。 ### 7. 后训练与强化学习 - \*\*监督微调\*\*：Tulu 3（Lambert et al., 2024）、FLAN（Wei et al., 2022）为指令数据构建提供基础。 - \*\*强化学习算法\*\*：GRPO（Shao et al., 2024）及其带有参考策略 KL 正则化的变体。 - \*\*训练稳定性技术\*\*：动态采样过滤零方差组（Yu et al., 2025）与长度惩罚被用于稳定长程优化。 ### 8. 评估基准 - \*\*数学推理\*\*：MATH-500、AIME 2024/2025/2026、HMMT 2025/2026、AGIEval SAT Math、AQuA-RAT、HARDMath-mini、IMO-AnswerBench、MATH-P-Hard、OlympiadBench、miniF2F。 - \*\*代码生成\*\*：HumanEval+、MBPP+、LiveCodeBench v6。 - \*\*知识与指令\*\*：MMLU、GPQA-Diamond、IFEval。 - \*\*智能体与工具使用\*\*：WebWalkerQA、BrowseComp、GAIA、Binary Function Search（本文自建）、SWE-bench、Terminal-Bench 2.0。 - \*\*抽象推理\*\*：ARC-AGI-1（Chollet, 2019）。 Q3: 论文如何解决这个问题？论文通过\*\*端到端的高效开放训练配方\*\*与\*\*“内部思考+外部工具”的能力范式\*\*，系统性解决了前述的双瓶颈。具体而言，解决方案可归纳为以下四个层面： --- ### 1. 架构与系统协同设计：以极端效率压缩训练成本为在学术级计算预算下完成 7B 模型的全周期训练，论文采用了多层级的效率优化，将预训练时间-损失（time-to-loss）提升约 4.2 倍： | 技术组件 | 具体方案 | 效率收益 | |---------|---------|---------| | \*\*混合注意力\*\* | 以 5:1 比例交错\*\*门控滑动窗口注意力（SWA）\*\*与全局注意力，27 层 SWA（128-token 窗口）+ 5 层全局注意力 | 256K 上下文下吞吐量提升 \*\*3.94×\*\*，KV 缓存减少 \*\*6.4×\*\* | | \*\*低精度训练\*\* | 矩阵乘法采用 FP8（E4M3 前向 / E5M2 后向）配合延迟缩放 | 约 \*\*1.5×\*\* 速度提升 | | \*\*优化器\*\* | 矩阵参数使用 \*\*Muon\*\*（谱缩放 + Newton–Schulz 迭代），标量参数使用 Adam | 约 \*\*1.8×\*\* 步级效率提升 | | \*\*激活正则化\*\* | \*\*TWEO\*\* 抑制中间激活异常值，防止 FP8 数值发散 | 支撑稳定训练，维持约 \*\*60% BF16 等效 MFU\*\* | | \*\*预层归一化\*\* | Pre-LN 架构 | 约 \*\*1.1×\*\* 数据效率提升 | 上述因素的累积效应为：

1.4 × 1.5 × 1.8 × 1.1 ≈ 4.2
—- ### 2. 渐进课程与 MDP 中期训练：激活长上下文与智能体能力为在 256K 长上下文上支持深度推理与工具交互，论文设计了三阶段中期训练（Mid-Training），在 **600B token** 上逐步扩展上下文长度，并重构监督信号： - **上下文渐进扩展**： 16K arrow 64K arrow 256K ，每阶段保留更短序列的回放数据，避免长上下文覆盖导致的常规任务退化。 - **MDP 状态-动作监督**：将多轮智能体交互轨迹重新建模为**马尔可夫决策过程（MDP）**的单步状态-动作转移，使模型在完整轨迹上下文之外，获得密集的局部决策监督。 - **能力导向的数据混合**：代码与数学各占约 20%，随长度增长逐步提升知识（10.5% → 14.2%）与智能体数据（1.5% → 3.3%）的比重。 —- ### 3. 后训练校准：以数据质量与双模式监督释放模型潜力后训练阶段通过精细的数据工程进一步弥补参数规模的局限： - **分层质量筛选（SFT）**：对候选 SFT 数据执行规则过滤与模型质量评分，**剪枝约 50% 的低质量样本**。实验表明，质量优先策略将六基准均值从 67.78 提升至 68.83，证明数据质量优于数量。 - **联合 think/no-think 训练**：故意混合同权重下的显式推理轨迹（think）与直接回答（no-think）样本，使单一模型能根据系统指令动态切换深度思考与高效应答。该联合训练还产生了正向跨模态迁移：结构化推理监督直接提升了直接应答模式下的数学与代码准确率。 - **混合强化学习（RL）**：采用 GRPO 与基于结果的二元/测试通过率奖励，辅以 KL 正则化与长度惩罚，在数学与代码等单领域任务上进一步获得能力增益。 —- ### 4. AI 原生研发与完全开源：消除透明度壁垒论文不仅发布最终模型，而是开放整个研发周期的全部资产，以消除“黑箱”障碍： - **全阶段权重开放**：预训练、中期训练、后训练各阶段权重及中间检查点。 - **训练基础设施开放**：训练代码、配置、每阶段数据配方与处理流水线、W&B 日志。 - **AI 原生 R&D 工作流**：将智能体集群（Agent Swarms）嵌入数据清洗、集群管理、自动实验与原子能力评估（ACE）中，形成可复现、可诊断的迭代闭环。 —- ### 5. 核心范式：用“思考+工具”突破参数容量边界针对规模壁垒的根本解在于**能力范式**的转换：ZGCM-1 不试图被动记忆整个开放网络，而是将自身定位为**具备长程内部推理与主动外部寻求能力的紧凑系统**。在评估中，该 7B 模型通过审慎的 CoT 推理与自主工具调用（网络搜索、终端交互、Ghidra 二进制分析），在 AIME 2026（75.0%）、WebWalkerQA（63.1%）和 Binary Function Search（62.0%）等任务上，达到了数量级更大的前沿模型（如 Qwen3-235B-A22B、GLM-5.1）的同等水平。 Q4: 论文做了哪些实验？论文围绕**架构效率**、**训练动态**、**数据策展**、**能力评估**与**AI原生研发**五个维度展开系统性实验，具体如下： —- ### 1. 架构设计与效率实验为选定生产级注意力方案，论文在固定算力预算下训练多种 7B 规模配置 10B token，对比了全注意力、Flash-GQA、MLA 以及三种滑动窗口注意力（SWA）比例（1:1、3:1、5:1）： - **优化质量 vs 吞吐量权衡**：以最后 50 步平均损失（tail loss）与每 GPU 每秒处理 token 数（tokens/s/GPU）为指标。SWA 5:1 达到最高吞吐量（9,566 tokens/s/GPU），同时 tail loss（1.93）与全注意力基线持平；SWA 3:1 tail loss 最低（1.92）但吞吐量略低；MLA 吞吐量显著下降（7,645）且未改善损失。生产模型最终采用 **SWA 5:1**。 - **长上下文吞吐量扩展**：在 4K 至 256K 序列长度上比较全注意力、SWA 3:1 与 SWA 5:1。SWA 5:1 的加速比随长度增长而扩大，从 4K 的 **1.13×** 提升至 256K 的 **3.94×**。 - **KV 缓存内存分析**：在 256K 上下文、batch=1、bf16 条件下，全注意力需 32.0 GiB，线性注意力基线（GDN 3:1）需 7.0 GiB，而混合 SWA 5:1 仅需 **5.0 GiB**，较全注意力减少 **6.4×**。 —- ### 2. 预训练与中期训练实验 #### 2.1 课程预训练（Curriculum Pre-training）使用 7B 探针模型训练 10B token，对比随机采样与基于**词汇复杂度**的课程排序： - 对非代码、非数学的通用语言文档按词汇复杂度由低到高排序，代码与数学独立交错。 - 结果：代码 BPB 从 1.99 降至 **0.81**，数学 BPB 从 0.97 降至 **0.94**；通用基准（MMLU、ARC、HellaSwag）BPB 上升 0.03–0.09。 - **Finding 1**：词汇复杂度是通用语言数据的有效排序信号，但对代码与数学不可靠，因其表面统计难以反映算法或推理深度。 #### 2.2 数据混合探索使用 **0.3B 参数代理模型**在约 30B token 上迭代搜索最优数据配比，以训练损失与领域评估信号为指导，最终将配方迁移至 7B 目标模型。 #### 2.3 中期训练上下文扩展对比两种达到 256K 的路径： - **直接路径**：在 256K 上下文上训练 30B token，最终损失 1.19。 - **分阶段路径**：先在 64K 上训练 10B token，再在 256K 上训练 20B token，最终损失 **1.16**（略优），故被选为生产方案。 #### 2.4 训练动态监控 - **基础检查点直接评估**：在通用预训练期间保存 54 个检查点（0.04T–4.19T token），直接以基座模型形式评估知识、问答、数学、代码、推理等能力。结果显示各领域能力发展**异质**且非单调，代码与较难数学基准在训练后期仍有提升，而通用知识在 2.5T 后边际收益递减。 - **短 SFT 探针**（Short-SFT Probe）：对中期训练里程碑进行轻量 SFT 后评估。从通用预训练结束到中期训练结束，MATH-500 从 35.83% 升至 **74.12%**，HumanEval+ 从 43.29% 升至 **73.78%**，验证中期训练对下游潜力的显著提升。 —- ### 3. 后训练与数据策展实验 #### 3.1 SFT 数据质量消融从同一基座检查点出发，使用三种过滤强度的数据快速微调： - 最小处理（~2.08M 样本）：六基准均值 **67.78** - 广泛过滤（~1.833M 样本）：均值 67.98 - 质量聚焦过滤（~1.145M 样本，剪枝约 50%）：均值 **68.83** **Finding 4**：在受控比较中，激进的分层过滤（剪枝约半数候选）优于未经筛选的完整语料，表明后训练中**数据质量优于数量**。 #### 3.2 长 CoT 比例校准通过网格搜索与候选运行评估，发现过度增加长链思维（CoT）轨迹比例会引入**冗长偏见**并损害指令遵循。通过动态校准长推理、直接回答与格式指令的比例，找到帕累托最优混合。 **Finding 5**：长 CoT 监督呈现**非单调效用**；过饱和会降解通用指令遵循性能，需动态校准以兼顾推理增益与指令遵循。 #### 3.3 Think/No-Think 联合训练对比仅使用直接回答数据的检查点与混合理性轨迹后的检查点，在 **no-think 模式**下评估： - AIME 2025 从 10.00% 提升至 **43.33%**（+33.33） - AIME 2024 从 6.67% 提升至 **33.33%**（+26.66） - MBPP+ 从 55.03% 提升至 **75.93%**（+20.90） **Finding 7**：中期训练建立的长上下文基础使得在 SFT 阶段**无需昂贵的 256K 长推理轨迹**，即可在 256K 尺度上激活长序列能力。 #### 3.4 通用与智能体数据调度对比对比两种 SFT 调度： - **顺序调度**：先通用数据，后智能体数据。 - **联合调度**：全程交错通用与智能体样本。候选运行评估显示**联合调度**下游性能更优，被用于最终训练。 **Finding 6**：仅使用智能体数据微调会导致交互保真度下降；与通用指令数据**联合训练**为逐步推理与精确约束遵循提供不可或缺的基础。 #### 3.5 混合强化学习实验在数学、代码与通用能力任务上执行 GRPO： - 探索两种采样规模：（a）24 问题 × 16 回复 = 384 轨迹/步；（b）384 问题 × 8 回复 = 3,072 轨迹/步。 - 最大生成长度 65,536 token，总上下文预算 98,304 token。 - 结合结果奖励、KL 正则化与长度惩罚以稳定长程优化。 —- ### 4. 模型评估实验 #### 4.1 7B–8B 规模基准对比在 20 个基准上与 6 个同规模推理模型对比（表 2），涵盖： - **数学与推理**：MATH-500（**97.13%**）、AIME 2024/2025/2026、HMMT 2025/2026、AGIEval SAT Math、AQuA-RAT、HARDMath-mini、IMO-AnswerBench、MATH-P-Hard、OlympiadBench、miniF2F、ARC-AGI-1。 - **代码**：HumanEval+（**90.24%**）、MBPP+、LiveCodeBench v6。 - **知识**：MMLU、GPQA-Diamond。 - **指令遵循**：IFEval。 ZGCM-1-7B 在 14 项推理基准的**平均排名**上位列 7B–8B 规模模型首位。 #### 4.2 智能体评估 - **网络环境深度研究**：在 WebWalkerQA（**63.09%**）、BrowseComp（**19.43%**）与 GAIA text-only（**42.52%**）上评估，使用思考启用的 ReAct 策略，最多 64 步搜索-阅读。 - **二进制函数搜索**（Binary Function Search）：在自建的 50 任务基准上，使用 Ghidra 支持的交互协议。模型需根据行为描述在剥离符号的 ELF 二进制中定位目标函数入口地址。ZGCM-1-7B 达到 **62%** 准确率，远超同规模基线（Qwen3-8B 为 12%），并接近 GLM-5.1（66%）与 Kimi-K2（62%）。 #### 4.3 原子能力评估（ACE）构建包含 **2,503 个探针**、**183 项原子能力**、**18 个类别**的诊断框架。对 ZGCM-7B 与多个参考模型进行对比： - ZGCM-7B 最强项：数学（98.18%）、规划（97.78%）、逻辑（96.08%）。 - 最弱项：指令遵循（49.74%）、真实性（68.91%）、代码（72.24%）。 - 相比同规模 Qwen3-8B-256K，在长上下文（+31.11pp）、算术（+20.74pp）与世界建模（+10.00pp）上优势最大。 —- ### 5. AI 原生研发与自主性评估 #### 5.1 训练系统监控实验在训练过程中，通过持续监控发现：恢复训练后吞吐量在约 80 步内从 585 逐渐下降至 554 TFLOP/s/GPU。实验性诊断表明释放重叠缓冲区不足以消除衰退；最终通过**每 100 次迭代周期性清理 CUDA 分配器缓存与垃圾回收**，恢复并稳定了 585 TFLOP/s/GPU 的吞吐量。 #### 5.2 AI4AI 自主性评级实验 9 名核心贡献者依据 5 级自主性量表（L1–L5）对 11 个 R&D 任务类别进行评级（共 99 个评级），结果： - **L4（高自主性）**：实验与监控、部署工程。 - **L3（条件自主）**：数据清洗、数据获取、合成数据生成、算子设计、训练/推理框架设计、资源平台管理、模型评估。 - **L2（部分自动化）**：模型架构设计、学习算法设计（无任何评级超过 L3）。 **Finding 8**：AI 自主性具有**任务依赖性**；运营类任务可达 L4，而架构与算法设计停留在 L2，表明设计类任务仍高度依赖人类判断。 —- ### 6. 关键效率验证实验论文将 16K 预训练的 time-to-loss 与 OLMo 3 风格的 BF16/AdamW 基线进行估算：

