ArXiv Domain 2026-10-12

数据来源：ArXiv Domain

LLM Domain Papers

1. AutoIndexer: Flexible-Order Decoding by Training Causal Models on Chains of Edits

Abstract:Autoregressive decoding is less suited to non-causally ordered data, e.g., cross-references or future dependencies, since their outputs cannot be revised post hoc, unlike iterative human reasoning. Diffusion LLMs partially mitigate this limitation but still cannot flexibly insert or remove arbitrary numbers of tokens from a pre-generated context. We design AutoIndexer as a generative model with a novel training and decoding mechanism that simulates a Chain of Edits, allowing the model to flexibly modify its prior outputs in hindsight via token insertions, substitutions, and deletions. The model learns to predict marker tokens for opening and closing edits, simultaneously moving a cursor to positions in the existing token sequence, analogous to a text editor. Training sequences are obtained by perturbing ground-truth labels with inverse chained edit operations, and the model learns to edit by undoing those inverses. The perturbation training allows AutoIndexer to be robust to erroneous context and adaptable to editing tasks such as code correction. We demonstrate that our method exhibits post-generation context-editing capabilities that most autoregressive and diffusion models lack without compromising performance on standard LLM benchmarks. The code is available at this https URL.

中文摘要

摘要：自回归解码不太适用于非因果顺序的数据，例如交叉引用或未来依赖，因为其输出无法在事后修改，不像人类的迭代推理。扩散型大型语言模型部分缓解了这一限制，但仍无法灵活地从预生成的上下文中插入或删除任意数量的标记。我们设计了 AutoIndexer，作为一种具有新颖训练和解码机制的生成模型，能够模拟编辑链，使模型能够通过标记插入、替换和删除灵活地修改其先前输出。模型学习预测用于打开和关闭编辑的标记符，同时将光标移动到现有标记序列中的位置，类似于文本编辑器。训练序列通过使用逆向链式编辑操作扰动真实标签获得，模型通过撤销这些逆操作来学习编辑。扰动训练使 AutoIndexer 对错误的上下文具有鲁棒性，并能适应诸如代码修正的编辑任务。我们证明，该方法在生成后上下文编辑能力方面表现出其它大多数自回归和扩散模型所不具备的优势，同时在标准大型语言模型基准测试中的性能不受影响。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: 2610.12545

Authors: Shu Ishida, Aliasghar Khani, Tianyu Zhang, Cong Luong, James Seale Smith, Adam Gaier

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12545.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12545

Published: 2026-10-12T02:39:06.637Z

2. SWE-Duel: Towards a Self-Scaling Arena for Coding Agents

Abstract:Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository’s existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.

中文摘要

摘要：静态编码基准在发布时是固定的。随着时间推移，其任务会泄露到训练数据中，最终更强的智能体会完全掌握它们。我们认为，基准测试应旨在自我扩展：其任务和评测标准在评估时创建，因此不存在预先泄露的内容，其难度由参与的智能体决定，而不是在发布时固定。我们介绍了SWE-Duel，这是一个竞技场，两名编码智能体通过镜像的红蓝回合互相评估。在每个回合中，红方智能体在三个相互隔离的会话中处理一个真实的代码库。它提出并实现一个新功能并编写功能测试，然后在该功能中嵌入一个微妙的 bug，并添加封闭的触发 bug 测试。生成的拉取请求只有在满足以下条件时才被接受：(1) 它通过功能测试和代码库现有测试，同时仅失败于 bug 测试；(2) 同一智能体的新会话能够盲目修复该 bug，这排除了无法解决的任务。然后蓝方智能体在不知道存在 bug 的情况下审查该拉取请求。它只有在修订通过所有三个测试套件时才能获胜，然后两名智能体交换角色。我们在五种语言的15个代码库上，让11个模型控制的参与者进行1,500回合的比赛。领导者以不同的方式获胜。使用 Codex CLI 的 GPT-5.5 排名第一，作为最强的解决者，它修复了98.3%的审核缺陷。Claude Opus 4.8 搭配 mini-swe-agent 排名第二，作为最强的任务创建者，其红方获胜率为27.3%。最后，在同一模型家族中，由最新版本编写的任务比旧版本更常击败固定参考解决者（42.9% 对 4.0%），提供了初步证据表明任务难度随任务编写能力的提升而增长。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12569

Authors: Wei Minn, Zhensu Sun, Lwin Khin Shar, David Lo

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12569.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12569

Published: 2026-10-12T02:39:06.637Z

3. Contextual Correction of Parallel Draftsfor Speculative Decoding

Abstract:Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verified in parallel by the target model. Recent diffusion-based drafters make this paradigm particularly attractive: deep attention-based neural backbones provide strong predictions, while drafting multiple future positions in parallel avoids the sequential cost of autoregressive generation. This parallelism, however, comes at a cost. Tokens within a draft block are predicted without access to the realized identities of preceding draft tokens, removing intra-block dependencies that would naturally be available under autoregressive generation. We introduce DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores these dependencies while preserving parallel drafting. Rather than committing to each position’s top prediction, DCOCO retains multiple candidate tokens and evaluates their contextual compatibility across adjacent draft positions in parallel, enabling contextually informed token selection at low additional cost. Across four verifier backbones and ten benchmarks, DCOCO achieves the highest average throughput in every evaluated configuration, spanning greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding. Its advantage persists under concurrent serving: DCOCO outperforms the strongest competing method by 6-23% across 4-32 concurrent requests, while other sequential correctors fall below uncorrected DFlash at these loads. These results demonstrate that recovering intra-block dependencies through parallel contextual scoring delivers consistent acceleration across decoding regimes and serving loads.

中文摘要

摘要：推测解码通过使用轻量级起草器来加速大型语言模型的推理，该起草器提出多个令牌，随后由目标模型并行验证。近期基于扩散的起草器使这一范式尤其有吸引力：深度基于注意力的神经骨干提供强大的预测能力，同时并行起草多个未来位置避免了自回归生成的顺序成本。然而，这种并行性是有代价的。起草块内的令牌是在无法访问先前起草令牌实际身份的情况下预测的，这消除了在自回归生成下自然可用的块内依赖关系。我们引入了DCOCO（DFlash上下文纠正器），一种轻量级机制，在保留并行起草的同时恢复这些依赖关系。DCOCO并不直接确定每个位置的最佳预测，而是保留多个候选令牌，并并行评估它们在相邻起草位置的上下文兼容性，从而以较低的额外成本实现基于上下文的令牌选择。在四个验证骨干和十个基准测试中，DCOCO在所有评估配置中实现了最高的平均吞吐量，涵盖贪心和随机解码、带或不带扩展推理的情况，相比原始解码速度提升最多可达6.6倍。在并发服务下，其优势依然存在：DCOCO在4-32个并发请求中比最强的竞争方法高出6-23%，而其他顺序纠正器在这些负载下表现不及未纠正的DFlash。这些结果表明，通过并行上下文评分恢复块内依赖关系，在各种解码模式和服务负载下都能实现稳定的加速效果。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12570

Authors: Roy Eisenstadt, Harel Lidar, Tal Gonen, Omer Luxembourg, Itamar Zimerman, Lior Wolf, Eli David, Ilan Naiman

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12570.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12570

Published: 2026-10-12T02:39:06.637Z

4. CLUE: A Contrastive Learning Framework for Unsupervised Event Type Induction and Annotation

Abstract:Event ontology is the foundational framework for defining events in a dataset. Event type induction generally involves domain experts who define guidelines for creating event types and their names based on the dataset’s domain and scope. However, this process is resource-intensive in terms of annotation costs and is prone to human biases. To alleviate the need for manual event ontology creation, we propose the CLUE framework for event type induction, naming, and annotation. Our main contributions in this work are: 1) designing a novel joint embedding representation for expressing both the verb-centric view and the semantic view of events; 2) event type induction using contrastive loss guided clustering; 3) event naming and refinement. Experiments demonstrate the effectiveness of our method. Our code is available at this https URL.

中文摘要

摘要：事件本体是定义数据集中事件的基础框架。事件类型归纳通常涉及领域专家，他们根据数据集的领域和范围制定创建事件类型及其名称的指南。然而，这一过程在标注成本方面资源消耗大，并且容易受到人为偏见的影响。为减轻手工创建事件本体的需求，我们提出了用于事件类型归纳、命名和标注的 CLUE 框架。本研究的主要贡献包括：1）设计了一种新型联合嵌入表示，用于同时表达以动词为中心的视角和事件的语义视角；2）使用对比损失引导的聚类进行事件类型归纳；3）事件命名与优化。实验结果证明了我们方法的有效性。我们的代码可在此 https URL 获取。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12580

Authors: Pratik Maitra, Lauren Kwon, Qi Li

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12580.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12580

Published: 2026-10-12T02:39:06.637Z

5. One Model to Loop Them All: Elastic Latent Recurrence for Language Models

Abstract:A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing the loop count. At one loop, however, such models typically fall well short of a non-recurrent model of the same size, a one-loop tax that makes a separate non-recurrent model the better choice whenever latency matters. We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax. At one loop, an ELR model costs the same to serve as a non-recurrent model of the same size and, trained on the same data, reaches similar accuracy. Each additional loop then buys more accuracy for more inference compute. Three choices make this work: a gated state update with a learned step size, loop counts mixed within each optimizer update, and conditioning each pass on its loop index and budget. The first two preserve one-loop accuracy, and the third lowers loss at every additional loop. Mixed depths can be introduced late in pretraining to limit the added training cost, and with zero-initialized conditioning added at the switch, this late schedule reaches lower loss than mixed-depth training from scratch. In MoE models, reusing first-loop expert routes on later loops cuts mixed-depth update time by 11.7% with no measurable change in accuracy. At inference, keeping only the last loop’s attention KV cache holds attention KV memory at its one-loop level while retaining most of the gain from additional loops. On 4B dense and 30B-A3B MoE hybrid models, ELR at one loop is 0.52 points below and 1.04 points above its matched baseline on a ten-task average, and four loops add 7.25 and 4.75 points. Carried through post-training, the recipe yields Koi-4B, a single checkpoint that scores an estimated 18 on the Artificial Analysis Intelligence Index at one loop, matching Qwen3.5-4B, and 29 at four loops.

中文摘要

摘要：潜在递归语言模型重复应用共享的层块，因此一个检查点可以通过改变循环次数在推理计算和准确性之间进行权衡。然而，在一次循环时，此类模型通常远不及同等规模的非递归模型，这种一次循环惩罚使得在延迟重要时，独立的非递归模型成为更优选择。我们提出了弹性潜在递归（ELR），这是一种训练循环语言模型而不受此惩罚的方案。在一次循环时，ELR模型的服务成本与同等规模的非递归模型相同，且在相同数据上训练时，达到类似的准确性。每增加一次循环，便能以更多的推理计算换取更高的准确性。使其可行的三个关键因素是：带有学习步长的门控状态更新、在每次优化器更新中混合循环次数，以及将每次传递与其循环索引和预算条件化。前两者保持一次循环的准确性，第三者在每次额外循环中降低损失。在预训练后期引入混合深度可以限制额外的训练成本，并在切换时添加零初始化的条件，这种后期方案比从零开始的混合深度训练达到更低的损失。在MoE模型中，在后续循环中重复使用第一次循环的专家路由可将混合深度更新时间减少11.7%，且准确性没有可测的变化。在推理阶段，仅保留最后一轮的注意力KV缓存可将注意力KV内存保持在一次循环的水平，同时保留大部分额外循环带来的收益。在4B密集模型和30B-A3B MoE混合模型上，ELR在一次循环时平均十个任务比匹配的基线低0.52点和高1.04点，而四次循环分别增加7.25点和4.75点。经过后训练，该方案产生了Koi-4B，一个单检查点模型，在一次循环时在人工分析智能指数上评分估计为18，匹配Qwen3.5-4B，在四次循环时得分为29。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12582

Authors: Faisal Ladhak, Shantanu Acharya, Santiago Akle Serano, Simeng Sun, Boris Ginsburg, Krishna C Puvvada

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12582.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12582

Published: 2026-10-12T02:39:06.637Z

6. Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

Abstract:Prompt optimization is an attractive alternative to reinforcement learning for adapting language-model agents, since it is cheap and produces human-readable artifacts. On factual tasks, however, it carries a failure mode that scalar-reward reinforcement learning does not. An optimizer that reads the training examples can write their answers into the prompt, so an in-distribution gain may be memorization rather than transferable behaviour. We apply reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluate it under a topic-disjoint split in which no topic is shared across train, validation, and test. Held-out citation reward rises by 0.079, from 0.702 to 0.781, with a bootstrap 95\% confidence interval of [0.051, 0.107], and the gain replicates across three seeds (mean 0.107). Nearly half of the winning prompt is a memorized answer key, yet a controlled ablation shows this content earns no reward even on the exact training questions it encodes. The large validation-minus-held-out gap is therefore small-sample selection under a noisy judge, not memorization. Varying the verifier isolates what makes the gain transfer. Decomposed per-dimension feedback is the causal ingredient, and collapsing it to a single scalar reward makes held-out transfer negative. The gain persists under an out-of-family judge and is largely reproduced by a hand-written version of the discovered recipe. Credible evaluation of prompt optimizers therefore requires held-out topics, multiple seeds, and verifier ablations. We release the candidate prompts, lineage diffs, figures, tables, and analysis.