1.4 × 1.5 × 1.8 × 1.1 ≈ 4.2
其中 1.4× 来自 SWA 5:1，1.5× 来自 FP8，1.8× 来自 Muon，1.1× 来自 Pre-LN，验证了**系统-算法协同设计**带来的约 **4.2×** 端到 Q5: 有什么可以进一步探索的点？基于论文的局限性分析与未来展望，可从以下几个层面展开进一步探索： —- ### 1. 架构与系统效率的延伸 - **向稀疏 MoE 架构扩展** 当前 ZGCM-1 采用 7.39B 参数的稠密架构。将混合 SWA、Muon 优化与 MDP 中期训练等配方迁移至稀疏混合专家（Mixture-of-Experts, MoE）主干，可在保持低推理 FLOPs 的前提下扩展参数容量与领域专精能力，验证该效率配方在更大激活参数空间下的可扩展性。 - **超长上下文（1M+）的注意力机制** 论文在 256K 上下文上验证了 SWA 5:1 的高效性。进一步探索在百万级甚至更长上下文中的注意力调度策略（如自适应窗口、动态全局层放置），以及如何维持 KV 缓存的次线性增长，是一个自然的延伸。 - **低精度训练的极限** TWEO + FP8 + 延迟缩放的组合在 7B 规模上实现了稳定训练。未来可探索 FP4/FP6 等更低精度格式，或将其与逐层/逐张量自适应精度策略结合，以进一步压缩训练与推理的显存和计算开销。 —- ### 2. 智能体与交互式学习 - **端到端交互式强化学习** 现有后训练仍主要依赖基于结果的奖励（answer correctness / test pass rate）与 token-level SFT。未来可转向在真实执行沙盒（终端、网页、编译器环境）中进行**多轮、交互式 RL**，使策略直接在环境反馈上优化长程工具调用序列，而非仅依赖离线轨迹监督。 - **自主动态知识检索的统一框架** 当前模型通过显式工具调用来补偿参数知识。下一步可探索将预训练表示与**动态检索子程序**深度融合：当模型检测到参数不确定性（如通过熵阈值或置信度估计）时，自主触发检索、验证与重推理的闭环，而非依赖硬编码的 ReAct 循环。 - **多智能体协作与任务分解** MDP 中期训练目前针对单智能体轨迹。将其扩展至多智能体协作场景（如分布式软件工程、多角色研究代理），建模状态-动作空间中的通信协议与角色分工，是提升复杂系统代理能力的重要方向。 —- ### 3. 数据与训练动态 - **课程学习信号的深化** 论文发现词汇复杂度对代码与数学并非可靠的难度代理。未来需开发**语义级或执行级难度度量**（如程序分支复杂度、证明长度、运行时资源消耗），以构建对技术领域更优的课程排序。 - **SFT 数据筛选的自动化** 当前质量筛选依赖规则与模型评分的混合管线。可探索基于**下游任务反馈的自适应数据选择**（如通过影响力函数或课程采样），使数据筛选本身成为一个可学习的优化问题，进一步压缩后训练成本。 - **智能体数据的规模与多样性** 智能体数据仅占中期训练的 1.5%–3.3%。系统性地扩大高质量交互轨迹的覆盖范围（如跨操作系统、跨编程语言、跨自然语言界面），并研究其对模型涌现长程规划能力的影响，具有显著潜力。 —- ### 4. AI 原生研发（AI4AI）的自主性提升 - **从 L2/L3 到 L4/L5 的设计任务自主化** 论文评估显示，模型架构设计与学习算法设计仍停留在 L2，且从未超过 L3。未来需探索如何让智能体具备**假设生成与算法发明**能力，例如通过自动化的符号回归、超参数搜索空间的形式化表示，以及跨阶段目标协调机制。 - **自我演进的合成环境设计** 将 AI 原生研发从“集群监控 + 数据清洗”扩展到**自动构建对抗性评估环境**与**生成合成训练任务**，形成闭环的自我改进系统。这包括自动发现模型弱点、针对性地生成增强数据、并验证修复效果。 - **可解释的失败模式诊断** Atomic Capability Evaluation（ACE）提供了细粒度诊断。下一步可让智能体基于 ACE 结果自主定位缺陷根因（如“注意力机制导致长上下文位置偏差”或“某类工具调用格式欠拟合”），并自动生成修复方案（数据重采样、架构微调或 RL 奖励重塑）。 —- ### 5. 评估与安全 - **闭卷知识与事实召回的极限** 7B 稠密模型在纯闭卷、高召回需求任务上仍受限于参数容量。未来需量化**参数知识边界**与**外部检索补偿**的精确权衡曲线，并开发更高效的参数-外部记忆混合架构。 - **推理冗长与指令遵循的帕累托前沿** 论文观察到重推理监督会轻微损害严格指令遵循。进一步探索**推理预算的动态分配机制**（如基于任务复杂度自适应思考深度），以及在 RL 阶段引入指令遵循与推理质量的多目标优化，可改善此权衡。 - **工具协议鲁棒性** 当前智能体对工具格式偏差与观察噪声较为脆弱。研究**格式容错推理**（如在不完美 JSON 或截断输出下恢复）与**延迟/噪声环境下的鲁棒决策**，对真实世界部署至关重要。 Q6: 总结一下论文的主要内容该论文提出了 **ZGCM-1**，一个完全开源的 7.39B 稠密基础模型，专为数学推理与智能体搜索任务设计，支持长达 256K 的上下文。以下是论文的主要内容总结： —- ### 1. 研究动机与核心论题论文针对基础模型发展的两大壁垒展开工作： - **规模壁垒**：前沿推理与深度智能体搜索长期被默认为百亿乃至千亿参数模型的专属领域，高昂计算成本将学术社区排除在外。 - **透明度壁垒**：主流开放权重模型并未公开训练配方、数据策展策略与中间过程，阻碍了系统性的基础研究。论文的核心论题为：**紧凑型模型虽受限于静态参数容量，但可通过“审慎的内部长链思考”与“主动的外部工具使用”之耦合，突破参数边界**，在不依赖被动记忆开放网络的前提下，弥补知识缺口。 —- ### 2. 端到端高效训练配方 #### 2.1 架构与系统协同设计 - **混合注意力机制**：以 5:1 比例交错**门控滑动窗口注意力（SWA，128-token 窗口）**与全局注意力（27 层 SWA + 5 层全局）。在 256K 上下文下，KV 缓存较全注意力减少 **6.4×**，推理吞吐量提升 **3.94×**。 - **低精度与稳定训练**：采用 FP8（E4M3 前向 / E5M2 后向）混合精度，结合 **TWEO** 激活正则化抑制异常值，并配合 **Muon 优化器**（矩阵参数）与 Adam（标量参数）。 - **效率增益**：上述设计与 Pre-LN 架构协同，在 16K 预训练中实现相对 BF16/AdamW 基线约 **4.2×** 的 time-to-loss 加速：

1.4 × 1.5 × 1.8 × 1.1 ≈ 4.2

2.2 渐进课程与 MDP 中期训练 - 上下文扩展：中期训练以 600B token 分三阶段扩展序列长度： 16K arrow 64K arrow 256K ，每阶段保留更短序列的回放数据。 - MDP 监督：将多轮智能体交互轨迹重构为马尔可夫决策过程（MDP）风格的状态-动作转移，提供密集的逐步决策监督，而非仅复现完整轨迹。 #### 2.3 后训练校准 - 分层 SFT 数据筛选：通过规则过滤与模型质量评分对候选数据执行约 50% 的激进剪枝，实证表明质量优于数量。 - 联合 Think/No-Think 训练：在 SFT 中混合同权重的显式推理轨迹（think）与直接回答（no-think），使模型能根据指令动态切换模式，且结构化推理监督对直接回答产生正向迁移。 - 混合强化学习：采用 GRPO 结合结果奖励、KL 正则化与长度惩罚，在数学与代码等单领域任务上进一步增益。 —- ### 3. 实验与评估结果 #### 3.1 与同规模模型对比（7B–8B）在 20 项标准基准的广泛评估中，ZGCM-1-7B 在 14 项推理基准上平均排名第一： - 数学推理：MATH-500 97.13%，AIME 2026 75.00%，HMMT 2025 70.42% - 代码生成：HumanEval+ 90.24%，MBPP+ 63.23% - 知识：MMLU 73.88%，GPQA-Diamond 47.87% #### 3.2 与前沿大模型对比（智能体任务）通过结合深度推理与自主工具调用，7B 模型在智能体任务上可与数量级更大的系统竞争： - WebWalkerQA：63.09% - BrowseComp：19.43% - Binary Function Search（自建基准）：62.00% 后两项接近或持平 GLM-5.1、Kimi-K2、Claude 4 Sonnet 等大规模模型。 —- ### 4. AI 原生研发（AI-Native R&D）论文构建了一套由研究者指导、智能体集群（Agent Swarms）执行的 R&D 工作流，覆盖数据清洗、集群管理、自动实验、原子能力评估（ACE）与部署。对 11 类 R&D 任务的自主性评估显示： - L4（高自主）：实验监控、部署工程 - L3（条件自主）：数据清洗、资源管理、模型评估等 - L2（部分自动化）：模型架构设计、学习算法设计（暂无评级超过 L3） —- ### 5. 八大经验发现（Findings）论文在全生命周期中提炼了八条可操作的实证发现，涵盖： 1. 词汇复杂度课程对通用语言有效，但对代码与数学不可靠； 2. TWEO 对 FP8 训练稳定性至关重要； 3. 系统-算法协同设计带来约 4.2× 预训练效率提升； 4. 后训练中数据质量优于数据数量； 5. 长 CoT 监督存在非单调效用，需动态校准； 6. 智能体训练必须与通用指令数据联合进行； 7. 中期训练建立的长上下文基础可规避超长的后训练轨迹需求； 8. AI 自主性高度依赖任务类型，设计与运营任务差距显著。 —- ### 6. 开放科学贡献为消除透明度壁垒，论文公开了： - 预训练、中期训练、后训练全阶段模型权重及中间检查点 - 训练代码、配置与超参数 - 每阶段数据配方、处理流水线与 W&B 日志 - 评估工具链与自建基准（Binary Function Search） Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Authors: Jiyan He, Guang Liang, Hao Liu, Haoxiang Guan, Jinbo Sun, Junyi Guo, Wenjun Feng, Yantai Xie, Yifei Shen, Bin Shao, Chuyang Wei, Kai Chen, Kexin Zhou, Minghang Zhu, Shuxin Zheng, Tie-Yan Liu, Taine Zhao, Wenhui Zhu, Xueyin Xu, Xiaoqing Zhang, Yatao Li, Yuxuan Ren

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13356.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13356

Published: 2026-09-16T01:39:37.206Z

2. Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation

Abstract:Multi-objective Bayesian optimisation (MOBO) is a sample-efficient approach for optimising expensive black-box functions with multiple objectives. In MOBO, the goal is to adequately approximate the Pareto front; that is, to obtain a high-quality solution set with 1) good convergence (closeness to the Pareto front) and 2) good diversity (spread across the Pareto front). Existing MOBO methods typically aim to accomplish these two tasks simultaneously, i.e., driving the search towards the Pareto front while maintaining a diverse set of nondominated solutions, such that the solutions, ideally, can gradually approach the entire front. When sufficient search budgets are available, this approach is effective. However, considering both convergence and diversity throughout the search is not easy and requires careful design. Under very tight budgets, there may not be enough solutions generated to be able to simultaneously approach the entire Pareto front. To address this issue, this paper proposes a \textit{converge-then-diversify} (CTD) approach that decouples convergence and diversity into two stages. In the first stage, CTD focuses on convergence, aiming to quickly drive the search toward a single point on the Pareto front. In the second stage, CTD focuses on diversity, aiming to spread solutions across the front. We present two simple instantiations of CTD by using widely adopted acquisition functions in the area. Experimental results show that, across all 446 pairwise comparisons, CTD statistically outperforms state-of-the-art methods in 72.9\% of the cases, performs equivalently in 21.1\%, and is statistically worse in only 6.1\%, with the advantage being particularly evident in settings with very tight evaluation budgets or in high-dimensional problems.

中文摘要

摘要：多目标贝叶斯优化（MOBO）是一种样本高效的方法，用于优化具有多个目标的昂贵黑箱函数。在 MOBO 中，目标是充分逼近帕累托前沿；也就是说，获得一个高质量的解集，其特征为 1) 良好的收敛性（接近帕累托前沿）和 2) 良好的多样性（在帕累托前沿上的分布广泛）。现有的 MOBO 方法通常旨在同时完成这两个任务，即在保持非支配解集多样性的同时将搜索引向帕累托前沿，使解理想情况下能够逐步接近整个前沿。当搜索预算充足时，这种方法是有效的。然而，在整个搜索过程中同时考虑收敛性和多样性并不容易，需要精心设计。在预算非常紧张的情况下，可能生成的解数量不足以同时逼近整个帕累托前沿。为了解决这一问题，本文提出了一种“先收敛后多样化”（CTD）方法，将收敛性和多样性分解为两个阶段。在第一阶段，CTD 侧重于收敛，目标是快速将搜索引向帕累托前沿的单一点。在第二阶段，CTD 侧重于多样性，目标是在前沿上分布解。我们通过使用该领域广泛采用的采集函数，提出了 CTD 的两个简单实例。实验结果表明，在所有 446 次成对比较中，CTD 在 72.9% 的情况下在统计上优于最先进的方法，21.1% 的情况下表现相当，仅在 6.1% 的情况下统计上较差，其优势在评估预算非常紧张或高维问题的情况下尤为明显。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13396 (HTTP 429)

Authors: Chao Jiang, Yueling Huang, Miqing Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13396.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13396

Published: 2026-09-16T01:39:37.206Z

3. Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement

Abstract:When we speak of recursive self-improvement (RSI), are we speaking of a phenomenon, a mechanism, or a prospect? Towards autonomous and evolving intelligence, RSI is being claimed at many scales, while no single framework that formally describes these emerging instances exists. Its counterpart in the classical realm, iterative policy improvement, is characterized by generalized policy iteration (GPI), a framework of broad applicability with well-understood theoretical properties, but only where the update principle and the evaluation base lie outside the agent. In this paper, we propose Generalized Agent Iteration (GAI), a formal framework that describes iterative policy improvement and RSI as two cases of a single learning paradigm. GAI defines the agent as a configuration of modifiable components within a system and models the learning process as a cycle of agent evaluation and agent improvement. Two pivotal dials then distinguish the instances: whether the improving mechanism is part of the agent and whether the standard it is measured against is grounded outside it. The former dial delineates the boundary between GPI and RSI, and the latter determines a system’s polarity as anchored, goal drift, or fully self-referential. Moreover, we use these coordinates to place existing systems on the same two axes and make the defects of recursive self-improvement statable one condition at a time. We see this paper as a first step toward exploring a formal characterization of RSI that rests on the classical account, makes existing systems comparable, and provides a principled basis for analyzing and designing new ones.

中文摘要

摘要：当我们谈论递归自我改进（RSI）时，我们是在谈论一种现象、一种机制，还是一种前景？迈向自主和不断进化的智能，RSI 在许多层面上被提出，但尚不存在一个能够正式描述这些新兴实例的统一框架。其在经典领域的对应物——迭代策略改进，则以广义策略迭代（GPI）为特征，这是一种具有广泛适用性的框架，并且其理论性质已被充分理解，但前提是更新原理和评估基础位于智能体之外。在本文中，我们提出了广义智能体迭代（GAI），这是一种正式框架，将迭代策略改进和 RSI 描述为单一学习范式的两种情况。GAI 将智能体定义为系统内可修改组件的配置，并将学习过程建模为智能体评估与智能体改进的循环。两个关键旋钮用于区分不同实例：改进机制是否是智能体的一部分，以及其所依据的标准是否建立在智能体之外。前者旋钮划定了 GPI 与 RSI 的边界，后者则决定系统的极性是锚定型、目标漂移型，还是完全自指型。此外，我们使用这些坐标将现有系统置于同一两轴上，并以逐步条件的方式陈述递归自我改进的缺陷。我们将本文视为探索 RSI 正式表征的第一步，这一表征基于经典理论，使现有系统可比，并为分析和设计新系统提供了原则性基础。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13406 (HTTP 429)

Authors: Hongyao Tang, Yi Ma, Pengyi Li, Yifu Yuan

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13406.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13406

Published: 2026-09-16T01:39:37.206Z

4. Vibe Patenting: Evaluating LLM Judges for Professional Patent-Drafting Agents

Abstract:LLM judges are increasingly used to evaluate and improve AI-generated outputs, yet their reliability for complex professional work remains unclear. We study this problem through Vibe Patenting, an end-to-end patent-drafting testbed for AI-agent evaluation. A separately-invoked LLM judge evaluates generated patent drafts and provides structured feedback for iterative revision. Across multiple inventions and drafting-agent configurations, judge-guided revision consistently improves judge-assessed quality, while unguided revision tends to saturate. Notably, iterative judge feedback enables a low-reasoning agent to approach the performance of a substantially more expensive high-reasoning agent. Stronger models and increased reasoning generally improve judge-assessed drafting quality, while domain-specific agentic workflows provide further gains. We validate the judge against independent evaluation by a professional patent attorney and find meaningful but strongly metric-dependent agreement and systematic calibration differences. These results highlight both the utility and limitations of LLM judges as evaluators and optimization signals for complex professional workflows.

中文摘要

摘要：大型语言模型（LLM）评审者越来越多地被用于评估和改进 AI 生成的输出，但它们在复杂专业工作中的可靠性仍不明确。我们通过 Vibe Patenting 研究这一问题，这是一种用于 AI 代理评估的端到端专利撰写测试平台。单独调用的 LLM 评审者会评估生成的专利草案，并提供结构化反馈以进行迭代修改。在多个发明和撰写代理配置中，评审者指导的修改始终提高了评审者评估的质量，而未指导的修改往往会达到饱和。值得注意的是，迭代的评审反馈使推理能力低的代理能够接近性能显著更高成本的高推理能力代理。更强大的模型和更高的推理能力通常会提高评审者评估的撰写质量，而特定领域的代理工作流程则带来进一步提升。我们将评审者的评估结果与专业专利律师的独立评估进行验证，发现存在有意义但高度依赖指标的一致性以及系统性的校准差异。这些结果突出了 LLM 评审者作为复杂专业工作流程的评估者和优化信号的实用性与局限性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13422 (HTTP 429)

Authors: Toshiaki Koike-Akino, Vlad Blaykhman, Ye Wang, Jing Liu, Gene V. Vinokur

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13422.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13422

Published: 2026-09-16T01:39:37.206Z

5. Toward Self-Adaptive Physical AI: Can LLM Agents Manage Long-Horizon Physical Tasks?

Abstract:Large Language Model (LLM) agents offer a promising path toward autonomously managing long-term physical tasks without human intervention. However, physical tasks require agents to continuously observe the environment, make consequential actions, and remain effective as the environment changes. Existing approaches either require substantial data and retraining, or primarily focus on agents operating in the virtual world. In this work, we explore the feasibility of building a self-adaptive physical AI agent that manages long-term physical tasks in a zero-shot manner and adapts to environmental changes without human intervention. We design a multi-agent framework that integrates planning, tool calling, observation, and verification, and evaluate it on agricultural tasks against reinforcement learning (RL) agents under different weather patterns. Our results show that zero-shot LLM agents can achieve comparable management outcomes to RL agents under the same weather pattern and adapt more effectively than RL when evaluated under a shifted environment, highlighting a promising path toward self-adaptive physical AI agents.