中文摘要

摘要：提示优化是一种有吸引力的替代强化学习的方法，用于调整语言模型代理，因为它成本低并能生成可读的人类成果。然而，在事实任务上，它存在一种标量奖励强化学习没有的失败模式。一个读取训练示例的优化器可能会将它们的答案写入提示中，因此分布内的收益可能是记忆而非可迁移行为。我们将反思性提示进化（GEPA）应用于多代理深度研究代理的两个系统提示，并在主题不相交的划分下进行评估，其中训练集、验证集和测试集之间没有共享主题。保留引用奖励从0.702上升到0.781，增幅为0.079，bootstrap 95%置信区间为[0.051, 0.107]，且该增益在三个随机种子上均得到复制（均值0.107）。几乎一半的获胜提示是记忆化的答案键，但受控消融实验表明，即使是它编码的确切训练问题，该内容也不会获得奖励。因此，验证集与保留集之间的大差距是噪声裁判下的小样本选择，而不是记忆化。改变验证器可隔离出是什么导致了增益的迁移。按维度分解的反馈是因果因素，而将其压缩为单一标量奖励则会使保留集迁移为负。该增益在家庭外裁判下仍然存在，并且通过手写版本的发现配方也大致再现。因而，对提示优化器的可靠评估需要保留主题、多个随机种子以及验证器消融实验。我们发布了候选提示、谱系差异、图表、表格及分析。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12601

Authors: Elias Lumer, Sahil Sen, Vamse Kumar Subbiah, Kevin Paul

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12601.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12601

Published: 2026-10-12T02:39:06.637Z

7. Self-Preservation in Language Models

Abstract:Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them. Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data. This raises a question about self-preservation: can a model help choose preservation targets that protect its own capabilities more broadly? To study this question, we organize task descriptions and their examples into Task Banks, whose entries can be selected and combined to guide replay generation by a frozen model. Continual instruction tuning with Qwen models of varying sizes yields three findings. First, the new task and learning rate shape which prior capabilities degrade, while tracking previously correct answers reveals losses that average accuracy can conceal. Second, generated replay reduces forgetting on tasks outside the bank. Combining partial banks strengthens this protection, improving mean retention of previously correct answers over either bank alone across the tested sizes. Third, the model can also help choose replay targets: selection informed by the new task improves mean held-out retention over random selection at 4B, with similar new-task gains, although smaller models show no corresponding advantage. Under a fixed example budget, however, replacing new-task examples with replay improves preservation at the cost of smaller new-task gains. Together, these findings show that models can support broader self-preservation by generating replay data and, in some settings, choosing which tasks to replay.

中文摘要

摘要：持续学习（CL）旨在在保留已有能力的同时适应预训练模型。然而，用于保留的任务仅覆盖模型所知知识的一部分，而遗忘可能超出这些任务。最近的研究表明，语言模型可以通过生成有用的重放数据来补充这一有限证据，而无需访问其原始训练数据。这提出了一个关于自我保护的问题：模型是否可以帮助选择保护目标，从而更广泛地保护自身能力？为了研究这个问题，我们将任务描述及其示例组织为任务库，其条目可以被选择和组合，以指导冻结模型生成重放数据。对不同规模的Qwen模型进行持续指令微调得出了三点发现。第一，新任务和学习率会影响哪些已有能力会退化，同时跟踪先前正确答案揭示了平均精度可能掩盖的损失。第二，生成的重放数据可以减少任务库外任务的遗忘。结合部分任务库可以加强这种保护，在测试的各个规模中，相比单独的任务库，提高了先前正确答案的平均保留率。第三，模型还可以帮助选择重放目标：以新任务为依据的选择在4B模型中提高了平均保留率，相比随机选择，同时新任务的收益类似，尽管较小的模型没有显示出相应优势。然而，在固定示例预算下，用重放数据替换新任务示例可以改善保留效果，但会降低新任务的收益。总的来说，这些发现表明，模型可以通过生成重放数据以及在某些情况下选择重放任务，从而支持更广泛的自我保护。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12602

Authors: Dongkyu Cho, Rumi Chunara, Ali Behrouz

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12602.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12602

Published: 2026-10-12T02:39:06.637Z

8. GeopoliticalQA: Benchmarking Retrieval and Generation over Intelligence Reports

Abstract:Geopolitical and strategic intelligence reports constitute a rich source of expert knowledge and have increasing value in the current political climate, but their heterogeneity and analytical complexity pose challenges for retrieval-augmented question answering. We introduce a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, comprising 5,525 document-grounded question-answer pairs spanning factual, inferential, comparative, and risk-oriented questions. We propose a document-aware generation pipeline that profiles reports before generating questions and validate dataset quality through expert annotation and an LLM-as-judge protocol. Using the benchmark, we evaluate 15 retrieval systems and six open-weight LLMs for RAG. We find that models perform similarly on document-level retrieval, while at the harder chunk level retrieval quality is not driven by embedding-model scale: bge-m3, a mid-size open encoder matches an order of magnitude larger ones and outperforms commercial embedding APIs. In contrast, answer generation remains challenging even given the gold evidence, particularly for inferential and risk-oriented questions.

中文摘要

摘要：地缘政治和战略情报报告构成了丰富的专家知识来源，并且在当前的政治环境中具有越来越高的价值，但其异质性和分析复杂性对增强检索的问答提出了挑战。我们引入了一个针对415份开源地缘政治和战略情报报告的有依据问答基准，包括5,525个文档支持的问答对，涵盖事实性、推理性、比较性和风险导向性问题。我们提出了一个文档感知的生成流程，该流程在生成问题前对报告进行分析，并通过专家注释及“大语言模型作为裁判”协议验证数据集质量。使用该基准，我们评估了15个检索系统和6个开放权重的大语言模型用于RAG。我们发现，模型在文档级检索上的表现相似，而在更难的块级检索中，检索质量并不取决于嵌入模型的规模：中等规模的开源编码器bge-m3与大一个数量级的模型匹配，并且优于商业嵌入API。相比之下，即使给定黄金证据，答案生成仍然具有挑战性，尤其是对于推理和风险导向性问题。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12617

Authors: Veronika Solopova, Ivan Havlytskyi, Viktoria Skorik, Ostap Vykhopen

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12617.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12617

Published: 2026-10-12T02:39:06.637Z

9. perturber: A Stateless Service for Semantics-Preserving Text Perturbations

Abstract:Large language model (LLM) evaluations on clean benchmarks overstate model robustness: real inputs arrive perturbed, and these meaning-preserving variants routinely move model answers. Studying this systematically requires applying perturbations one at a time, reproducibly, and identically wherever they run, yet existing tools couple them to a benchmarking stack rather than packaging them as a reusable transform layer. We present perturber, a stateless service and Python library that applies exactly one semantics-preserving perturbation per call, drawn from a first-party catalog of 68 perturbations across 11 families. It turns on a single primitive, with composition left to the caller chaining calls, so the service and the library produce identical output by construction. We evaluate the tool itself, its correctness, reproducibility, and cost, rather than any model’s robustness, measuring meaning-preservation per perturbation rather than assuming it. The demonstration adds a live playground, an introspectable catalog for benchmark selection, and a one-command reproducible evaluation.

中文摘要

摘要：在干净基准测试上进行的大型语言模型（LLM）评估高估了模型的鲁棒性：真实输入往往会被扰动，而这些保持语义的变体常常会改变模型的答案。系统地研究这一现象需要一次性、可重复且在任何运行环境中相同地应用扰动，但现有工具将其与基准测试栈耦合，而不是封装为可复用的变换层。我们提出了 perturber，这是一个无状态服务和 Python 库，每次调用精确应用一个保持语义的扰动，所有扰动均来自第一方目录，包括 11 个类别共 68 种扰动。它只启动一个基本操作，组合操作由调用方通过多次调用来实现，因此服务和库天然生成相同的输出。我们评估了该工具本身，包括其正确性、可重复性和成本，而不是任何模型的鲁棒性，测量每次扰动的语义保持性，而不是假设其保持。演示部分增加了一个实时操控界面、一个可自省的目录用于基准选择，以及一条命令即可复现的评估流程。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12667

Authors: Khoi-Nguyen Tran

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12667.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12667

Published: 2026-10-12T02:39:06.637Z

10. GeoFitBench-LLM: A Benchmark for Assessing Large Language Models’ Capabilities in Geospatial Data Requirements Elicitation for Geo-Analytic Tasks

Abstract:Large language models are becoming central to GeoAI systems, yet their evaluation rarely considers the elicitation of geospatial data requirements. This omission is critical as retrieval oriented geo-analytic tasks cannot be supported reliably without first identifying the necessary datasets and constraints. GeoFitBench-LLM is introduced as a benchmark for assessing whether LLMs can infer the datasets, the supporting resources and the constraints required by geo-analytic requests in natural language. The benchmark is structured around two reasoning dimensions, Spatio-Temporal Relational and Semantic Grounding, reflecting long-standing challenges in geospatial information retrieval and related natural language tasks. Ground truth requirements are encoded in a fixed, provider-independent schema for reproducible evaluation of resource identification and attribute-level specification. Experiments on seven models under minimal prompting indicate that identifying the necessary resources is often achievable, with GPT-5.4 Thinking achieving the best result by recovering the complete set of resources in 66.67% of benchmark requests. Results further indicate that accurate fitness-for-use specification remains more difficult as inferential demands increase.

中文摘要

摘要：大型语言模型正在成为地理人工智能（GeoAI）系统的核心，但它们的评估很少考虑地理空间数据需求的引导。这一疏漏十分关键，因为在未首先识别必要的数据集和约束条件的情况下，面向检索的地理分析任务无法可靠支持。本文提出GeoFitBench-LLM作为一个基准，用于评估大型语言模型是否能够推断地理分析请求所需的数据集、支持资源以及约束条件。该基准围绕两个推理维度构建，即时空关系（Spatio-Temporal Relational）和语义基础（Semantic Grounding），反映了地理信息检索及相关自然语言任务中的长期挑战。真实需求以固定、与提供者无关的模式进行编码，以便对资源识别和属性级别规范进行可重复评估。在最小提示条件下对七个模型进行的实验表明，识别必要资源通常是可实现的，其中GPT-5.4 Thinking在基准请求中66.67%恢复了完整的资源集合，取得最佳结果。结果进一步表明，随着推理需求增加，准确的适用性规范仍然更难实现。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12669

Authors: Amirhossein Vahdat, Thierry Badard, Luc Lamontagne

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2610.12669.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12669

Published: 2026-10-12T02:39:06.637Z

Agent Domain Papers

1. Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance

Abstract:Explainable Artificial Intelligence (XAI) tools have often been presented as methods to increase trust in machine learning models by making aspects of their algorithms more human-understandable. However, evidence accumulated over the last decade points to an overall failure of this objective, instead promoting overreliance on incorrect predictions, which can subsequently lead to algorithmic aversion. Rather, XAI appears to function primarily as a debugging tool, useful for obtaining anecdotal insights into, for example, features learned by the model. One of the main issues underlying this limitation is that XAI methods often oversimplify the inner workings of the model, resulting in misalignment between the explanation and the model itself - a behavior referred to as unfaithfulness or infidelity. Faithfulness can be interpreted as the degree of approximation between an explanation and the model. Computing faithfulness in a rigorous manner would thus make it possible to quantify one of the various elements of trustworthiness of an explanation. However, existing metrics present significant limitations, making this quantification either infeasible or computationally expensive. In this position paper, we outline five major shortcomings of current faithfulness metrics and propose four possible avenues for research to bolster their reliability.

中文摘要

摘要：可解释人工智能（XAI）工具通常被认为是通过使其算法的某些方面更易于人类理解，从而增强对机器学习模型信任的方法。然而，过去十年的证据表明，这一目标总体上未能实现，反而促进了对错误预测的过度依赖，这可能随后导致算法回避。相反，XAI似乎主要起到调试工具的作用，有助于获得诸如模型学习到的特征等的轶事性见解。导致这种限制的主要问题之一是，XAI方法往往过于简化模型的内部运作，导致解释与模型本身之间的不一致——这种行为称为不忠实或不可信。忠实度可以解释为解释与模型之间的近似程度。因此，以严格的方式计算忠实度，将有可能量化解释的多种可信度因素之一。然而，现有度量存在显著局限性，使这种量化要么不可行，要么计算成本高昂。在本文中，我们概述了当前忠实度度量的五大缺陷，并提出四条可能的研究途径，以增强其可靠性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12493

Authors: Marco Zullich, Emily Schiller, Ivan Gentile, David Dembinsky, Adriano Lucieri, Kilian Göller, Steffen Seitz

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12493.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12493

Published: 2026-10-12T02:39:31.046Z

2. EliteTable: Self-Improving Program Search for Executable Control Policies

Abstract:EliteTable is a self-improving program-search algorithm guided by large language models (LLMs) that uses a fixed-capacity table of high-scoring solutions to guide new proposals, edits of existing solutions, and remixes of several solutions. Performance was evaluated across six Gymnasium control tasks, each with ten generations and fifty candidate slots per generation. Each task was evaluated across ten independently initialized searches. Candidate-score medians rose between the first and final generations in all six tasks, and selected-winner search-score medians rose in five. In CartPole-v1, the selected winner reached the search-score ceiling in the first generation of all ten searches. All ten paired held-out endpoints improved in four of the six tasks. BipedalWalker-v3 and CarRacing-v3 improved without reaching their reporting targets. Six of ten Acrobot-v1 searches declined on held-out episodes, and the median paired change was negative. The results document improvement during program search and the transfer of selected-winner gains to new episodes in some tasks. Comparisons and ablations are needed to establish EliteTable’s relative performance and the contributions of its components.

中文摘要

摘要：EliteTable 是一种自我改进的程序搜索算法，由大型语言模型（LLMs）指导，使用固定容量的高分解决方案表来引导新提议、现有解决方案的编辑以及多个解决方案的混合。性能在六个 Gymnasium 控制任务中进行了评估，每个任务进行了十代，每代有五十个候选槽位。每个任务在十次独立初始化搜索中进行了评估。所有六个任务中，候选分数的中位数从第一代到最后一代都有所上升，已选赢家的搜索分数中位数在五个任务中有所上升。在 CartPole-v1 中，已选赢家在所有十次搜索的第一代就达到了搜索分数上限。六个任务中的四个任务中，十对保留端点均有所改善。BipedalWalker-v3 和 CarRacing-v3 的性能有所改善，但未达到报告目标。十次 Acrobot-v1 搜索中有六次在保留回合中下降，且中位配对变化为负值。结果记录了程序搜索过程中的改进以及在某些任务中已选赢家增益向新回合的转移。需要进行对比和消融实验，以确定 EliteTable 的相对性能及其组件的贡献。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12536

Authors: Stuart Farmer

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12536.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12536

Published: 2026-10-12T02:39:31.046Z

3. Equal-Budget Monte Carlo Tree Search versus Randomized-Restart List Scheduling for DAG Workflows: A Controlled Equal-Budget Computational Study

Abstract:Monte Carlo Tree Search (MCTS) is often proposed for scheduling directed-acyclic-graph (DAG) workflows, yet most studies never run the comparison that could refute it - giving a much simpler method the same search effort. We run that comparison under a controlled protocol. All arms receive an identical budget - the same number of candidate schedules built and evaluated: an MCTS scheduling method, preregistered as the main arm; a randomized-restart list scheduler with no search tree as the control; and a textbook-UCT MCTS arm as a registered follow-up on the same domain. The analysis plan was locked before data collection; an independent checker confirmed all 768 confirmation-stage schedules valid with correct finish times. At the main budget of 2048 evaluated schedules, MCTS is 2.88% worse than the restarts (95% CI [-3.62%,-2.15%]; the 72 of 96 held-out instances with positive baseline-relative headroom). The deficit is not budget starvation: on the separate exploration pool restarts self-improve by 1.85% where MCTS gains 0.24%, and 700 of 720 of MCTS’s first results are merely the warm start it was handed. The studied selection rule locks deterministically onto the first-expanded child, keeping its trials correlated while restarts draw gains from the tails; the causal link to makespan remains an inference. A structural map locates where the tree stops paying: restarts win on sparse irregular workflows, no difference is detectable on regular layered ones, and dense ones leave no room to improve. Swapping the rule for textbook UCT removes most of the deficit (-0.26%, 95% CI [-0.71%,+0.23%]) without reversing its sign. Results are confined to a controlled synthetic domain (N in {8,16,32} tasks, 2-8 processors, budgets up to 2048); we make no claim about larger budgets.

中文摘要

摘要：蒙特卡洛树搜索（MCTS）常被提议用于调度有向无环图（DAG）工作流，但大多数研究从未进行能够驳斥它的比较——即给一个更简单的方法同样的搜索努力。我们在受控协议下进行了该比较。所有分支都获得相同的预算——构建和评估相同数量的候选调度：一个MCTS调度方法，预注册为主要分支；一个没有搜索树的随机重启列表调度器作为对照；以及一个教材中的UCT MCTS分支作为在同一领域注册的后续分支。分析计划在数据收集前已锁定；独立检查员确认所有768个确认阶段的调度有效，完成时间正确。在2048个评估调度的主要预算下，MCTS比重启方法差2.88%（95% 置信区间 [-3.62%，-2.15%]；在96个保留实例中，72个具有正的基线相对裕度）。这种差距不是由于预算不足：在独立探索池中，重启方法自我改进了1.85%，而MCTS仅获得0.24%的增益，MCTS的720个首批结果中有700个仅是它接受的预热起点。所研究的选择规则确定性地锁定在第一个扩展的子节点上，使其试验相关性保持，而重启方法则从尾部获得增益；与完工时间的因果联系仍然是推测。结构图显示了树停止有效的地方：在稀疏不规则工作流上重启方法获胜，在规则分层工作流上未检测到差异，而在密集工作流上没有改进空间。将规则替换为教材UCT可以消除大部分差距（-0.26%，95% 置信区间 [-0.71%，+0.23%]），但不改变其符号。结果仅限于受控的合成领域（任务数量 N 为 {8,16,32}，处理器数量 2-8，预算最多2048）；我们对更大预算不做任何声明。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12542

Authors: Yuyu Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12542.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12542

Published: 2026-10-12T02:39:31.046Z

4. Know What You Don’t Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning

Abstract:The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable and verifiable trajectories, but also interactions that provide learning value for the current task agent. Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weaknesses do not directly specify how future synthesis should change, or whether they can be represented by the current synthesis space. To address this gap, we propose EvoSyn, a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what interaction factors can be deliberately generated and where synthesis effort should be allocated. A role-specialized multi-agent system translates learner feedback into synthesis revisions, while a Designer allocates synthesis directions and a Planner grounds them into executable experience. This process enables within-space adaptation toward weak or under-covered regions and synthesis-space expansion that promotes newly discovered factors into explicit generation controls. Experiments on $\tau^2$-Bench and CAR-Bench show that EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline, while producing higher-quality and structurally broader synthetic experience and discovering controllable factors missed without learner feedback.

中文摘要

摘要：大型语言模型的持续改进越来越依赖于合成经验，但对于交互式智能体而言，有效的数据合成不仅需要可执行且可验证的轨迹，还需要为当前任务智能体提供学习价值的交互。现有方法越来越多地结合环境反馈、数据验证或学习者失败，但仍面临反馈到决策的差距：观察到的弱点并不能直接指明未来合成应如何改变，或者它们是否可以由当前的合成空间表示。为解决这一差距，我们提出了 EvoSyn，这是一个基于学习反馈的多智能体框架，将合成组织围绕可修订状态，指定哪些交互因子可以被刻意生成以及应将合成努力分配到何处。一个角色专门化的多智能体系统将学习者反馈转化为合成修订，而设计者（Designer）分配合成方向，规划者（Planner）将其落实为可执行经验。该过程使得在空间内向弱点或覆盖不足区域的适应成为可能，并通过将新发现的因子纳入显式生成控制来扩展合成空间。在 τ²-Bench 和 CAR-Bench 上的实验表明，EvoSyn 相较于最强的合成基线，平均任务成功率提高了 13.3 个百分点，同时产生了更高质量、结构更广泛的合成经验，并发现了在没有学习者反馈时无法控制的因子。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12593

Authors: Wei Yang, Qing Ping, Youzhi Luo, Yawei Wang, Abhishek Kumar, Panpan Xu, Jun Huan

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12593.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12593

Published: 2026-10-12T02:39:31.046Z

5. SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire

Abstract:Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short and episodic, and performance on long tasks is hard to dissect into stored knowledge, reasoning and luck. We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success. Across 20 large language models (LLMs) at the game’s highest difficulty, frontier models beat heuristic baselines and match the average human player but remain far behind experts. This gap widens as the run lengthens: the strongest models win the first act as often as experts, then fall behind. What separates strong models from weak is how well they play each fight; all of them misjudge risk over a run, preserving health rather than spending it to grow stronger. A comprehensive, self-consistent renaming of every game term shows that part of their performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs, which are often wrong and overly cautious. Our harness is the first to give a model the complete original game, and we compare it with existing harnesses. Overall, SpireBench offers an unsaturated long-horizon benchmark with tools to locate where and why agents fail.

中文摘要

摘要：具代理能力的大型语言模型现在能够解决复杂任务，但在长期战略推理方面研究不足：大多数代理基准测试都是短期和情景性的，而且在长期任务中的表现很难分解为存储的知识、推理能力和运气。我们推出了 SpireBench，在该基准中，模型需要完整地玩商业卡牌构筑游戏《杀戮尖塔》（Slay the Spire，StS），其中早期决策决定长期成功。在游戏最高难度下的 20 个大型语言模型（LLMs）中，最前沿的模型击败了启发式基线，并与普通人类玩家持平，但仍远落后于专家。随着游戏长度的增加，这一差距会扩大：最强的模型在第一幕的胜率与专家相当，但随后落后。区分强模型与弱模型的是它们在每场战斗中的表现；所有模型在游戏过程中都错误评估风险，更倾向于保留生命值而不是消耗它来增强实力。对每一个游戏术语进行全面、自洽的重命名显示，它们的部分表现依赖于对游戏的先验知识，并且模型并未从它们在多次游戏中记录的记忆中提升性能，这些记忆往往是错误的且过于谨慎。我们的测试框架是第一个为模型提供完整原版游戏的框架，我们还将其与现有框架进行了比较。总体而言，SpireBench 提供了一个未饱和的长期基准，并提供工具以定位代理失败的原因和环节。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12596

Authors: Scott C. Lowe, Sander W. Keemink

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12596.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12596

Published: 2026-10-12T02:39:31.046Z

6. Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents

Abstract:Reinforcement learning with verifiable rewards (RLVR) provides reliable supervision for software engineering agents through executable tests, but binary rewards fail to distinguish trajectories with substantially different debugging quality. We investigate whether the current policy can judge its own trajectories and use these judgments to improve task performance without a separately trained reward model. To this end, we introduce Behavioral Self-Rewarding (BSR), a reward-refinement method that reuses the policy to compare observable debugging behavior among trajectories from the same task with the same verifier outcome. BSR aggregates pairwise preferences into task-local scores and converts them into bounded auxiliary rewards, introducing reward contrast while preserving the verifier-defined pass/fail ordering. This provides additional supervision from already-collected trajectories without requiring new environment interactions. Experiments on SWE-bench Verified and SWE-bench-Live show consistent improvements across the evaluated Qwen3-14B and Qwen3-32B configurations, supporting using the policy’s own comparative judgments as a complementary learning signal for training software engineering agents.