中文摘要

摘要：大型语言模型（LLM）智能体为自主管理长期物理任务而无需人类干预提供了一个有前景的途径。然而，物理任务要求智能体持续观察环境、执行重要操作，并在环境变化时保持有效性。现有方法要么需要大量数据和重新训练，要么主要关注智能体在虚拟世界中的操作。在本研究中，我们探讨了构建一个自适应物理 AI 智能体的可行性，该智能体能够以零样本方式管理长期物理任务，并在无需人类干预的情况下适应环境变化。我们设计了一个多智能体框架，整合了规划、工具调用、观察和验证，并在农业任务中对其进行评估，同时与强化学习（RL）智能体在不同天气模式下进行比较。我们的结果显示，零样本 LLM 智能体在相同天气模式下可以实现与 RL 智能体相当的管理效果，并且在环境变化时比 RL 更有效地适应，这突显了实现自适应物理 AI 智能体的有前景途径。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13436 (HTTP 429)

Authors: Varun Kaushik, Yayun Tan, Xiaofan Yu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13436.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13436

Published: 2026-09-16T01:39:37.206Z

6. LabAgent: Customize Any Research Hubs for Scientific Discoveries Using AI Agents

Abstract:Scientific research is a continuous process that emphasizes inheritance. Methods developed by predecessors are often expanded upon by new researchers to explore more novel and in-depth scientific questions. However, the change of lab staff, such as student graduation, leads to a lack of personnel capable of replicating methods. Methods that have been developed with significant effort and resources cannot be continued. To address these limitations, we propose LabAgent, a reproduce and discovery harness tailored for a lab’s continuous work. LabAgent employs two mechanisms to guarantee that all skills can be executed and verified and to record the corrective methods and experiences, allowing for direct correction or avoidance of similar errors. We applied LabAgent to drug property prediction, biomedical problem analysis, protein variant effect prediction, and statistical genetics in life science domains. LabAgent ranks first over commercial generalist agents in every domain, and demonstrates accurate reproduction of a published figure. Overall, these results demonstrate that LabAgent can effectively integrate and reasonably expand laboratory knowledge.

中文摘要

摘要：科学研究是一个强调传承的连续过程。前人开发的方法通常会被新的研究人员加以扩展，以探索更新颖、更深入的科学问题。然而，实验室人员的变化，例如学生毕业，会导致缺乏能够复现方法的人员。那些投入了大量努力和资源开发的方法可能无法继续使用。为了解决这些限制，我们提出了LabAgent，一种为实验室持续工作量身定制的复现与发现工具。LabAgent采用两种机制来保证所有技能的执行和验证，并记录修正方法和经验，从而可以直接纠正或避免类似错误。我们将LabAgent应用于药物性质预测、生物医学问题分析、蛋白质变体效应预测以及生命科学领域的统计遗传学。LabAgent在每个领域均排名优于商业通用代理，并展示了已发表图表的准确复现。总体而言，这些结果表明LabAgent可以有效整合并合理扩展实验室知识。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13437 (HTTP 429)

Authors: Lei Liu, Yikun Zhang, Jialin Chen, Wanjia Zhao, Rex Ying, Wengong Jin, Hua Xu, James Zou, Tianyu Liu, Hongyu Zhao

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13437.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13437

Published: 2026-09-16T01:39:37.206Z

7. TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models

Abstract:Timeseries multimodal large language models (TS-MLLMs) have recently begun leveraging the reasoning capabilities of large language models (LLMs) for question-answering tasks. However, these models often fail to capture dynamic temporal patterns, providing only implicit reasoning that lacks the underlying explanations critical for high-stakes applications like healthcare. While reinforcement learning (RL)-based timeseries language models aim to address this, they often fall short because they are trained on narrow, in-distribution data and struggle with out-of-distribution compositional questions. To address these challenges, we present TimeThink, a synthetic framework for eliciting compositional timeseries reasoning. Core timeseries primitives (e.g., trend, seasonality) are domain-independent and can be deterministically generated. Guided by this premise, TimeThink first designs a synthetic data generator that produces atomic and composite question-answer pairs, providing objective ground truth with reasoning traces. Building on this framework, TimeThink employs a reinforcement learning with verifiable rewards (RLVR) training strategy that encourages explicit reasoning. Unlike template-reliant methods, this approach enables the model to learn the underlying logic of composition rather than simply imitating traces. Extensive experiments show that TimeThink, trained only on synthetic data, significantly outperforms strong baselines on both synthetic and real-world benchmarks.

中文摘要

摘要：时间序列多模态大型语言模型（TS-MLLMs）近年来开始利用大型语言模型（LLMs）的推理能力来执行问答任务。然而，这些模型常常无法捕捉动态的时间模式，只提供缺乏关键解释的隐式推理，而这些解释对于医疗等高风险应用至关重要。虽然基于强化学习（RL）的时间序列语言模型旨在解决这一问题，但它们往往表现不佳，因为它们仅在有限的分布内数据上进行训练，并且在分布外的组合性问题上存在困难。为了应对这些挑战，我们提出了TimeThink，这是一个用于引导组合性时间序列推理的合成框架。核心时间序列原语（例如趋势、季节性）是领域无关的，并且可以确定性地生成。基于这一前提，TimeThink首先设计了一个合成数据生成器，生成原子和复合的问答对，并提供具有推理轨迹的客观真实数据。在此框架的基础上，TimeThink采用可验证奖励的强化学习（RLVR）训练策略，以鼓励显式推理。与依赖模板的方法不同，该方法使模型能够学习组合的底层逻辑，而不仅仅是模仿推理轨迹。大量实验表明，仅在合成数据上训练的TimeThink，在合成和真实世界基准测试中均显著优于强基线模型。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13457 (HTTP 429)

Authors: Sudarshan Regmi, Arvind Pillai, Yu Yvonne Wu, Yuliang Chen, Bibek Panthi, Tess Z. Griffin, Michael V. Heinz, Lisa Marsch, Nicholas C. Jacobson, Andrew Campbell

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13457.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13457

Published: 2026-09-16T01:39:37.206Z

8. Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures

Abstract:The increasing deployment of AI agents in long-horizon tasks yields massive execution logs. Diagnosing failures within these records is crucial for reliability, as it transforms outcome-level signals into actionable interventions. The sheer scale of the data renders human review impractical, driving the need for automated root-cause attribution (RCA). However, automated RCA methods using LLMs suffer from low diagnostic accuracy, especially as execution traces grow larger. They struggle because relevant information is often sparse, distributed across distant actions, and disconnected from the visible failure, reducing root-cause attribution to a massive search problem. Existing RCA methods typically rely on one-shot LLM judgments to diagnose failures from execution traces. While effective for shorter trajectories, these judges tend to settle on a plausible diagnosis early, leaving critical evidence in longer traces unexamined. We introduce Continual Search, an iterative framework that nudges the judge, over successive turns, to keep searching for unresolved diagnostic evidence. We evaluate Continual Search across four existing RCA benchmarks. Recognizing the lack of massive execution traces in current benchmarks, we introduce MegaRCA-Mix to evaluate RCA at scale. MegaRCA-Mix provides a challenging testbed of 50 human-annotated failure trials spanning long-horizon, execution-heavy tasks. Across multiple benchmark suites and model families, Continual Search consistently improves attribution performance. On MegaRCA-Mix, for example, it improves GPT-5.5’s F1 score by more than 40\%, from $0.349$ to $0.498$. More interestingly, within the same model family, lower-tier models can even surpass their higher-tier counterparts, demonstrating that effective search supersedes raw model scale.

中文摘要

摘要：在长周期任务中，AI代理的日益部署产生了大量的执行日志。在这些记录中诊断故障对于可靠性至关重要，因为它将结果级信号转化为可执行的干预措施。数据的庞大规模使人工审查不切实际，从而推动了自动化根因分析（RCA）的需求。然而，使用大型语言模型（LLM）的自动RCA方法在诊断准确性上表现不佳，尤其是当执行轨迹变得更长时。原因在于相关信息通常稀疏，分布在相距较远的动作中，并且与可见故障脱节，使根因分析变成了一个巨大的搜索问题。现有的RCA方法通常依赖一次性LLM判断来从执行轨迹中诊断故障。虽然在较短轨迹中有效，但这些判断者往往过早做出一个看似合理的诊断，导致长轨迹中的关键证据未被检查。我们提出了持续搜索（Continual Search）框架，这是一个迭代方法，通过多次回合推动判断者不断搜索未解决的诊断证据。我们在四个现有的RCA基准上评估了持续搜索。鉴于当前基准中缺乏大量执行轨迹，我们引入了MegaRCA-Mix，以大规模评估RCA。MegaRCA-Mix提供了一个具有挑战性的测试平台，包括50个人工标注的失败实验，涵盖长周期、执行密集型任务。在多个基准套件和模型家族中，持续搜索始终提升了根因分析性能。例如，在MegaRCA-Mix上，它提升了GPT-5.5的F1分数超过40%，从0.349提高到0.498。更有趣的是，在同一模型家族中，较低等级的模型甚至可以超越其高等级对手，这表明有效的搜索优于模型规模本身。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13463 (HTTP 429)

Authors: Harsh Raj, David Lee, Anas Mahmoud, Renxiong Wang, Razvan-Gabriel Dumitru, Chenguang Wang, Tong Zhao, Yunzhong He, Darvin Yi, Vipul Gupta

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13463.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13463

Published: 2026-09-16T01:39:37.206Z

9. Governing at Machine Speed: An Adaptive Intelligence Architecture for Real-Time AI Policy Enforcement

Abstract:Enterprise AI adoption has reached 78% of organizations globally, yet the infrastructure to govern that adoption has not kept pace. This paper identifies and characterizes the attestation deficit, a structural condition in which organizations maintain governance policies but cannot produce auditable, tamper-evident evidence of enforcement within regulatory timelines. Drawing on empirical data from the Stanford 2026 AI Index Report (362 documented incidents), the IBM/Ponemon 2026 Cost of a Data Breach study (USD 4.99M average cost, 92% lacking access controls), and the EY/AIUC-1 Consortium survey (38% end-to-end monitoring, 17% agent-to-agent coverage), this paper demonstrates that the governance failure is organizational and architectural rather than technical. To address this deficit, we propose AGIL (Adaptive Governance Intelligence Layer), a conceptual five-layer architecture designed to use machine learning for real-time AI governance enforcement. The proposed layers include: (1) Autonomous Discovery for shadow AI detection via behavioral fingerprinting, (2) Behavioral Risk Classification unifying security, hallucination, privacy, and accountability scoring, (3) a Policy Enforcement Gateway for inline permit/deny/modify decisions at sub-100ms latency, (4) a Continuous Attestation Engine generating tamper-evident audit trails as a byproduct of enforcement, and (5) Adaptive Policy Intelligence for ML-driven policy evolution across jurisdictions. AGIL is presented as a theoretical framework and architectural proposal; empirical validation through controlled deployment remains a direction for future work.