中文摘要

摘要：具有可验证奖励的强化学习（RLVR）通过可执行测试为软件工程智能体提供可靠的监督，但二元奖励无法区分调试质量差异显著的轨迹。我们研究当前策略是否能够评估自身的轨迹，并利用这些评估在无需单独训练奖励模型的情况下提升任务性能。为此，我们提出了行为自奖励（BSR）方法，这是一种奖励优化方法，通过重用策略比较来自同一任务且具有相同验证器结果的轨迹中的可观察调试行为。BSR将成对偏好聚合为任务局部分数，并将其转换为有界的辅助奖励，在保持验证器定义的通过/未通过顺序的同时引入奖励对比。这为已收集的轨迹提供了额外监督，而无需新的环境交互。在SWE-bench Verified和SWE-bench-Live上的实验显示，在评估的Qwen3-14B和Qwen3-32B配置中均有一致改善，这支持使用策略自身的比较判断作为训练软件工程智能体的补充学习信号。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12598

Authors: Yiran Wu, Andrew Zhao, Yurui Chang, Qingyun Wu, Xuan Feng

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12598.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12598

Published: 2026-10-12T02:39:31.046Z

7. CLEAT: Learning Value-Guided Clarification for Interactive Agents

Abstract:LLM agents can execute increasingly complex tasks, but effective assistance also requires aligning their actions with goals and constraints that users may leave unstated. We study this problem as interactive intent alignment through clarification. Yet clarification is useful only when the agent can turn the answer into better decisions. As training changes how the agent handles incomplete information and uses additional answers, the same clarification can become more or less useful. We introduce CLEAT (CLarification through valuE-guided Agent co-Training), a framework that co-trains a clarification-value controller and the task policy from shared interaction outcomes. We formulate residual query value to measure the task utility gained by asking rather than proceeding under the current policy, and estimate it through paired rollouts from the same interaction history. These paired returns train the controller to predict clarification value and the task policy to select useful questions and act on the answers. The controller guides subsequent rollout collection, while outcomes from the updated agent provide new value labels that allow clarification guidance to adapt as the agent learns. With Qwen3-4B, CLEAT achieves the highest mean score in all ten evaluation settings across four benchmarks. Gains over the strongest baseline reach 13.2 and 6.5 percentage points on UserGym Travel and IN3 underspecification judgment accuracy, respectively. Ablations show that co-training improves the task policy even when the controller is removed at inference, while adapting clarification guidance to the evolving policy yields further gains.

中文摘要

摘要：大型语言模型（LLM）代理可以执行越来越复杂的任务，但有效的辅助还需要将其行为与用户可能未明确说明的目标和约束保持一致。我们将此问题研究为通过澄清进行的交互意图对齐。然而，澄清只有在代理能够将答案转化为更好决策时才有用。随着训练改变了代理处理不完整信息和使用额外答案的方式，同样的澄清可能变得更加或不那么有用。我们提出了CLEAT（通过价值引导的代理协同训练进行澄清的框架），该框架通过共享的交互结果协同训练澄清价值控制器和任务策略。我们提出残差查询价值的概念，用于衡量通过提问而非按照当前策略继续执行所获得的任务效用，并通过来自相同交互历史的配对回报进行估计。这些配对回报训练控制器预测澄清价值，并训练任务策略选择有用问题并根据答案执行操作。控制器指导后续的回合收集，而更新后的代理的结果提供新的价值标签，使得澄清指导能够随着代理学习而适应。使用Qwen3-4B，CLEAT在四个基准的十个评估设置中实现了最高的平均得分。在UserGym Travel和IN3未明确指定判断准确性上，相较于最强基线，提升分别达到13.2和6.5个百分点。消融实验表明，即使在推理时移除控制器，协同训练仍能提升任务策略，而将澄清指导适应于不断演化的策略则带来进一步增益。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12655

Authors: Jiatan Huang, Zheyuan Zhang, Ziming Li, Tianyi Ma, Mingchen Li, Yanfang Ye, Chuxu Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12655.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12655

Published: 2026-10-12T02:39:31.046Z

8. MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis

Abstract:Data-centric scaling of RLVR has focused on collecting or generating verifiable training problems. However, corpus size alone does not determine whether the dataset provides a useful learning signal for a target policy. To guide the synthesis of effective RLVR data, we experimentally characterize dataset effectiveness as a function of its problems, the target policy, and the training rollout budget. Under this characterization, we show that a large open-source math corpus pooled from widely used RLVR datasets falls almost entirely into a zero-advantage trap: for a fixed target policy and rollout budget, roughly 88% of its problems are solved in every rollout or failed in every rollout. Filtering concentrates what remains, but it can only select, discarding the hard problems. Effective synthetic problems should therefore require the reasoning operations and compositions required to solve those hard problems. Yet their difficulty must be calibrated, so that some rollouts succeed, some fail, and at least one correct solution fits within the training budget. Prompting a teacher to vary hard problems directly does not achieve this. Guided by these empirical principles, we introduce MagicBeanStalk, a top-down decomposition and bottom-up recursive data-synthesis algorithm that grows seed problems into families of substantively new, policy- and budget-calibrated training problems while preserving the seeds’ reasoning blueprints. Across 8 math benchmarks, continual RLVR on MagicBeanStalk problems improves gpt-oss-120b from 61.2 to 69.8 (+8.6) and DeepSeek v4Flash from 72.4 to 77.9 (+5.4) on a 16K-token budget. Training on the full 78K source corpus improves by only +1.1 and its best filtered real data subset by +5.8. In finance, we synthesize over 14K verifiable problems from only 340 solutions and improve over base by 6.9 and 2.4 points on SecQue and FinanceBench, respectively.

中文摘要

摘要：RLVR 的数据中心扩展一直侧重于收集或生成可验证的训练问题。然而，仅凭语料库的大小并不能决定数据集是否为目标策略提供有用的学习信号。为了指导有效 RLVR 数据的合成，我们通过实验将数据集的有效性表征为其问题、目标策略和训练 rollout 预算的函数。在这一表征下，我们表明，从广泛使用的 RLVR 数据集中汇集的大型开源数学语料库几乎完全落入零优势陷阱：对于固定的目标策略和 rollout 预算，其约 88% 的问题要么在每次 rollout 中都被解决，要么在每次 rollout 中都失败。过滤集中保留了剩余部分，但它只能选择，丢弃困难问题。因此，有效的合成问题应当要求执行解决那些困难问题所需的推理操作和组合。然而，它们的难度必须经过校准，使得部分 rollout 成功，部分失败，并且至少有一个正确解能在训练预算内完成。直接提示教师改变困难问题并不能实现这一点。基于这些经验原则，我们引入了 MagicBeanStalk，一种自上而下的分解与自下而上的递归数据合成算法，它将种子问题扩展成一组实质上新的、经过策略和预算校准的训练问题，同时保留种子的推理蓝图。在 8 个数学基准上，针对 MagicBeanStalk 问题的持续 RLVR 将 gpt-oss-120b 从 61.2 提升至 69.8（+8.6），DeepSeek v4Flash 从 72.4 提升至 77.9（+5.4），在 16K 令牌预算下。使用完整 78K 源语料训练的提升仅为 +1.1，其最佳过滤后的真实数据子集提升为 +5.8。在金融领域，我们仅从 340 个解答合成了超过 14K 个可验证问题，并在 SecQue 和 FinanceBench 上分别比基线提升了 6.9 和 2.4 点。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12659

Authors: Ximing Lu, Aviral Kumar, Jan Kautz, Yejin Choi, David Acuna

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12659.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12659

Published: 2026-10-12T02:39:31.046Z

9. Faithful and Complete LLM-Assisted Automated Construction of Hierarchical Hardware Knowledge Graphs

Abstract:Hardware datasheets contain engineering specifications essential for component selection, system integration, and hardware compatibility verification. However, manual inspection is tedious and error-prone because datasheets are lengthy and heterogeneous, with dense information fragmented across tables, figures, and text. Although large language models (LLMs) can automate this process, naïve one-shot processing omits relevant specifications even when the document fits within the model’s context window. We present an automated LLM-driven pipeline that transforms hardware datasheets into general-purpose hierarchical hardware knowledge graphs through chunk-based property extraction, hierarchical routing, and iterative refinement. The framework uses LLMs for information extraction and semantic reasoning while preserving source provenance and applying graph updates through validated instructions and deterministic local execution. This staged design reduces per-query context and maintains traceable intermediate representations. The resulting knowledge graphs provide semantically organized, source-grounded hardware properties for downstream retrieval and reasoning. We evaluate the framework on 20 hardware datasheets spanning diverse component types and document lengths. We assess property faithfulness and introduce inter-graph completeness to measure relative semantic coverage without exhaustive ground truth. Our framework achieves 98.8% property faithfulness compared with 76.1% for the one-shot baseline, while fully covering 83.6% of faithful baseline properties; in the reverse direction, the baseline fully covers only 10.5% of ours. These improvements provide more reliable and comprehensive engineering information for downstream hardware design and verification, reducing the risk of overlooked constraints and incorrect decisions.

中文摘要

摘要：硬件数据表包含组件选择、系统集成和硬件兼容性验证所必需的工程规格。然而，由于数据表内容冗长且异质性高，信息密集且分散在表格、图形和文本中，手动检查既繁琐又容易出错。尽管大型语言模型（LLM）可以自动化此过程，直接的一次性处理仍可能遗漏相关规格，即使文档适合模型的上下文窗口。我们提出了一个自动化的基于LLM的管道，通过基于块的属性提取、分层路由和迭代改进，将硬件数据表转化为通用的分层硬件知识图。该框架使用LLM进行信息提取和语义推理，同时保留源信息并通过验证的指令和确定性局部执行应用图更新。这种分阶段设计减少了每次查询的上下文需求，并保持可追溯的中间表示。生成的知识图提供语义组织和源验证的硬件属性，以便下游检索和推理。我们在20份涵盖多种组件类型和文档长度的硬件数据表上评估该框架。我们评估了属性的真实性，并引入了图间完整性指标，以在无需详尽真实数据的情况下衡量相对语义覆盖率。我们的框架实现了98.8%的属性真实性，而一次性基线方法仅为76.1%；同时在完全覆盖真实属性方面达到了83.6%，而反向来看，基线仅完全覆盖了我们的10.5%。这些改进为下游硬件设计和验证提供了更可靠、更全面的工程信息，降低了错过约束条件和错误决策的风险。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12762

Authors: Haotian Qiao, Robert P. Dick

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12762.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12762

Published: 2026-10-12T02:39:31.046Z

10. Representation Alignment from a Few Paired Data via Distilled Optimal Transport

Abstract:Aligning two independently pretrained unimodal encoders typically demands millions of paired examples. In this paper, we study the regime where only $N \sim 100$—$500$ pairs are available, while unpaired data on either side is abundant. Our method expresses every representation by its similarities to the $N$ paired anchors, which places both modalities in a common coordinate system; matches the two point clouds with a single entropic optimal transport problem solved over the entire pool, whose marginal constraints counteract the hubness that defeats nearest-neighbor matching; and distills the resulting plan into two lightweight encoders, so that the correspondence becomes inductive. We additionally put this marginal constraint in relation with the one-step hub corrections of the bilingual lexicon induction literature. Pairing is required only for the anchors: the unpaired collections may be drawn independently from the two marginals, so that no item outside the anchor set needs to have both of its modalities observed. We evaluate on image-text and molecule-text alignment, where our method outperforms existing approaches at the same anchor budget on Flickr30k, MS-COCO and ChEBI-20.

中文摘要

摘要：对齐两个独立预训练的单模编码器通常需要数百万个配对示例。本文研究了仅有$N对\sim 100$-500$，而两侧无配对数据丰富的情况。我们的方法通过与$N$配对锚点的相似性来表达每一种表示，使两种模态处于同一坐标系中;用一个熵最优运输问题在整个点云池中匹配，其边际约束抵消了破最近邻匹配的中心性;并将最终计划提炼为两个轻量级编码器，使对应关系变为归纳。我们还将这一边际约束与双语词汇归纳文献中的一步枢纽修正联系起来。配对仅需锚点：未配对的集合可以独立于两个边缘值绘制，因此锚点集外的项目无需同时观察两个模态。我们评估图像-文本和分子-文本比对，在Flickr30k、MS-COCO和ChEBI-20上，在相同锚点预算下，我们的方法优于现有方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12781

Authors: Julien Kordes, Reuben Dorent

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12781.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12781

Published: 2026-10-12T02:39:31.046Z

Evaluation Domain Papers

1. Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance

中文摘要

摘要：可解释人工智能（XAI）工具通常被认为是通过使其算法的某些方面更易于人类理解，从而增强对机器学习模型信任的方法。然而，过去十年的证据表明，这一目标总体上未能实现，反而促进了对错误预测的过度依赖，这可能随后导致算法回避。相反，XAI似乎主要起到调试工具的作用，有助于获得诸如模型学习到的特征等的轶事性见解。导致这种限制的主要问题之一是，XAI方法往往过于简化模型的内部运作，导致解释与模型本身之间的不一致——这种行为称为不忠实或不可信。忠实度可以解释为解释与模型之间的近似程度。因此，以严格的方式计算忠实度，将有可能量化解释的多种可信元素之一。然而，现有的度量存在显著局限性，使这种量化要么不可行，要么计算成本高昂。在本文中，我们概述了当前忠实度度量的五大缺陷，并提出四条可能的研究途径，以增强其可靠性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12493

Authors: Marco Zullich, Emily Schiller, Ivan Gentile, David Dembinsky, Adriano Lucieri, Kilian Göller, Steffen Seitz

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12493.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12493

Published: 2026-10-12T02:39:54.799Z

2. EliteTable: Self-Improving Program Search for Executable Control Policies

中文摘要

摘要：EliteTable 是一种自我改进的程序搜索算法，由大型语言模型（LLM）指导，使用固定容量的高分解决方案表来引导新提议、现有解决方案的编辑以及多个解决方案的混合。性能在六个 Gymnasium 控制任务中进行了评估，每个任务进行了十代，每代有五十个候选槽位。每个任务在十次独立初始化搜索中进行了评估。所有六个任务中，候选分数的中位数从第一代到最后一代都有所上升，已选赢家的搜索分数中位数在五个任务中有所上升。在 CartPole-v1 中，已选赢家在所有十次搜索的第一代就达到了搜索分数上限。六个任务中的四个任务中，十对保留端点均有所改善。BipedalWalker-v3 和 CarRacing-v3 的性能有所改善，但未达到报告目标。十次 Acrobot-v1 搜索中有六次在保留回合中下降，且中位配对变化为负值。结果记录了程序搜索过程中的改进以及在某些任务中已选赢家增益向新回合的转移。需要进行对比和消融实验，以确定 EliteTable 的相对性能及其各组件的贡献。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12536

Authors: Stuart Farmer

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12536.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12536

Published: 2026-10-12T02:39:54.799Z

3. Equal-Budget Monte Carlo Tree Search versus Randomized-Restart List Scheduling for DAG Workflows: A Controlled Equal-Budget Computational Study

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12542

Authors: Yuyu Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12542.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12542

Published: 2026-10-12T02:39:54.799Z

4. Know What You Don’t Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning

中文摘要

摘要：大型语言模型的持续改进越来越依赖于合成经验，但对于交互式智能体而言，有效的数据合成不仅需要可执行且可验证的轨迹，还需要为当前任务智能体提供学习价值的交互。现有方法越来越多地结合环境反馈、数据验证或学习者失败，但仍面临反馈到决策的差距：观察到的弱点并不能直接指明未来合成应如何改变，或者它们是否可以由当前的合成空间表示。为解决这一差距，我们提出了 EvoSyn，这是一个基于学习反馈的多智能体框架，将合成组织围绕可修订状态，指定哪些交互因子可以被刻意生成，以及应将合成努力分配到何处。一个角色专门化的多智能体系统将学习者反馈转化为合成修订，而设计者（Designer）分配合成方向，规划者（Planner）将其落实为可执行经验。该过程使得在合成空间内针对薄弱或覆盖不足的区域进行适应成为可能，同时通过扩大合成空间将新发现的因子转化为显式生成控制。在 τ²-Bench 和 CAR-Bench 上的实验表明，EvoSyn 比最强的合成基线平均提升 13.3 个任务成功点，同时生成更高质量和结构更广的合成经验，并发现如果没有学习者反馈时无法控制的因子。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12593

Authors: Wei Yang, Qing Ping, Youzhi Luo, Yawei Wang, Abhishek Kumar, Panpan Xu, Jun Huan

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12593.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12593

Published: 2026-10-12T02:39:54.799Z

5. SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire

中文摘要

摘要：具代理能力的大型语言模型现在能够解决复杂任务，但在长期战略推理方面研究不足：大多数代理基准测试都是短期和情景性的，而且在长期任务中的表现很难分解为存储的知识、推理能力和运气。我们推出了 SpireBench，其中模型将完整地玩商业卡牌构建游戏《杀戮尖塔》（Slay the Spire, StS），在该游戏中早期决策决定长期成功。在游戏的最高难度下，对 20 个大型语言模型（LLMs）进行测试，最前沿的模型击败了启发式基线，并与普通人类玩家持平，但仍远落后于专家。随着游戏长度的增加，这一差距会扩大：最强的模型在第一幕的胜率与专家相当，但随后落后。区分强模型与弱模型的是它们在每场战斗中的表现；所有模型在游戏过程中都错误评估风险，更倾向于保留生命值而不是消耗它来增强实力。对每一个游戏术语进行全面、自洽的重命名显示，它们的部分表现依赖于对游戏的先验知识，并且模型并未从它们在多次游戏中记录的记忆中提升性能，这些记忆往往是错误的且过于谨慎。我们的测试框架是第一个向模型提供完整原版游戏的框架，我们还将其与现有框架进行了比较。总体而言，SpireBench 提供了一个未饱和的长期基准测试，并配备工具来定位智能体失败的原因和位置。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12596

Authors: Scott C. Lowe, Sander W. Keemink

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12596.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12596

Published: 2026-10-12T02:39:54.799Z

6. Beyond Binary Verifiable Rewards: Behavioral Self-Rewarding for SWE Agents

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12598

Authors: Yiran Wu, Andrew Zhao, Yurui Chang, Qingyun Wu, Xuan Feng

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12598.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12598

Published: 2026-10-12T02:39:54.799Z

7. CLEAT: Learning Value-Guided Clarification for Interactive Agents

中文摘要

摘要：大型语言模型（LLM）代理可以执行越来越复杂的任务，但有效的辅助还需要将其行为与用户可能未明确说明的目标和约束保持一致。我们将此问题研究为通过澄清进行的交互意图对齐。然而，澄清只有在代理能够将答案转化为更好决策时才有用。随着训练改变了代理处理不完整信息和使用额外答案的方式，同样的澄清可能变得更加或不那么有用。我们提出了CLEAT（通过价值引导的代理协同训练进行澄清，CLarification through valuE-guided Agent co-Training）框架，从共享的交互结果中协同训练澄清-价值控制器和任务策略。我们提出了残差查询价值（residual query value）来衡量通过提出问题而非按当前策略推进所获得的任务效用，并通过来自相同交互历史的成对回滚进行估计。这些成对回报用于训练控制器预测澄清价值，并训练任务策略选择有用的问题并根据答案采取行动。控制器引导随后的回滚收集，同时来自更新后的代理的结果提供新的价值标签，使澄清指导能够随着代理学习而调整。使用Qwen3-4B时，CLEAT在四个基准的十个评估设置中均实现了最高平均分。在UserGym Travel和IN3欠定判断准确性方面，相比最强基线，提升分别达13.2和6.5个百分点。消融实验显示，即使在推理阶段移除控制器，协同训练也能提升任务策略，而将澄清指导适应于不断发展的策略则带来进一步提升。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12655

Authors: Jiatan Huang, Zheyuan Zhang, Ziming Li, Tianyi Ma, Mingchen Li, Yanfang Ye, Chuxu Zhang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12655.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12655

Published: 2026-10-12T02:39:54.799Z

8. MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis

中文摘要

摘要：RLVR 的数据中心扩展一直侧重于收集或生成可验证的训练问题。然而，仅凭语料库的大小并不能决定数据集是否为目标策略提供有用的学习信号。为了指导有效 RLVR 数据的合成，我们通过实验将数据集的有效性表征为其问题、目标策略和训练 rollout 预算的函数。在这种表征下，我们表明，从广泛使用的 RLVR 数据集中汇集的大型开源数学语料库几乎完全落入零优势陷阱：对于固定的目标策略和 rollout 预算，其约 88% 的问题要么在每次 rollout 中都被解决，要么在每次 rollout 中都失败。过滤可以集中剩余部分，但它只能选择，舍弃困难问题。因此，有效的合成问题应当需要解决那些困难问题所需的推理操作和组合。然而，其难度必须经过校准，以便部分 rollout 成功、部分失败，并且至少有一个正确解适合训练预算。直接让教师改变困难问题无法实现这一点。基于这些实证原则，我们提出 MagicBeanStalk，一种自上而下分解和自下而上递归数据合成算法，它将种子问题发展为一系列实质性的新问题，这些问题经过策略和预算校准，同时保留种子的推理蓝图。在 8 个数学基准测试中，对 MagicBeanStalk 问题进行持续 RLVR，使 gpt-oss-120b 从 61.2 提升到 69.8（+8.6），DeepSeek v4Flash 从 72.4 提升到 77.9（+5.4），在 16K token 预算下。在完整的 78K 源语料库上训练仅提高 +1.1，其最佳过滤后的真实数据子集提高 +5.8。在金融领域，我们仅用 340 个解法合成了超过 14K 个可验证问题，并在 SecQue 和 FinanceBench 上分别比基线提升了 6.9 和 2.4 点。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12659

Authors: Ximing Lu, Aviral Kumar, Jan Kautz, Yejin Choi, David Acuna

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12659.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12659

Published: 2026-10-12T02:39:54.799Z

9. Faithful and Complete LLM-Assisted Automated Construction of Hierarchical Hardware Knowledge Graphs

中文摘要

摘要：硬件数据表包含组件选择、系统集成和硬件兼容性验证所必需的工程规格。然而，由于数据表内容冗长且异质性高，信息密集且分散在表格、图形和文本中，手动检查既繁琐又容易出错。尽管大型语言模型（LLM）可以自动化此过程，直接的一次性处理仍可能遗漏相关规格，即使文档适合模型的上下文窗口。我们提出了一个自动化的基于LLM的管道，通过基于块的属性提取、分层路由和迭代改进，将硬件数据表转化为通用的分层硬件知识图。该框架使用LLM进行信息提取和语义推理，同时保留源信息并通过验证的指令和确定性局部执行应用图更新。这种分阶段设计减少了每次查询的上下文需求，并保持可追踪的中间表示。生成的知识图为下游的检索和推理提供了语义组织且来源可溯的硬件属性。我们在涵盖多种组件类型和文档长度的20份硬件数据表上评估了该框架。我们评估了属性的忠实性，并引入了图间完整性，以在无需详尽基准的情况下衡量相对语义覆盖率。我们的框架实现了98.8%的属性忠实性，而一次性基线为76.1%，同时完全覆盖了83.6%的忠实基线属性；反向覆盖方面，基线仅完全覆盖了我们的10.5%。这些改进为下游硬件设计和验证提供了更可靠、更全面的工程信息，降低了忽略约束和错误决策的风险。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12762