中文摘要

摘要：企业人工智能的采用在全球范围内已达到78%的组织，但用于管理这一采用的基础设施并未同步发展。本文识别并描述了认证赤字，这是一种结构性状况，组织保持治理政策，但无法在监管时间范围内提供可审计、防篡改的执行证据。本文借助《斯坦福2026年人工智能指数报告》（记录362起事件）、IBM/Ponemon 2026年数据泄露成本研究（平均成本499万美元，92%缺乏访问控制）以及EY/AIUC-1联盟调查（38%端到端监控，17%代理对代理覆盖率）的实证数据，表明治理失败是组织和架构层面的，而非技术层面的。为应对这一赤字，我们提出AGIL（自适应治理智能层），这是一个概念性的五层架构，旨在利用机器学习实现人工智能治理的实时执行。拟议的五个层包括：（1）自主发现层，通过行为指纹检测影子人工智能；（2）行为风险分类层，统一安全、幻觉、隐私和问责评分；（3）政策执行网关，在不到100毫秒的延迟内实现在线允许/拒绝/修改决策；（4）持续认证引擎，作为执行的副产品生成防篡改审计轨迹；（5）自适应政策智能层，实现跨司法管辖区的机器学习驱动政策演化。AGIL作为理论框架和架构提案提出；通过受控部署进行实证验证将是未来工作的方向。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13466 (HTTP 429)

Authors: Sandeep Bokkasam, B. Durgalakshmi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13466.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13466

Published: 2026-09-16T01:39:37.206Z

10. OrchSLM: Probing the Dynamics of Small Language Model Orchestration

Abstract:Although large language models (LLMs) have demonstrated remarkable capabilities, their reliance on cloud-scale infrastructure poses fundamental challenges for deployment in agentic pipelines, including latency, privacy, connectivity, and substantial computational cost. Small language models (SLMs) offer a compelling alternative: recent studies suggest that many repetitive and narrowly scoped subtasks in agentic workloads may be better served by specialized SLMs than by monolithic LLMs. However, the limited capacity and context windows of SLMs can constrain long-horizon reasoning and interaction-heavy orchestration strategies such as iterative verification and debate. This motivates a complementary, non-interactive paradigm in which heterogeneous SLMs independently generate candidate solutions and a router orchestrates their cached samples without further model interaction. To further understand the mechanisms of such orchestration, we introduce OrchSLM, a routing framework that unifies existing non-interactive orchestration methods and exposes their underlying design choices as controllable parameters. Using OrchSLM as a systematic probe, we reveal how orchestration behavior emerges from diverse knobs, including the task structure, model-pool composition, and multi-agent consensus.

中文摘要

摘要：尽管大型语言模型（LLM）已经展示了显著的能力，但它们对云级基础设施的依赖在代理流水线的部署中带来了根本性挑战，包括延迟、隐私、连接性以及巨大的计算成本。小型语言模型（SLM）提供了一种有吸引力的替代方案：最近的研究表明，在代理工作负载中，许多重复且范围狭窄的子任务可能由专门的SLM比由单一的LLM更好地完成。然而，SLM的容量和上下文窗口有限，这可能限制长周期推理和以交互为主的编排策略，如迭代验证和辩论。这促使提出一种互补的非交互范式，其中异构SLM独立生成候选解，而路由器在不进一步模型交互的情况下编排它们的缓存样本。为了进一步理解这种编排的机制，我们引入了OrchSLM，这是一个路由框架，它统一现有的非交互编排方法，并将其潜在设计选择作为可控参数暴露出来。通过将OrchSLM作为系统化探测工具，我们揭示了编排行为如何从多样化的旋钮中出现，包括任务结构、模型池组成和多代理共识。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13470 (HTTP 429)

Authors: Chengxi Zhang, Yu Yao

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13470.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13470

Published: 2026-09-16T01:39:37.206Z

Evaluation Domain Papers

1. ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

中文摘要

LLM Analysis

1.4 × 1.5 × 1.8 × 1.1 ≈ 4.2

2.2 渐进课程与 MDP 中期训练 - 上下文扩展：中期训练以 600B token 分三阶段扩展序列长度： 16K arrow 64K arrow 256K ，每阶段保留更短序列的回放数据。 - MDP 监督：将多轮智能体交互轨迹重构为马尔可夫决策过程（MDP）风格的状态-动作转移，提供密集的逐步决策监督，而非仅复现完整轨迹。 #### 2.3 后训练校准 - 分层 SFT 数据筛选：通过规则过滤与模型质量评分对候选数据执行约 50% 的激进剪枝，实证表明质量优于数量。 - 联合 Think/No-Think 训练：在 SFT 中混合同权重的显式推理轨迹（think）与直接回答（no-think），使模型能根据指令动态切换模式，且结构化推理监督对直接回答产生正向迁移。 - 混合强化学习：采用 GRPO 结合结果奖励、KL 正则化与长度惩罚，在数学与代码等单领域任务上进一步增益。 —- ### 3. 实验与评估结果 #### 3.1 与同规模模型对比（7B–8B）在 20 项标准基准的广泛评估中，ZGCM-1-7B 在 14 项推理基准上平均排名第一： - 数学推理：MATH-500 97.13%，AIME 2026 75.00%，HMMT 2025 70.42% - 代码生成：HumanEval+ 90.24%，MBPP+ 63.23% - 知识：MMLU 73.88%，GPQA-Diamond 47.87% #### 3.2 与前沿大模型对比（智能体任务）通过结合深度推理与自主工具调用，7B 模型在智能体任务上可与数量级更大的系统竞争： - WebWalkerQA：63.09% - BrowseComp：19.43% - Binary Function Search（自建基准）：62.00% 后两项接近或持平 GLM-5.1、Kimi-K2、Claude 4 Sonnet 等大规模模型。 —- ### 4. AI 原生研发（AI-Native R&D）论文构建了一套由研究者指导、智能体集群（Agent Swarms）执行的 R&D 工作流，覆盖数据清洗、集群管理、自动实验、原子能力评估（ACE）与部署。对 11 类 R&D 任务的自主性评估显示： - L4（高自主）：实验监控、部署工程 - L3（条件自主）：数据清洗、资源管理、模型评估等 - L2（部分自动化）：模型架构设计、学习算法设计（暂无评级超过 L3） —- ### 5. 八大经验发现（Findings）论文在全生命周期中提炼了八条可操作的实证发现，涵盖： 1. 词汇复杂度课程对通用语言有效，但对代码与数学不可靠； 2. TWEO 对 FP8 训练稳定性至关重要； 3. 系统-算法协同设计带来约 4.2× 预训练效率提升； 4. 后训练中数据质量优于数据数量； 5. 长 CoT 监督存在非单调效用，需动态校准； 6. 智能体训练必须与通用指令数据联合进行； 7. 中期训练建立的长上下文基础可规避超长的后训练轨迹需求； 8. AI 自主性高度依赖任务类型，设计与运营任务差距显著。 —- ### 6. 开放科学贡献为消除透明度壁垒，论文公开了： - 预训练、中期训练、后训练全阶段模型权重及中间检查点 - 训练代码、配置与超参数 - 每阶段数据配方、处理流水线与 W&B 日志 - 评估工具链与自建基准（Binary Function Search） Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13356.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13356

Published: 2026-09-16T01:40:01.394Z

2. Converge Then Diversify: Decoupling Convergence and Diversity in Multi-Objective Bayesian Optimisation

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13396 (HTTP 429)

Authors: Chao Jiang, Yueling Huang, Miqing Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13396.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13396

Published: 2026-09-16T01:40:01.394Z

3. Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement

中文摘要

摘要：当我们谈论递归自我改进（RSI）时，我们是在谈论一种现象、一种机制，还是一种前景？迈向自主和不断进化的智能，RSI 在许多层面上被提出，但尚不存在一个能够正式描述这些新兴实例的统一框架。其在经典领域的对应物——迭代策略改进，则以广义策略迭代（GPI）为特征，这是一种具有广泛适用性的框架，且其理论性质已被充分理解，但前提是更新原理和评估基础位于智能体之外。在本文中，我们提出了广义智能体迭代（GAI），这是一种正式框架，将迭代策略改进和 RSI 描述为单一学习范式的两种情况。GAI 将智能体定义为系统内可修改组件的配置，并将学习过程建模为智能体评估与智能体改进的循环。两个关键旋钮用于区分不同实例：改进机制是否是智能体的一部分，以及其所依据的标准是否建立于智能体之外。前者旋钮划定了 GPI 与 RSI 的边界，后者则决定系统的极性是锚定型、目标漂移型，还是完全自我参照型。此外，我们利用这些坐标将现有系统置于同一两个轴上，并逐一条件说明递归自我改进的缺陷。我们将本文视为探索 RSI 正式特征的第一步，该特征基于经典理论，使现有系统可以相互比较，并为分析和设计新系统提供原则性基础。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13406 (HTTP 429)

Authors: Hongyao Tang, Yi Ma, Pengyi Li, Yifu Yuan

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13406.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13406

Published: 2026-09-16T01:40:01.394Z

4. Vibe Patenting: Evaluating LLM Judges for Professional Patent-Drafting Agents

中文摘要

摘要：大型语言模型（LLM）评审者越来越多地被用于评估和改进 AI 生成的输出，但它们在复杂专业工作中的可靠性仍不明确。我们通过 Vibe Patenting 研究这一问题，这是一种用于 AI 代理评估的端到端专利撰写测试平台。单独调用的 LLM 评审者会评估生成的专利草案，并提供结构化反馈以进行迭代修改。在多个发明和撰稿代理配置中，评审者指导的修改始终提高了评审者评估的质量，而未指导的修改往往达到饱和。值得注意的是，迭代的评审反馈使低推理能力的代理能够接近性能明显更高的高推理能力代理。更强的模型和增加的推理能力通常会提升评审者评估的撰写质量，而特定领域的代理工作流则带来进一步提升。我们通过独立的专业专利律师评估验证了评审者，发现有显著但强烈依赖指标的一致性以及系统性的校准差异。这些结果突显了 LLM 评审者作为复杂专业工作流评估者和优化信号的用途与局限性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13422 (HTTP 429)

Authors: Toshiaki Koike-Akino, Vlad Blaykhman, Ye Wang, Jing Liu, Gene V. Vinokur

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13422.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13422

Published: 2026-09-16T01:40:01.394Z

5. Toward Self-Adaptive Physical AI: Can LLM Agents Manage Long-Horizon Physical Tasks?

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13436 (HTTP 429)

Authors: Varun Kaushik, Yayun Tan, Xiaofan Yu

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13436.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13436

Published: 2026-09-16T01:40:01.394Z

6. LabAgent: Customize Any Research Hubs for Scientific Discoveries Using AI Agents

中文摘要

摘要：科学研究是一个强调传承的连续过程。前人开发的方法通常会被新的研究人员加以扩展，以探索更新颖、更深入的科学问题。然而，实验室人员的变化，例如学生毕业，会导致缺乏能够复现方法的人员。那些投入了大量努力和资源开发的方法可能无法继续使用。为了解决这些限制，我们提出了LabAgent，一种为实验室持续工作量身定制的复现与发现工具。LabAgent采用两种机制来保证所有技能的执行和验证，并记录修正方法和经验，从而可以直接纠正或避免类似错误。我们将LabAgent应用于药物性质预测、生命科学领域的生物医药问题分析、蛋白质变体效应预测以及统计遗传学。LabAgent在每个领域都排名第一，优于商业通用代理，并展示了对已发表图表的准确复现。总体而言，这些结果表明，LabAgent能够有效整合并合理扩展实验室知识。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13437 (HTTP 429)

Authors: Lei Liu, Yikun Zhang, Jialin Chen, Wanjia Zhao, Rex Ying, Wengong Jin, Hua Xu, James Zou, Tianyu Liu, Hongyu Zhao

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13437.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13437

Published: 2026-09-16T01:40:01.394Z

7. TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13457 (HTTP 429)

Authors: Sudarshan Regmi, Arvind Pillai, Yu Yvonne Wu, Yuliang Chen, Bibek Panthi, Tess Z. Griffin, Michael V. Heinz, Lisa Marsch, Nicholas C. Jacobson, Andrew Campbell

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13457.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13457