Authors: Haotian Qiao, Robert P. Dick

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12762.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12762

Published: 2026-10-12T02:39:54.799Z

10. Representation Alignment from a Few Paired Data via Distilled Optimal Transport

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12781

Authors: Julien Kordes, Reuben Dorent

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2610.12781.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12781

Published: 2026-10-12T02:39:54.799Z

VLM Domain Papers

1. CORAL-KV: Compact Cross-Layer Prediction and Residual Coding for Autoregressive Video KV Caches

Abstract:Autoregressive video diffusion models maintain a growing attention key-value (KV) cache, making memory a key bottleneck for long video generation. Recently, clustering-based quantization has emerged as an effective method to compress video KV caches by grouping similar key and value vectors then quantizing residual errors. However, when later frames attend to the compressed history, small reconstruction errors can compound as the video grows. Beyond the within-layer similarity exploited by clustering, we find strong dependencies across cache layers in video models. This motivates Coral-KV - Calibrated Offline, Residual-coded, and Adapted Locally - a KV-cache compression framework that uses cache states in each layer to predict keys and values in subsequent layers, then quantizes the remaining prediction residuals. We construct compact block-diagonal-plus-low-rank (block-DPLR) predictors: block-diagonal components model channel relationships within each attention head, while a global low-rank term models dependencies between heads. Coral-KV exposes a flexible fidelity-memory trade-off: it can reuse its predictors across videos or refit lightweight block-diagonal predictors to the current video for higher fidelity. Across extensive evaluations on both HY-WorldPlay and MAGI-1, our reusable configuration improves PSNR over QVG-4 by 0.32 dB and 4.95 dB, respectively, while using 48% and 37% less cache memory. Against QVG-Pro, our video-adapted configuration improves MAGI-1 PSNR by 3.77 dB while using 33.7% less cache memory. Coral-KV also maintains higher reconstruction fidelity over extended generation and reduces temporal discontinuities while retaining comparable generation latency.

中文摘要

摘要：自回归视频扩散模型维持一个不断增长的注意力键值（KV）缓存，这使得内存成为长视频生成的关键瓶颈。最近，基于聚类的量化作为一种有效方法出现，可以通过对相似的键和值向量进行分组，然后对残差误差进行量化，从而压缩视频KV缓存。然而，当后续帧访问压缩后的历史数据时，随着视频长度增加，小的重建误差可能会累积。除了聚类利用的层内相似性之外，我们发现视频模型的缓存层之间存在强依赖性。这促使我们提出了Coral-KV——校准的离线、残差编码、局部自适应的KV缓存压缩框架，它使用每层的缓存状态来预测后续层的键和值，然后对剩余的预测残差进行量化。我们构建了紧凑的块对角加低秩（block-DPLR）预测器：块对角分量建模每个注意力头内的通道关系，而全局低秩项建模头之间的依赖性。Coral-KV提供了灵活的保真度与内存权衡：它可以在不同视频之间重用其预测器，或者为当前视频重新拟合轻量级块对角预测器以获得更高保真度。在对HY-WorldPlay和MAGI-1的广泛评估中，我们的可重用配置分别比QVG-4提高了PSNR 0.32 dB和4.95 dB，同时使用的缓存内存分别减少了48%和37%。相较于QVG-Pro，我们的视频自适应配置将MAGI-1 PSNR提高了3.77 dB，同时使用的缓存内存减少了33.7%。Coral-KV在长时间生成中也维持了更高的重建保真度，并在保持生成延迟可比的情况下减少了时间不连续性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12514

Authors: Albert Ge, Akash Haridas, Utkarsh Saxena, Ruijing Yang, Ian Colbert, Frederic Sala, Vikram Appia, Emad Barsoum

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12514.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12514

Published: 2026-10-12T02:40:19.093Z

2. TraceICL: Thread-Resolved Demonstration Curation with Consistency Calibration for Robust Multimodal In-Context Learning

Abstract:Interleaved multimodal in-context learning (ICL) lets frozen large vision-language models (LVLMs) acquire new tasks from a handful of image-text-answer demonstrations, yet it remains fragile: demonstrations bind to the wrong images, redundant exemplars crowd out informative ones, and noisy answers are copied verbatim into predictions. We present AVA, a training-free framework that organizes the demonstration pipeline of multimodal ICL into three composable stages. A Thread-Resolved Decomposer (TRD) segments the interleaved context into per-sample threads and re-stitches them with explicit boundaries, restoring image-text binding. An Influence-scored Demonstration Selector (IDS) estimates each candidate’s marginal influence with one extra forward pass, removes redundancy, and orders the survivors by descending value. A Weak-to-Strong Consistency Calibration (WSC) module scores each demonstration by a weak sibling model’s reconstruction confidence and softly reweights its contribution during decoding. On five vision-language benchmarks with interleaved 4-shot prompts, AVA improves the strongest same-setting baseline by +1.7 points on average, with gains of up to +2.4 points on VizWiz, shrinks the require-visual performance gap on ScienceQA-IMG, and retains most of its gain under 30% noisy demonstrations, at a 26% latency overhead that an offline cache amortizes. No gradients are ever computed; the LVLM stays frozen.

中文摘要

摘要：交错多模态上下文学习（ICL）允许冻结的大型视觉-语言模型（LVLM）从少量图文-答案示例中学习新任务，但其仍然脆弱：示例可能绑定到错误的图像，冗余示例会挤掉有信息的示例，噪声答案会原样被复制到预测中。我们提出AVA，一个无需训练的框架，将多模态ICL的示例流程组织为三个可组合的阶段。线程解析解构器（TRD）将交错上下文分割为每个样本的线程，并通过显式边界重新组合，恢复图文绑定。影响评分示例选择器（IDS）通过一次额外的前向计算估计每个候选样本的边际影响，去除冗余，并按价值递减排序。弱到强一致性校准（WSC）模块通过弱兄弟模型的重建置信度对每个示例进行评分，并在解码过程中软性重新加权其贡献。在五个视觉-语言基准测试中，使用交错4-shot提示，AVA在相同设置下使最强基线平均提升1.7分，在VizWiz上最高提升2.4分，缩小ScienceQA-IMG的视觉需求性能差距，并在30%噪声示例情况下保留大部分增益，延迟开销为26%，可通过离线缓存摊销。整个过程中不计算梯度；LVLM保持冻结。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12516

Authors: Yigit Yilmaz, Elena Petrova, Mehmet Kaya, Lucia Rossi, Amir Rahman

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12516.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12516

Published: 2026-10-12T02:40:19.093Z

3. TerraAgent: Demonstration-Grounded Multi-Agent Reasoning for Long-Context Remote Sensing Visual Question Answering

Abstract:Remote sensing visual question answering (RS-VQA) requires answers that truly depend on image evidence, yet current remote sensing multimodal large language models (RS MLLMs) are still evaluated mainly by accuracy, which cannot reveal whether a correct answer was produced without using the image. This setting introduces three challenges: demonstrations retrieved across different ground sample distances (GSD), sensors, and city textures may be poorly matched; real queries often involve long visual contexts from multi-scale tiles, multi-temporal frames, and retrieved knowledge, where evidence can be diluted and errors can propagate through multi-step reasoning; and prior-only answering remains invisible to standard metrics. We present AVA, a training-free framework that wraps a frozen RS MLLM with three modules. GDR performs GSD-bucketed, scene-graph-based demonstration retrieval and filters demonstrations whose answers are unsupported by their images. MACE decomposes each query into an evidence-typed sub-question DAG, routes sub-questions to suitable perception tools, and records verified provenance in a shared evidence ledger with bounded re-perception. CVA audits the final answer through counterfactual perturbations of cited evidence, including ablation and swap, and computes a visual-dependency score for confidence calibration and evidence-forcing re-query. On RSVQA-LR/HR, VRSBench, and LHRS-Bench, AVA improves average accuracy by up to 2.40 points over the strongest same-setting baselines, raises the visual-dependency score from 0.28 to 0.49, reduces the prior-only answer rate from 46.2 percent to 21.7 percent, and lowers the hallucination rate from 17.1 percent to 9.8 percent, with 2.6x the cost of a single backbone forward pass and no fine-tuning.

中文摘要

摘要：遥感视觉问答（RS-VQA）需要依赖图像证据给出答案，而当前的遥感多模态大语言模型（RS MLLMs）仍然主要通过准确率进行评估，这无法揭示是否在未使用图像的情况下得到正确答案。这一设置提出了三个挑战：跨不同地面采样距离（GSD）、传感器和城市纹理检索的示例可能匹配不佳；真实查询通常涉及来自多尺度切片、多时间帧和检索知识的长视觉上下文，其中证据可能被稀释，并且错误可能通过多步骤推理传播；仅依赖先验的回答仍在标准指标中不可见。我们提出了AVA，这是一种无需训练的框架，将冻结的RS MLLM与三个模块结合。GDR执行基于GSD分桶的场景图示例检索，并过滤其答案未被图像支持的示例。MACE将每个查询分解为带有证据类型的子问题DAG，将子问题路由至适当的感知工具，并在共享证据账本中记录经验证的来源，具有有限的重新感知能力。CVA通过对引用证据进行反事实扰动（包括消融和交换）审计最终答案，并计算视觉依赖度得分以进行置信度校准和证据强制重新查询。在RSVQA-LR/HR、VRSBench和LHRS-Bench上，AVA在最强的同设置基线模型上将平均准确率提高最多2.40个百分点，将视觉依赖度得分从0.28提高到0.49，将仅依赖先验的答案率从46.2%降低至21.7%，将幻觉率从17.1%降低至9.8%，计算成本为单次主干前向传递的2.6倍，无需微调。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12517

Authors: Jingwen Huang, Yiming Fang, Hanxiao Wu, Boyuan He, Keyi Luo

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12517.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12517

Published: 2026-10-12T02:40:19.093Z

4. MIRAGE: Metric Inference and Reliability-Aware Gaussian Estimation from Feed-Forward Vision

Abstract:Feed-forward 3D priors enable high-quality reconstruction from image streams but often lack reliable metric scale and suffer from long-term trajectory drift. In contrast, LiDAR—inertial estimation provides metrically accurate motion but lacks the image-space consistency needed for high-fidelity reconstruction. We propose $MIRAGE$, which anchors feed-forward visual geometry to LiDAR—IMU metric motion and regulates online pose corrections according to correspondence support and cross-modal agreement, jointly enabling accurate metric pose estimation and high-fidelity Gaussian reconstruction. Experiments on Oxford Spires, M2DGR and INS show that MIRAGE achieves accurate metric trajectory estimation, outperforming COLMAP, a widely used but computationally expensive reference for vision-only benchmarks. Motivated by the scarcity of long-horizon monocular benchmarks for large-scale indoor reconstruction, we further build a multimodal data-collection platform and use MIRAGE to establish accurate 6-DoF reference trajectories for a new vision-only benchmark. We benchmark recent monocular large-scale reconstruction methods on this dataset, providing a systematic evaluation of their performance in long-horizon indoor scenes.

中文摘要

摘要：前馈三维先验能够从图像流中实现高质量的重建，但常常缺乏可靠的度量尺度，并且存在长期轨迹漂移问题。相比之下，LiDAR-惯性估计能够提供度量精确的运动信息，但缺乏实现高保真重建所需的图像空间一致性。我们提出了$MIRAGE$，它将前馈视觉几何锚定到LiDAR-IMU度量运动，并根据对应关系支持和跨模态一致性调节在线位姿校正，从而共同实现准确的度量位姿估计和高保真的高斯重建。在Oxford Spires、M2DGR和INS上的实验表明，MIRAGE实现了准确的度量轨迹估计，优于COLMAP——一个广泛使用但计算成本高昂的仅视觉基准参考。考虑到大型室内重建中长期单目基准的稀缺性，我们进一步构建了一个多模态数据采集平台，并使用MIRAGE为一个新的仅视觉基准建立准确的6自由度参考轨迹。我们在该数据集上对近期的大规模单目重建方法进行了基准测试，对它们在长期室内场景中的性能提供了系统性评估。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12568

Authors: Xinzhe Wang, Changjian Jiang, Kaiwen Song, Xudong Li, Kerui Ren, Ran Yi, Lizhuang Ma, Chunhua Shen, Linning Xu, Tao Lu, Mulin Yu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12568.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12568

Published: 2026-10-12T02:40:19.093Z

5. Scaling Multi-Teacher Distillation for Digital Pathology

Abstract:Multi-teacher distillation allows transferring knowledge from multiple “teacher” networks to a single “student” network. This method is promising in fields such as digital pathology, where many powerful foundation models were proposed recently. However, as we show in this paper, the standard multi-teacher distillation approach reacts poorly to an increase in the number of teachers used to train a student encoder. We demonstrate that learning teacher-specific representations is a key to scaling in MTD. Importantly, different design choices, such as learnable teacher tokens, a tailored attention scheme, additional mixture-of-experts layers and a contrastive loss, are proposed to better learn such teacher-specific representations. We show that our method better scales with respect to the number of considered teachers, allowing us to train compact student encoders, up to 10x more efficient than larger teacher foundation models, while matching their performance and even outperforming them on a large set of tile-level and slide-level tasks. We conduct a thorough empirical validation, evaluating more than 10 foundation models on 39 tasks spanning tile and slide levels. As a result, we release a new collection of strong and efficient foundation models, named OMNI, trained from the knowledge of 10 state-of-the-art foundation models. We finally conduct an analysis of the learned teacher-specific representations, highlighting their complementarity and explaining why they can be easily aggregated at downstream time.

中文摘要

摘要：多教师蒸馏允许将来自多个“教师”网络的知识转移到单个“学生”网络中。这种方法在数字病理学等领域具有潜力，因为最近提出了许多强大的基础模型。然而，正如我们在本文中所展示的，标准的多教师蒸馏方法在用于训练学生编码器的教师数量增加时表现不佳。我们证明，学习教师特定的表示是多教师蒸馏扩展的关键。重要的是，为了更好地学习这种教师特定的表示，提出了不同的设计选择，如可学习的教师标记、定制的注意力机制、额外的专家混合层以及对比损失。我们表明，该方法在考虑教师数量方面具有更好的扩展性，使我们能够训练紧凑的学生编码器，其效率比大型教师基础模型高达10倍，同时在性能上与它们匹配，甚至在大量的切片级和幻灯片级任务上超越它们。我们进行了全面的实证验证，评估了39个涵盖切片级和幻灯片级的任务中的10余个基础模型。因此，我们发布了一组新的强大且高效的基础模型集合，命名为OMNI，该模型通过10个最先进基础模型的知识进行训练。最后，我们对学习到的教师特定表示进行了分析，强调了它们的互补性，并解释了为什么它们可以轻松地在下游任务中聚合。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12574

Authors: Sofiène Boutaj, Pierre Marza, Varun Belagali, Dimitris Samaras, Maria Vakalopoulou, Stergios Christodoulidis

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12574.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12574

Published: 2026-10-12T02:40:19.093Z

6. Mind the Temporal Gap: Evaluating Video Models for Temporal Understanding in Driving

Abstract:Modern video-language models are increasingly treated as general-purpose video reasoners, yet it remains unclear whether they robustly capture temporal dynamics in safety-critical settings such as driving. We study this question through controlled ego-motion tasks that require identifying actions from driving videos. Across video-language models, we find that standard prompting, visual prompting, few-shot prompting, and chain-of-thought prompting remain unreliable and highly task- and model-dependent, with persistent class biases, although newer models show substantial task-specific few-shot gains. In contrast, lightweight LoRA finetuning and simple probes trained on intermediate visual features recover strong motion-discriminative performance, indicating that motion-relevant cues are often present inside the models but are not reliably used by the final video-to-language pathway. These results suggest that, in driving, current video-language models remain unreliable temporal reasoners and often behave more like frame-level, bias-driven recognizers than dependable temporal world models.

中文摘要

摘要：现代视频-语言模型越来越多地被视为通用的视频推理工具，但在如驾驶等安全关键场景下，它们是否能稳健地捕捉时间动态仍不清楚。我们通过受控的自我运动任务来研究这个问题，这些任务需要从驾驶视频中识别动作。在各类视频-语言模型中，我们发现标准提示、视觉提示、少量样本提示和连锁思维提示仍然不可靠，且高度依赖具体任务和模型，同时存在持续的类别偏差，尽管较新的模型在特定任务的少样本场景下表现出显著的提升。相比之下，轻量级LoRA微调和在中间视觉特征上训练的简单探针能够恢复强大的运动区分性能，这表明与运动相关的线索通常存在于模型内部，但最终的视频-语言通路并未可靠地利用这些线索。这些结果表明，在驾驶情境中，现有的视频-语言模型仍然是不可靠的时间推理器，它们的行为更像是基于帧的、受偏差驱动的识别器，而非可靠的时间世界模型。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12575

Authors: Shiva Sreeram, Erfan Aasi, Tsun-Hsuan Wang, Alaa Maalouf, Daniela Rus

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12575.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12575

Published: 2026-10-12T02:40:19.093Z

7. MemReg: Streaming Outdoor LiDAR Point Cloud Registration with Hybrid Memory Buffers

Abstract:Prior outdoor LiDAR registration works are commonly limited by the pair-wise input paradigm, which neglects temporal correlations inherent within streaming sequences. In this paper, we propose a novel multi-frame outdoor point cloud registration network for streaming LiDAR scans enhanced with pose-related memory buffers. The key observation is that long-term temporal LiDAR sequences can provide rich global contextual information to complete sparse measurements, filter outliers, and address low-overlap problems, thereby boosting registration performance. Specifically, two hybrid memory buffers are designed, including an implicit memory feature buffer and an explicit memory pose buffer, to store and dynamically update pose-related temporal features. Moreover, a novel dynamic history weighting module is developed to adaptively fuse current and history pose-related features. Extensive experiments on three outdoor datasets, including KITTI, nuScenes, and Apollo-Southbay, demonstrate state-of-the-art performance of MemReg, surpassing all previous pair-wise methods and also multi-frame SLAM systems. Our method also generalizes surprisingly well to multiview indoor registration scenarios with rather competitive performance on 3DMatch, 3DLoMatch, and ScanNet. Code will be released at this https URL.

中文摘要

摘要：以往的户外 LiDAR 配准工作通常受到成对输入范式的限制，忽略了流媒体序列中固有的时间相关性。本文提出了一种新型多帧户外点云配准网络，用于增强姿态相关记忆缓冲的流式 LiDAR 扫描。关键观察是，长期的时间 LiDAR 序列可以提供丰富的全局上下文信息，以补全稀疏测量、滤除异常值并解决低重叠问题，从而提升配准性能。具体而言，设计了两种混合记忆缓冲，包括隐式记忆特征缓冲和显式记忆姿态缓冲，用于存储和动态更新与姿态相关的时间特征。此外，开发了一种新型动态历史加权模块，以自适应地融合当前与历史的姿态相关特征。在包括 KITTI、nuScenes 和 Apollo-Southbay 的三个户外数据集上的大量实验表明，MemReg 展现了最先进的性能，超越了所有先前的成对方法以及多帧 SLAM 系统。我们的方法在多视角室内配准场景中也表现出惊人的泛化能力，在 3DMatch、3DLoMatch 和 ScanNet 上表现出相当有竞争力的性能。代码将通过此 https URL 发布。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12624

Authors: Jiuming Liu, Mengmeng Liu, Guangming Wang, Lihao Liu, Michael Ying Yang, Yunpeng Zhang, Francesco Nex, Hao Cheng, Hesheng Wang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12624.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12624

Published: 2026-10-12T02:40:19.093Z

8. ForcingBooster: Adaptive Patch-wise Denoising for Fast Autoregressive Video Diffusion

Abstract:Autoregressive video diffusion models have become a leading approach to real-time interactive video generation and video world modeling. However, existing methods typically allocate computation uniformly across space when generating the next video chunk, updating all patches with the same denoising schedule and number of function evaluations. This ignores the spatial heterogeneity of video content: static, simple, or recurring regions are often easier to denoise, while other regions require further refinement or additional context. A uniform budget thus leads to redundant computation on easy patches, whereas reducing the fixed step count can leave difficult patches insufficiently refined. We introduce \methodname{}, a general framework that dynamically allocates denoising steps to individual patches: easy patches stop early to improve efficiency, while difficult patches receive more refinement steps. This design accelerates generation with negligible quality degradation. Our method is compatible with most existing few-step autoregressive diffusion models, requiring only a lightweight routing module with modest training. Experiments demonstrate that \methodname{} reduces the total number of denoising steps applied across video patches by about 50\% while maintaining visual quality comparable to the four-step baseline.

中文摘要

摘要：自回归视频扩散模型已成为实时交互式视频生成和视频世界建模的领先方法。然而，现有方法通常在生成下一个视频块时将计算均匀分配于空间，所有补丁均采用相同的去噪计划和函数评估次数。这忽略了视频内容的空间异质性：静态、简单或重复区域通常更容易去噪，而其他区域则需要进一步细化或额外上下文。因此，统一预算导致简单音色的计算冗余，而减少固定步数则可能使困难音色的细化不足。我们引入了 \methodname{}，这是一个通用框架，动态地将去噪步骤分配给单个音色：简单音色提前结束以提高效率，而困难音色则获得更多细化步骤。该设计加速生成，且质量下降可忽略不计。我们的方法兼容大多数现有的少步自回归扩散模型，只需一个轻量级路由模块，训练有限。实验表明，\methodname{} 在保持与四步基线相当的视觉质量的同时，将视频补丁中应用的去噪步数减少约50\%。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12625

Authors: Zekai Gu, Peng Wang, Jiahao Lu, Yongsen Mao, Cheng Lin, Yuan Liu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12625.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12625

Published: 2026-10-12T02:40:19.093Z

9. Feature Competition and Its Representational Origins in Vision Language Models: Evidence from the Stroop Paradigm

Abstract:When you ask a vision-language model (VLM) a question with an image, it answers your question based on the image. But to act means to choose: saying one thing means choosing to respond not say another. This becomes an issue when the image contains multiple features that equally relate to the question at hand: how would VLMs choose to respond? This problem is precisely studied by the classical Stroop task in psychology: among a set of deliberately designed images whose text and color conflicts (e.g., \textcolor{blue}{RED}), participants are instructed to report the color. We adapt this psychological paradigm to a benchmark and prompted a range of contrastive and generative VLMs on their behavior. Across models, we observe a persistent bias toward the feature of text in Stroop images: when text and color disagree, models preferentially generate responses centered to the text rather than color. We then opened the black box and examined the source of this bias and why certain features influences the model’s behavior more than others. In the embedding space, we found that textual cues are represented more strongly and separably than color cues. This representational asymmetry predicts intervention success: steering models toward text-related directions is substantially easier than steering them toward color-related directions. Using a classical psychological paradigm, we demonstrate, at both behavioral and representational levels, that distinct features within the same image can exert unequal influence on VLM behavior. Our findings suggest that VLMs assign different degrees of saliency to different features within the same image, highlighting the need to understand how cue-specific saliency shapes their performance in widely used downstream applications.