Published: 2026-09-16T01:40:01.394Z

8. Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures

中文摘要

摘要：在长周期任务中，AI代理的日益部署产生了大量的执行日志。在这些记录中诊断故障对于可靠性至关重要，因为它将结果级信号转化为可执行的干预措施。数据的庞大规模使人工审查不切实际，从而推动了自动化根因分析（RCA）的需求。然而，使用大型语言模型（LLM）的自动RCA方法在诊断准确性上表现不佳，尤其是当执行轨迹变得更长时。原因在于相关信息通常稀疏，分布在相距较远的动作中，并且与可见故障脱节，使根因分析变成了一个巨大的搜索问题。现有的RCA方法通常依赖一次性LLM判断来从执行轨迹中诊断故障。虽然在较短轨迹中有效，但这些判断者往往过早做出一个看似合理的诊断，导致长轨迹中的关键证据未被检查。我们提出了持续搜索（Continual Search）框架，这是一个迭代方法，通过多次回合推动判断者不断搜索未解决的诊断证据。我们在四个现有的RCA基准上评估了持续搜索。鉴于当前基准中缺乏大规模执行轨迹，我们引入了MegaRCA-Mix来评估大规模RCA。MegaRCA-Mix提供了一个具有挑战性的测试平台，其中包含50个经过人工标注的故障实验，涵盖长周期、执行密集型任务。在多个基准测试集和模型系列中，持续搜索始终提升了归因表现。例如，在MegaRCA-Mix上，它将GPT-5.5的F1分数从$0.349$提高到$0.498$，提升超过40%。更有趣的是，在同一模型系列中，低端模型甚至可能超越高端模型，显示出有效的搜索优于单纯的模型规模。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13463 (HTTP 429)

Authors: Harsh Raj, David Lee, Anas Mahmoud, Renxiong Wang, Razvan-Gabriel Dumitru, Chenguang Wang, Tong Zhao, Yunzhong He, Darvin Yi, Vipul Gupta

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13463.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13463

Published: 2026-09-16T01:40:01.394Z

9. Governing at Machine Speed: An Adaptive Intelligence Architecture for Real-Time AI Policy Enforcement

中文摘要

摘要：企业人工智能的采用在全球范围内已达到78%的组织，但用于管理这一采用的基础设施并未同步发展。本文识别并描述了认证赤字，这是一种结构性状况，组织保持治理政策，但无法在监管时间内生成可审计、可防篡改的执行证据。本文借助《斯坦福2026人工智能指数报告》（记录362起事件）、IBM/Ponemon 2026《数据泄露成本研究》（平均成本499万美元，92%缺乏访问控制）以及EY/AIUC-1联盟调查（38%端到端监控，17%代理对代理覆盖）的实证数据，表明治理失败更多是组织和架构问题，而非技术问题。为解决这一赤字，我们提出了AGIL（自适应治理智能层），这是一种概念性五层架构，旨在利用机器学习实现实时人工智能治理执行。拟议的层包括：(1) 自主发现层，通过行为指纹检测影子AI，(2) 行为风险分类层，统一安全、幻觉、隐私和问责评分，(3) 政策执行网关层，在不到100毫秒的延迟内进行在线允许/拒绝/修改决策，(4) 持续认证引擎层，作为执行副产品生成可防篡改的审计轨迹，以及(5) 自适应政策智能层，用于跨司法管辖区的机器学习驱动的政策演化。AGIL作为理论框架和架构提案呈现；通过受控部署进行的实证验证仍是未来工作的方向。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13466 (HTTP 429)

Authors: Sandeep Bokkasam, B. Durgalakshmi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13466.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13466

Published: 2026-09-16T01:40:01.394Z

10. OrchSLM: Probing the Dynamics of Small Language Model Orchestration

中文摘要

摘要：尽管大型语言模型（LLM）展现了显著的能力，但它们对云级基础设施的依赖在部署于智能代理管道中带来了基本性挑战，包括延迟、隐私、连接性以及巨大的计算成本。小型语言模型（SLM）提供了一种有吸引力的替代方案：最近的研究表明，在智能代理工作负载中，许多重复性和范围狭窄的子任务可能由专门的SLM比由单一的LLM更好地完成。然而，SLM的有限容量和上下文窗口可能会限制长距离推理和互动密集型的编排策略，例如迭代验证和辩论。这激励了一种互补的非互动范式，其中异构的SLM独立生成候选解决方案，而路由器对它们缓存的样本进行编排，而无需进一步的模型互动。为了进一步理解这种编排的机制，我们引入了OrchSLM，这是一种路由框架，它统一了现有的非互动编排方法，并将它们的底层设计选择暴露为可控参数。利用OrchSLM作为系统化的探针，我们揭示了编排行为如何从多样化的调节因素中产生，包括任务结构、模型池组成和多代理共识。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13470 (HTTP 429)

Authors: Chengxi Zhang, Yu Yao

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.13470.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13470

Published: 2026-09-16T01:40:01.394Z

VLM Domain Papers

1. Personalized and Explainable Blood Pressure Estimation from PPG via Hybrid CNN—Morphological Features

Abstract:Continuous cuffless blood pressure (BP) monitoring using photoplethysmography (PPG) offers a promising solution for personalized healthcare. However, existing methods have two major limitations. Handcrafted feature-based approaches rely on precise fiducial point detection and are limited to short-term analysis, while deep learning models, despite their accuracy, often operate as black boxes with limited physiological interpretability. To address these challenges, we propose a physiology-guided hybrid framework for personalized BP estimation that couples a convolutional neural network (CNN) branch capturing global and local waveform dynamics with a morphology-prior branch that explicitly encodes person-specific vascular characteristics. By embedding a morphology-based feature set that explicitly encodes individual vascular characteristics, the proposed framework enhances personalization and reduces dependence on large-scale training datasets. Evaluated on a subset of the MIMIC-III database under a subject-specific (personalized) testing protocol, the proposed personalized physiology-guided hybrid approach achieved mean absolute errors (MAEs) of 3.77 mmHg for systolic BP and 2.36 mmHg for diastolic BP, corresponding to relative improvements of 43.7% and 32.4% over a subject-specific (personalized) CNN-only baseline. SHAP-based analysis confirmed that the introduced morphology-prior features align with individual vascular characteristics, reinforcing per-subject interpretability. These findings highlight the potential of personalized, physiology-guided hybrid learning with novel morphological descriptors for accurate and explainable BP monitoring in real-world settings.

中文摘要

摘要：使用光电容积描记术（PPG）的连续无袖带血压（BP）监测为个性化医疗提供了有前景的解决方案。然而，现有方法存在两大局限性。基于手工特征的方法依赖于精确的基准点检测，且仅限于短期分析；而深度学习模型尽管准确，但通常作为黑箱操作，其生理可解释性有限。为应对这些挑战，我们提出了一种基于生理指导的混合框架，用于个性化血压估计，该框架将捕捉全局和局部波形动态的卷积神经网络（CNN）分支与明确编码个体血管特征的形态先验分支结合。通过嵌入明确编码个体血管特征的基于形态的特征集，所提出的框架增强了个性化能力，并减少了对大规模训练数据集的依赖。在MIMIC-III数据库的子集上，采用个体特定（个性化）测试协议评估，所提出的个性化生理指导混合方法实现了收缩压平均绝对误差（MAE）为3.77 mmHg，舒张压MAE为2.36 mmHg，相较于个体特定（个性化）仅CNN基线，分别提高了43.7%和32.4%。基于SHAP的分析证实，引入的形态先验特征与个体血管特征一致，加强了对每个受试者的可解释性。这些发现突显了基于新型形态描述符的个性化生理指导混合学习在真实环境中实现准确且可解释血压监测的潜力。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13190 (HTTP 429)

Authors: Myung-Kyu Yi, Jongshill Lee, Jeyeon Lee, In Young Kim

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13190.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13190

Published: 2026-09-16T01:40:25.728Z

2. Part Grounding, Not Action Knowledge: Locating the Bottleneck in VLM Affordance Prediction

Abstract:Benchmarks agree that vision-language models reason poorly about low-level manipulation, but an aggregate accuracy score does not say which step fails. We separate two steps that affordance questions conflate: identifying which part of an object to act on, and knowing what action that part requires. Across 19 articulated objects we asked eight models, spanning three developers, what motion a robot should apply. Under an open prompt, push was produced once in 64 evaluations where it was correct, despite being correct for 8 of 19 objects and appearing in the offered label set every time. Inspecting the outputs showed why: models described a different part than the one being scored, e.g. explaining how to pick up a camera rather than press its button. Naming the target part raises action accuracy by 0.32 to 0.63 for every model, from 0.158-0.474 to 0.684-0.947, and push recall from 0-1/8 to 7-8/8. No model beats a constant answer that ignores the image under the open prompt; once the part is named, all eight do. Asked to describe the same part in free prose with no label set, models produce pressing language for 6 to 8 of 8. These results are hard to reconcile with missing action knowledge, and instead point to part grounding as the dominant bottleneck, a pattern that holds across all three model families and does not diminish with capability. Naming the part supplies the grounding variable, so this bounds what a perfect part detector would offer rather than demonstrating a general model of mechanics. Two supporting results agree: on real photographs only three of eight models localize grasp points better than a constant baseline, and on rendered objects none do. We also document two measurement errors of our own, a threshold that let a constant baseline score 0.929 and a labelling rule wrong on 4 of 19 objects, both caught only by testing our numbers against trivial alternatives.

中文摘要

摘要：基准测试一致认为视觉-语言模型在低级操作推理方面表现不佳，但一个综合的准确率得分并不能说明是哪一步出错。我们将可供性问题混淆的两个步骤分开：识别对象的哪一部分需要操作，以及知道该部分需要执行什么动作。在对19个关节物体进行测试时，我们向三个不同开发者的八个模型询问机器人应该应用什么动作。在开放式提示下，在64次评估中正确的情况下，“推动”仅出现一次，尽管它对19个对象中的8个是正确的，并且每次都出现在提供的标签集中。检查输出结果显示原因：模型描述了与评分部件不同的部分，例如解释如何拿起相机，而不是按下它的按钮。指明目标部件将每个模型的动作准确率从0.158-0.474提高到0.684-0.947，增幅为0.32至0.63，而推动的召回率从0-1/8提高到7-8/8。在开放式提示下，没有任何模型超过忽略图像的固定答案；一旦指定了部件，八个模型都能超过。被要求以自由文字描述相同部件且不使用标签集时，模型在8个中能产生6到8个与按压相关的描述。这些结果难以与缺失动作知识相协调，反而指向部件定位是主要瓶颈，这一模式在所有三类模型中保持不变，并且不会随着能力提升而减弱。指定部件提供了定位变量，因此这限定了完美部件检测器能提供的信息，而不是证明通用的机械模型。两个支持性结果也一致：在真实照片上，只有八个模型中有三个位点定位抓取点优于固定基准，而在渲染物体上则没有一个模型做到。同时，我们还记录了两个自身的测量错误：一个阈值导致固定基准得分为0.929，以及一个标注规则在19个对象中有4个错误，这两者都是通过将我们的数据与简单替代方案进行测试才发现的。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13225 (HTTP 429)

Authors: Sarthak Sattigeri

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13225.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13225

Published: 2026-09-16T01:40:25.728Z

3. Synthetic Leprosy Image Generation Using Mask-Conditioned Latent Diffusion and Transfer Learning from Large Chronic Wound Datasets

Abstract:Machine learning for neglected tropical diseases is limited by data, not algorithms: public annotated image sets for leprosy (Hansen’s disease) number in the hundreds, orders of magnitude below what generative models require. We ask whether a model trained on abundant chronic wound photography transfers to this low-data regime. We build a three-stage pipeline. First, a DeepLabV3-ResNet50 segmentation network (validation Dice 0.876, IoU 0.799) supplies lesion masks for two wound datasets that ship without them. Second, we assemble a mask-conditioned latent diffusion model from Stable Diffusion 1.5 components and train it on 3,280 region-of-interest wound crops, widening the UNet input convolution from 4 to 11 channels to admit three mask feature maps and a blurred low-frequency context latent. Third, we fine-tune this model on 708 leprosy image-mask pairs drawn from 764 images of approximately 150 patients. We evaluate with LPIPS perceptual distance, anchored by a real-versus-real baseline computed on the same 242 anchor images as the cross-set comparisons; without that reference the cross-set distances cannot be interpreted. The generated set shows no mode collapse: its internal perceptual diversity (0.662) is statistically indistinguishable from that of the real leprosy set (0.672, 95% CI [0.664, 0.680]). Generated images sit 0.044 LPIPS outside the real distribution - measurably apart, but under half of one standard deviation. Fine-tuning shifted the output distribution only marginally, which we trace to lesion geometry reaching the network through input concatenation alone. Chronic wound photography is therefore a viable donor domain for leprosy lesion synthesis: low-level appearance transfers well, and the remaining barrier is semantic control rather than image quality.

中文摘要

摘要：被忽视的热带病的机器学习受到数据限制，而非算法限制：公开的麻风病（Hansen病）标注图像集数量仅有数百张，远低于生成模型所需的数据量。我们探讨了一个在丰富的慢性伤口摄影数据上训练的模型是否可以迁移到这种低数据环境中。我们构建了一个三阶段管道。首先，一个 DeepLabV3-ResNet50 分割网络（验证集 Dice 为 0.876，IoU 为 0.799）为两个未附带标注的伤口数据集提供病灶掩码。其次，我们使用 Stable Diffusion 1.5 组件组建了一个掩码条件潜变量扩散模型，并在 3,280 张感兴趣区域 (ROI) 的伤口裁剪图像上进行训练，将 UNet 输入卷积从 4 通道扩大到 11 通道，以接纳三个掩码特征图和一个模糊的低频上下文潜变量。第三，我们在从约 150 名患者的 764 张图像中提取的 708 对麻风病图像-掩码对上微调该模型。我们使用 LPIPS 感知距离进行评估，其基准为在相同的 242 张锚图像上计算的真实对真实基线；没有该参考，跨数据集距离无法解释。生成的数据集没有出现模式坍塌：其内部感知多样性（0.662）在统计上与真实麻风病数据集（0.672，95% 置信区间 [0.664, 0.680]）无显著差异。生成图像相对于真实分布偏离 0.044 LPIPS——可测量地不同，但低于半个标准差。微调仅略微改变了输出分布，我们将其归因于病灶几何信息通过输入连接传递给网络。因而，慢性伤口摄影可以作为麻风病病灶合成的可行源域：低级外观能够良好迁移，剩下的障碍是语义控制，而非图像质量。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13226 (HTTP 429)

Authors: Yusuf Abdulkadir

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13226.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13226

Published: 2026-09-16T01:40:25.728Z

4. Don’t Just Look, Intervene: Perturbation Based Region Labeling for VQA Images

Abstract:Vision Language Models (VLMs) should rely on visual evidence that directly determines the correct answer, but supervision for grounding visual reasoning is often expensive to obtain manually or tied to dataset-specific annotation primitives. We instead introduce model-causal visual evidence as an annotation target, defined as the set of image regions whose counterfactual intervention changes a model’s answer distribution for a given image-question pair. Based on this principle, we introduce Counterfactual Search for Grounding Regions (CSGR). CSGR is a scalable pipeline that proposes candidate regions, perturbs them, measures their effect on answer sensitivity, and aggregates this evidence across multiple judges to approximate answer-critical regions in VQA data. To assess whether CSGR annotations contain a useful supervision signal, we plug them into three existing grounding-aware training routines: attention steering, Visual CoTfinetuning, and latent visual reasoning. These experiments test whether the proposed annotation scheme can provide a useful supervision signal across multiple ways of consuming region labels, rather than introducing a new way of using them. Across competing automatic region-labeling mechanisms, CSGR annotations provide the most consistent gains over Cross Entropy-only finetuning in both in-domain and out-of-domain evaluations, indicating that the proposed labeling scheme captures useful region-level information.

中文摘要

摘要：视觉语言模型（VLMs）应依赖直接决定正确答案的视觉证据，但用于视觉推理定位的监督通常难以通过人工获得，或者绑定于特定数据集的标注原语。我们提出将模型因果视觉证据作为标注目标，该目标定义为一组图像区域，其反事实干预会改变模型对给定图像-问题对的答案分布。基于这一原理，我们提出了用于定位区域的反事实搜索（CSGR）。CSGR 是一个可扩展的流程，它提议候选区域，对其进行扰动，测量其对答案敏感性的影响，并通过多个评判者聚合这些证据，以近似 VQA 数据中的关键答案区域。为了评估 CSGR 标注是否包含有用的监督信号，我们将其引入三种现有的知晓区域的训练方法：注意力引导、视觉链式微调（Visual CoTfinetuning）和潜在视觉推理。这些实验测试所提出的标注方案是否能在多种利用区域标签的方式中提供有用的监督信号，而不是引入利用它们的新方法。在竞争的自动区域标注机制中，CSGR 标注在域内和域外评估中均相比仅使用交叉熵微调显示出最一致的提升，这表明所提出的标注方案捕捉了有用的区域级信息。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13228 (HTTP 429)

Authors: Marko Jojic, Zhaonan Li, Ben Zhou

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13228.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13228

Published: 2026-09-16T01:40:25.728Z

5. What Does the Encoder Actually Decide? A Controlled Comparison of Vision Backbones on Joint Tree Segmentation and Stereo Depth

Abstract:A robot pruning trees needs two facts per pixel: whether it belongs to a tree, and its distance. Both are usually obtained via task heads attached to a vision backbone chosen by reputation rather than measurement. Holding dataset, decoders, losses, schedule, and evaluation fixed, we ask: how much does the encoder choice change joint semantic segmentation and stereo depth on thin vegetation? We build a hard parameter-sharing network with one encoder feeding both branches, swapping only the encoder without downstream retuning. We evaluate [N] encoders across [M] architecture families (CNNs, transformers, hybrids, MLP-mixers, state-space models) near a ~25M budget, trained from scratch. Depth is evaluated on tree pixels only; segmentation uses boundary F1 and background IoU to prevent “label-everything-tree” shortcuts. Three findings stand out. First, the strongest encoders are convolutional and hybrid, not transformers: [BestEncoder] leads with [MIoU] segmentation mIoU and [Delta] depth $\delta_1$, while [X] of [Y] plain vision transformers collapse when trained from scratch. Second, parameter count does not predict quality —- [SmallEncoder] at only [P]M parameters outranks models two orders of magnitude larger. Third, segmentation and depth rankings agree strongly (Spearman $\rho$ = [RhoValue]), showing no task conflict. Finally, [K] of [N] encoders collapse to degenerate all-tree segmentation —- exposed by boundary F1 but hidden by region IoU.

中文摘要

摘要：机器人修剪树木每个像素需要两个事实：是否属于树，以及其距离。这两者通常通过连接在视觉骨干上的任务头获得，这些数据基于声誉而非测量。固定数据集、解码器、损失、调度和评估后，我们问：编码器的选择对稀疏植被的联合语义分割和立体深度有多少影响？我们构建了一个硬参数共享网络，一个编码器供给两个分支，仅交换编码器，无需下游重调。我们在[M]架构家族（卷积神经网络、变压器、混合器、MLP混合器、状态空间模型）中评估[N]个编码器，预算约为2500万，从零训练。深度仅在树像素上评估;分割使用边界F1和背景IoU防止“标记所有树”捷径。有三个发现尤为突出。首先，最强编码器是卷积和混合编码器，而非变换器：[BestEncoder] 以 [MIoU] 分割 mIoU 和 [Delta] 深度 $\delta_1$ 领先，而 [X] 的 [Y] 平视变换器在从零训练时会崩溃。其次，参数计数不预测 [SmallEncoder] 仅在 [P]M 参数下—-质量，比大两个数量级的模型更重要。第三，分割和深度排名高度一致（Spearman $\rho$ = [RhoValue]），显示无任务冲突。最后，[N] 编码器的 [K] 崩溃，导致退化的全树分割，—-边界 F1 暴露，但被 IoU 区域隐藏。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13232 (HTTP 429)

Authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13232.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13232

Published: 2026-09-16T01:40:25.728Z

6. EMCStereo: Attention-Enhanced Stereo Matching for Thin-Structure Depth Estimation with a Synthetic Tree-Branch Benchmark

Abstract:Thin structures such as tree branches are among the hardest cases for stereo matching: a branch is only a few pixels wide, the background is cluttered, and dense ground truth for real branches is nearly impossible to label by hand. We make three contributions. First, EMCStereo integrates three lightweight attention modules into a PSMNet-style cost-volume backbone: Efficient Multi-scale Attention (EMA) on deep semantic features, a Multi-Scale Fusion block (MSFblock) learning spatial pyramid weights instead of concatenating them, and Coordinate Attention (CoordAtt) on final matching features. Because MSFblock collapses four pyramid branches into one, the modules leave the network 2.0% smaller and add only 1.7% inference time overhead. Second, VirtualTree is a synthetic stereo dataset rendered in Unreal Engine 5 with a simulated ZED Mini rig, providing 5,520 pairs with exact disparity for thin branches. Third, an eight-way ablation establishes a run-to-run noise floor of 0.009 px end-point error (EPE). EMCStereo achieves 1.31 px EPE (5.96% D1-all) on the VirtualTree test split, 1.00 px on SceneFlow, and 0.80, 0.73, 0.62, and 3.19 px on KITTI 2012, KITTI 2015, ETH3D, and Middlebury, with depth accuracy delta_1 from 92.6% to 98.7%. Evaluated against the noise floor, the attention stack is accuracy-neutral at a 100-epoch budget, while MSFblock and CoordAtt cost 0.03-0.05 px unless EMA is present.

中文摘要

摘要：像树枝这样细的结构是立体匹配中最难处理的情况之一：树枝宽度仅几像素，背景杂乱，而且真实树枝的密集标注几乎不可能手工完成。我们做出了三个贡献。首先，EMCStereo将三个轻量级注意力模块集成到类似PSMNet的代价体积主干中：在深层语义特征上的高效多尺度注意力（EMA）、学习空间金字塔权重而非将其拼接的多尺度融合模块（MSFblock）以及最终匹配特征上的坐标注意力（CoordAtt）。由于MSFblock将四条金字塔支路折叠为一条，这些模块使网络体积缩小2.0%，推理时间仅增加1.7%。其次，VirtualTree是一个在Unreal Engine 5中使用模拟ZED Mini装置渲染的合成立体数据集，提供5,520对具有精确视差的细树枝图像。第三，八向消融实验建立了0.009像素端点误差（EPE）的运行间噪声基线。在VirtualTree测试集上，EMCStereo达到了1.31像素EPE（D1-all 5.96%），在SceneFlow上为1.00像素，而在KITTI 2012、KITTI 2015、ETH3D和Middlebury上分别为0.80、0.73、0.62和3.19像素，深度精度delta_1在92.6%到98.7%之间。与噪声基线相比，在100轮训练预算下，注意力堆栈对精度没有显著影响，而MSFblock和CoordAtt在没有EMA的情况下会增加0.03-0.05像素误差。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13233 (HTTP 429)

Authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13233.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13233

Published: 2026-09-16T01:40:25.728Z

7. Occlusal Geometry in Closed Form for Orthodontic Report Generation

Abstract:Orthodontic report generation from intraoral data is normally cast as multimodal captioning, yet the released Bite2Text scan pairs are supplied already registered in occlusion, which makes several core occlusal quantities directly measurable rather than inferable. The system reported here exploits that property: an anatomical frame is recovered per case from arch taper and arch closure instead of the stated RAS convention, which does not hold across the release, and each arch is reduced to an occlusal ridge profile in arch-angle coordinates yielding overbite, overjet, midline deviation, transverse overlap, crossbite extent, cusp interdigitation lag, and the occlusal curves in closed form. Gradient boosting maps 31 such measurements onto 13 template fields, a field being predicted only where patient-level cross-validation beats its own majority baseline, and a deterministic renderer emits the corpus six-part narrative; a ConvNeXt-Tiny classifier over the five standardised photographic views is fused per field, raising mean field accuracy from 0.601 to 0.683. Reimplementation of the challenge evaluator shows that its BLEU-4 and METEOR are local variants whose F-mean weights recall nine to one, that two clinicians agree on 47 percent of findings for the same patient, and that a constant report consequently outscores a genuine second clinician report by 0.165 captioning. Held-out scores reach BLEU-4 0.458 and METEOR 0.677 against intraoral scan references and 0.278 and 0.507 against photograph references, and the submitted system placed third in the ODIN 2026 Bite2Text test phase at 0.2680 and 0.4629, within 0.022 BLEU-4 of first, running on CPU in under ten seconds per case. The dataset and code are available at this https URL

中文摘要

摘要：从口内数据生成正畸报告通常被视为多模态标注任务，但发布的 Bite2Text 扫描对已经在咬合状态下配准，这使得若干核心咬合量可以直接测量而非推断。本文报道的系统利用了这一特性：每个病例的解剖框架是从牙弓锥度和牙弓闭合情况恢复的，而非采用声明的 RAS 公约（该公约在数据集内不适用），每个牙弓被简化为牙弓角坐标系下的咬合嵴轮廓，从而得到覆盖、距、正中线偏移、横向覆盖、反牙合范围、牙尖互咬滞后以及闭合形式下的咬合曲线。梯度提升将 31 项此类测量映射到 13 个模板字段，一个字段仅在患者级交叉验证超过自身多数基线时进行预测，确定性渲染器输出语料库六部分叙述；对五个标准化摄影视角的 ConvNeXt-Tiny 分类器按字段融合，将平均字段准确率从 0.601 提升至 0.683。挑战评估器的重新实现显示，其 BLEU-4 与 METEOR 为局部变体，F-mean 权重是召回九比一，两名临床医生对同一患者的发现一致率为 47%，因此一个固定报告比真正的第二位医生报告的标注高 0.165。保留集分数在针对口内扫描参考时达到 BLEU-4 0.458 和 METEOR 0.677，对照片参考则为 0.278 和 0.507；提交系统在 ODIN 2026 Bite2Text 测试阶段排名第三，分别为 0.2680 和 0.4629，比第一名 BLEU-4 仅差 0.022，每例在 CPU 上运行时间不足十秒。数据集和代码可通过此 https 链接获得。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13237 (HTTP 429)

Authors: Ajo Babu George, Govind Arun, Sidharth N Krishna, Uma Ranjan

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13237.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13237

Published: 2026-09-16T01:40:25.728Z

8. Abstract-LoRA: Unlocking Single-Image Style Transfer through Targeted U-Net Block Training

Abstract:Diffusion models represent one of the most advanced paradigms in generative modeling. Leveraging their development, a growing number of style transfer methods based on diffusion models have been proposed. However, among these methods, multi-image style transfer approaches that require at least five to ten style examples tend to achieve more satisfactory results. Single-image methods, by contrast, often struggle with either insufficient content preservation or inadequate style fidelity. This greatly limits style extraction from scarce artworks and undermines their artistic value. To address this, we propose Abstract-LoRA, a method that pushes the boundaries of single-image style transfer through lightweight LoRA training on specific U-Net blocks in diffusion models. Specifically, our work is inspired by B-LoRA, a style transfer method that achieves basic style-content disentanglement by training specific U-Net blocks. However, it suffers from a critical limitation: the inability to capture complex backgrounds. Building upon B-LoRA, our method conducts a more refined analysis of U-Net blocks, employing additional U-Net blocks and clustering-based abstraction of style images to better disentangle and balance style and content. Extensive experiments demonstrate that our proposed method not only generates visually more harmonious and satisfying artistic images but also quantitatively improves the preservation of both style and content in the final outputs.

中文摘要

摘要：扩散模型是生成式建模中最先进的范式之一。利用其发展，越来越多基于扩散模型的风格迁移方法被提出。然而，在这些方法中，需要至少五到十个风格示例的多图像风格迁移方法往往能够取得更令人满意的结果。相比之下，单图像方法通常在内容保留不足或风格忠实度不够方面存在问题。这极大地限制了从稀缺艺术作品中提取风格的能力，并削弱了它们的艺术价值。为了解决这一问题，我们提出了Abstract-LoRA，一种通过对扩散模型中特定U-Net模块进行轻量级LoRA训练来推动单图像风格迁移边界的方法。具体而言，我们的工作受到B-LoRA的启发，B-LoRA是一种通过训练特定U-Net模块实现基本风格-内容解耦的风格迁移方法。然而，它存在一个关键限制：无法捕捉复杂背景。在B-LoRA的基础上，我们的方法对U-Net模块进行了更精细的分析，采用额外的U-Net模块和基于聚类的风格图像抽象，以更好地解耦和平衡风格与内容。大量实验表明，我们提出的方法不仅生成了视觉上更和谐和令人满意的艺术图像，而且在定量上提升了最终输出中风格和内容的保留效果。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13239 (HTTP 429)

Authors: Xinglin Hu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13239.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13239

Published: 2026-09-16T01:40:25.728Z

9. ArtSociety: Multi-Agent Multimodal Collaboration for Art Emotion Understanding

Abstract:The AffectiveArt Multidimensional Art Emotion Understanding task asks to jointly predict an artwork’s fine-grained emotion (12 classes, 1549:1 head-to-tail ratio), binary valence/arousal, and five attribute-grounded descriptions — sub-tasks that exhibit strong empirical trade-offs, so the single-model solutions we tried do not jointly optimize all of them well. We present ArtSociety, a multi-agent framework that assembles heterogeneous multimodal experts — a DINOv2-Giant vision agent (A1), a scene-grounded CoT fine-tuned MLLM (A2), and three closed-source reasoning agents (A3-A5) — and coordinates them with two training-free controllers: (i) a rare-class-aware voting arbiter that lowers the agreement threshold for tail emotions, exploiting decorrelated error patterns across agent families; and (ii) a description-first reasoning agent whose DESCRIBE-then-CLASSIFY chain of thought forces visual evidence before label commitment, yielding near-perfect grounded descriptions. A task-routing policy directs the hard emotion task to the full five-agent ensemble while assigning the near-saturated valence/arousal and generative description tasks to the single strongest reasoning agent. On the official test set (1,000 artworks), ArtSociety achieves an Overall Score of 0.8870 (Classification 0.7789, Description 0.9952). An eleven-variant ablation study reveals that, once method and scale saturate at around 0.76, the decisive gains come from agent collaboration and data-side supervision — a 30B MoE model trained on older data does not outperform an 8B model trained on better data. Code is available at this https URL

中文摘要

摘要：AffectiveArt 多维艺术情感理解任务要求联合预测艺术作品的细粒度情感（12 类，头尾比 1549:1）、二元情绪价值/激活度，以及五个基于属性的描述——这些子任务表现出明显的经验权衡，因此我们尝试的单模型解决方案无法同时很好地优化所有任务。我们提出了 ArtSociety，这是一个多智能体框架，它整合了异构的多模态专家——一个 DINOv2-Giant 视觉智能体（A1）、一个场景驱动的 CoT 微调多模态大语言模型（A2）以及三个闭源推理智能体（A3-A5）——并通过两个无需训练的控制器进行协调：（i）一个稀有类别感知投票仲裁器，可降低尾部情绪的意见一致性阈值，利用不同智能体群体间的去相关错误模式；以及（ii）一个描述优先的推理智能体，其“先描述后分类”的思维链在承诺标签前强制获取视觉证据，从而产生几乎完美的基于证据的描述。任务路由策略将困难的情感任务分配给完整的五智能体集成，同时将接近饱和的情绪价值/激活度和生成描述任务分配给单个最强推理智能体。在官方测试集（1000 件艺术作品）上，ArtSociety 实现了总体分数 0.8870（分类 0.7789，描述 0.9952）。十一种变体的消融研究显示，一旦方法和规模达到约 0.76 的饱和点，决定性的提升来自智能体协作和数据端监督——一个在较旧数据上训练的 30B MoE 模型并不优于在更好数据上训练的 8B 模型。代码可在此 https URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13240 (HTTP 429)

Authors: Jian Li, Fanfan Ji, Jinxiang Lai, Ying Tai, Jian Yang, Xiao-Tong Yuan, Chengjie Wang, Yabiao Wang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13240.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13240

Published: 2026-09-16T01:40:25.728Z

10. SJD-SV: Speculative Jacobi Decoding with Semantics Verification for Autoregressive Image Generation

Abstract:Speculative Jacobi Decoding (SJD) is an important approach for accelerating autoregressive image generation. Although SJD has shown superior performance, recent studies point out that it usually suffers from a token ambiguity issue during token verification but its reason can not be well explained. To figure out this reason, in this paper, we conduct a visualization analysis on vision token and find that different from text tokens, vision tokens generally corresponds to some local, small, and unclear vision details, which means only using single token is difficult to accurately express a certain semantic, thereby causing token ambiguity issue. To this end, we propose a novel Speculative Jacobi Decoding with Semantics Verification (called SJD-SV), for accelerating autoregressive image generation. The key idea is that leveraging the strong correction characters between tokens to recognize semantic-aware token subsequence and then instead of perform token-by-token verification, turning to perform verification on semantic-aware token subsequence level for accelerating image generation. In particular, our method is plug-in, which can be directly integrated into existing SJD and its variants. Extensive experiments on various datasets show that existing SJD methods achieve significant performance improvement after integrating our SJD-SV method.

中文摘要

摘要：投机性雅可比解码（Speculative Jacobi Decoding, SJD）是加速自回归图像生成的重要方法。虽然SJD已经表现出优越的性能，但近期研究指出它在令牌验证过程中通常存在令牌歧义问题，其原因尚无法很好解释。为弄清这一原因，本文对视觉令牌进行了可视化分析，并发现与文本令牌不同，视觉令牌通常对应一些局部、小型且不清晰的视觉细节，这意味着仅使用单个令牌难以准确表达某一特定语义，从而导致令牌歧义问题。为此，我们提出了一种新颖的具有语义验证的投机性雅可比解码（简称SJD-SV），用于加速自回归图像生成。其核心思想是利用令牌之间的强纠错特性识别语义感知的令牌子序列，然后不再进行逐令牌验证，而是在语义感知的令牌子序列层面进行验证，以加速图像生成。特别地，我们的方法为插件型，可直接集成到现有的SJD及其变体中。在各种数据集上的大量实验表明，在集成我们的SJD-SV方法后，现有的SJD方法实现了显著的性能提升。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.13245 (HTTP 429)

Authors: Baoquan Zhang, Bingqi Shan, Shihao Fang, Kenghong Lin, Xutao Li, Yunming Ye

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.13245.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.13245

Published: 2026-09-16T01:40:25.728Z