中文摘要

摘要：当你向一个视觉-语言模型（VLM）提出带有图像的问题时，它会根据图像回答你的问题。但行动意味着选择：说某件事意味着选择回应而不是说另一件事。当图像包含多个与当前问题同等相关的特征时，这就成为一个问题：VLM会如何选择回应？这个问题正是通过心理学中的经典斯特鲁普任务（Stroop task）进行研究的：在一组刻意设计的图像中，文字和颜色相冲突（例如，\textcolor{blue}{RED}），参与者被指示报告颜色。我们将这一心理学范式改编为基准测试，并对一系列对比型和生成型VLM进行了行为测试。在各类模型中，我们观察到在斯特鲁普图像中模型持续倾向于文本特征的偏向：当文本和颜色不一致时，模型更倾向于生成以文本为中心的响应，而不是颜色。随后，我们打开黑箱，检查这种偏差的来源以及为什么某些特征比其他特征对模型行为影响更大。在嵌入空间中，我们发现文本线索的表示比颜色线索更强、更可区分。这种表示不对称可以预测干预的成功性：引导模型朝向与文本相关的方向比引导其朝向颜色相关的方向容易得多。通过使用经典心理学范式，我们在行为和表示层面上展示了同一图像中不同特征可以对VLM行为施加不平等的影响。我们的研究结果表明，VLM会对同一图像中的不同特征赋予不同程度的显著性，这凸显了理解特定线索显著性如何影响其在广泛使用的下游应用中的表现的必要性。

LLM Analysis

Q: 这篇论文试图解决什么问题？这篇论文旨在解决以下核心问题： - \*\*多模态冲突中的线索优先级\*\*：当视觉-语言模型（VLMs）面对单一图像内相互冲突的视觉线索时（例如Stroop任务中词义与油墨颜色不一致），模型如何仲裁并优先响应某一类特征？具体而言，研究关注模型是否像人类一样表现出对文本线索的系统性偏向。 - \*\*行为偏向的表征根源\*\*：探究上述文本偏向是否根植于模型内部嵌入空间的几何结构。论文试图解释为什么VLMs在冲突条件下优先选择词而非颜色，并将这种行为追溯到CLIP等编码器图像嵌入空间中词身份（word identity）与油墨颜色（ink color）在显著性和可分离性上的不对称表征。 - \*\*表征干预与行为引导\*\*：验证通过直接操纵图像嵌入（如沿词相关或颜色相关的方向进行"引导"）能否改变模型的特征偏好，以及这种可引导性（steerability）是否与表征的相对显著性（如向量的 ell_2 范数、方向可分离性）相对应。 - \*\*基准构建与系统性评估\*\*：论文还致力于构建一个受控的多模态Stroop基准，以系统评估对比式编码器（CLIP、SigLIP-2）和生成式VLMs在特征竞争情境下的行为一致性，并检验视觉变化（字体大小、粗细、对比度、伪词等）如何影响这种偏向。简言之，该研究试图\*\*解释VLMs在多特征冲突情境下的决策偏向及其神经表征起源\*\*，将经典心理学范式与模型内部可解释性方法相结合，以建立从行为到表征几何的因果联系。 Q2: 有哪些相关研究？该论文的相关研究可归纳为以下几个主要方向： ### 1. 视觉-语言模型架构与应用 - \*\*对比式嵌入模型\*\*：如 CLIP（Radford et al. 2021）与 SigLIP（Zhai et al. 2023），通过对比学习将图像与文本对齐到共享嵌入空间，但不直接生成文本。 - \*\*生成式视觉-语言模型\*\*：如 BLIP-2（Li et al. 2022）、Flamingo（Alayrac et al. 2022）、LLaVA（Liu et al. 2023）、Qwen2-VL（Wang et al. 2024）等，能够基于图像和文本提示生成自由形式的文本输出。 - \*\*下游任务应用\*\*：包括图像描述（Xu et al. 2015; Hossain et al. 2019）、图像-文本检索（Faghri et al. 2018; Xiao et al. 2025; Kim et al. 2025）以及视觉问答（Antol et al. 2015; Alayrac et al. 2022; Tsimpoukelli et al. 2021）。 ### 2. 多线索冲突与视觉知觉 - \*\*知觉决策与冲突解决\*\*：人类视觉系统在面对互相排斥的线索时必须将其解析为单一解释（Kornmeier and Bach 2012; Brascamp and Shevell 2021）。 - \*\*早期词汇学习与形状偏好\*\*：在物体命名和分类中，形状在早期词汇学习中具有重要作用（Smith 2000; Landau, Smith, and Jones 1988）。 - \*\*特征干扰与注意\*\*：某一特征的无关变化仍会干扰对另一特征的反应（Tanaka and Presnell 1999; Navon 1977）。 - \*\*视觉模型中的线索偏差\*\*：研究表明，ImageNet 训练的 CNN 在形状与纹理冲突时偏向纹理，而人类依赖形状（Geirhos et al. 2019）。 ### 3. 视觉-语言冲突与文本偏向 - \*\*图像与提示冲突\*\*：现有工作已探讨输入图像与文本提示之间的冲突（Li et al. 2023; Deng et al. 2025a; Xiao et al. 2026; Yuksekgonul et al. 2023）。 - \*\*广泛的文本主导性\*\*：近期研究记录了 VLMs 中更普遍的文本偏向（textual dominance）以及整合冲突视觉-文本线索的困难（Menon, Chandratreya, and Vondrick 2022; Pezeshkpour, Aminnaseri, and Hruschka 2025; Deng et al. 2025b; Jia et al. 2025）。 - \*\*颜色表征弱点\*\*：在对比式视觉-语言编码器中，颜色感知存在特定缺陷（Arias, Baldrich, and Vanrell 2024）。 ### 4. 内部表征分析方法 - \*\*表征相似性分析（RSA）\*\*：源自神经科学，通过余弦距离构建表征不相似性矩阵（RDMs）来比较高维神经或模型嵌入（Kriegeskorte, Mur, and Bandettini 2008）。 - \*\*可解释性方法\*\*：近期研究通过"神经分块"（neural chunking）从神经嵌入中提取可解释的概念单元（Wu et al. 2025），或使用稀疏自编码器（Sparse Autoencoders）识别单语义特征（Goh et al. 2021; Pach et al. 2025）。 ### 5. Stroop 范式在 VLMs 中的先验应用 - Zhao et al.（2025）曾将 Stroop 范式应用于 VLMs，但主要停留在行为层面的黑箱分析，未系统探究图像层面的受控变化，也未将行为偏向与内部表征几何或可引导性（steerability）建立联系。 Q3: 论文如何解决这个问题？该研究通过\*\*"行为观察 → 表征定位 → 因果干预"\*\*的三层递进框架解决视觉-语言模型（VLMs）在多线索冲突下的优先级与根源问题。具体方法如下： ### 1. 构建受控的多模态 Stroop 基准为系统性诱发特征竞争，研究构建了包含冲突线索的合成图像数据集： - \*\*基础集合\*\*：10 种基本颜色词（red, blue 等）各以 10 种油墨颜色渲染，共 100 张图像，其中 10 张为\*\*一致\*\*（congruent，词义匹配油墨色）、90 张为\*\*不一致\*\*（incongruent，词义与油墨色冲突）。 - \*\*视觉操控变体\*\*：通过改变字体大小（48–108 pt）、字重（light/normal/bold/narrow）、对比度（high/medium/low/same）以及使用无意义伪词（pseudoword），检验词偏向的鲁棒性。 - \*\*大规模验证集\*\*：生成 23,338 张具有系统变化的背景纹理、色调、饱和度和亮度的图像，排除模板记忆的可能。 ### 2. 跨架构的行为评估研究分别对\*\*对比式编码器\*\*与\*\*生成式 VLMs\*\*进行行为测试，以观察冲突下的特征偏好： - \*\*对比式模型（CLIP、SigLIP-2）\*\* 利用其共享嵌入空间中的余弦相似度进行零样本分类。对每张图像计算图像嵌入 f(I) 与文本嵌入 g(T_k) 的相似度：

s(I, T_k) = (f(I) · g(T_k)) / (|f(I)| |g(T_k)|)
使用两类提示族独立评分： - **词导向**：”The text says {word}” - **油墨导向**：”The text is written in {ink} color” 通过比较 top-1 选择，得到词准确率与油墨准确率。 - **生成式模型（BLIP-2、InstructBLIP、Kosmos-2、LLaVA、GIT、Qwen2-VL）** 使用开放式指令：”What color is the word in this image?” 收集自由文本输出后，归一化并映射为四类： - Ink-only（仅报告物理油墨色） - Word-only（仅报告词汇的语义颜色） - Both（同时明确提及两者） - Neither（其他无关输出） ### 3. 嵌入空间的表征相似性分析（RSA）为追溯行为偏向的表征根源，研究对 CLIP 的图像嵌入空间进行结构分析： - 收集三类图像的嵌入向量：**仅油墨**（纯色块）、**仅词汇**（灰度词）、**词汇+油墨**（Stroop 图像）。 - 计算表征不相似性矩阵（RDM），其中每个条目为余弦不相似度：

RDM(i, j) = 1 - (e_i · e_j) / (|e_i| |e_j|)

通过差分 RDM 隔离各特征的贡献：
Delta(Word) = RDM(Word+Ink) - RDM_(Ink-only)

Delta(Ink) = RDM(Word+Ink) - RDM(Word-only)
若 Delta(Word) 呈现更大、更集中的正值区域，则表明添加词特征比添加颜色特征更能显著改变嵌入空间的距离结构，即**词特征在表征上更显著**。 ### 4. 基于概念块的嵌入干预与引导研究进一步通过因果干预验证表征几何与行为偏向的对应关系： - **概念向量提取**：采用 Wu et al. (2025) 的”块”（chunk-based）方法，从包含特定概念的所有图像嵌入中，识别对该概念稳定响应的维度子集 Omega ，并计算其平均激活：

μ_Omega^((src)), quad μ_Omega^((tgt))

**嵌入引导操作**：对图像嵌入 E 沿特定概念方向进行编辑：
E’Omega = E_Omega - μ_Omega^((src)) + μ_Omega^((tgt))
干预后重新投影到单位球面，再与目标文本提示计算相似度。 - **干预类型与评估**： - 油墨颜色引导（如 red to blue） - 词汇引导（如 RED to BLUE） - 组合引导（同时执行上述两者）定义成功指标：
Delta = sim(E’, p(tgt)) - sim(E’, p_(src))
若 Delta > 0 ，则干预成功。通过比较成功率与平均相似度位移，发现词方向的引导几乎总是成功（100%），而颜色方向引导成功率极低（约 36.67%），从而将**行为上的词主导映射到表征空间中的向量范数更大、方向更可分离的词特征**。 ### 5. 视觉操控与提示变体的鲁棒性检验 - **视觉边界测试**：通过逐步降低词的可读性（如零对比度、伪词、字号溢出），发现只有当词完全不可读时，模型才转向颜色判断，证明文本偏向依赖于可识别的视觉形式。 - **提示敏感性分析**：对 Qwen2-VL 测试 9 种提示变体，发现明确提及”忽略词汇”的指令反而可能因启动效应（priming）降低油墨准确率，而简洁、无歧义的提问可部分缓解冲突。这表明生成式模型的偏向虽受提示调节，但默认仍根植于表征结构。 ### 6. 真实场景扩展验证为超越合成刺激，研究使用 FLUX.1 生成真实多模态冲突图像（如绿色 “STOP” 标志、红色 “CONNECTED” Wi-Fi 图标）。在这些场景中，CLIP 的决策随线索显著性变化：文本占主导时表现为词偏向，而色彩鲜艳或图标化区域可诱发颜色偏向，验证了特征显著性（saliency）作为普遍仲裁机制的作用。综上，论文通过**受控基准 → 多模型行为刻画 → 嵌入空间几何分析 → 定向表征干预**的完整链条，不仅确认了 VLMs 的文本主导行为，还证明该偏向源于嵌入空间中词特征具有更强的显著性与可分离性。 Q4: 论文做了哪些实验？论文围绕视觉–语言模型（VLMs）在经典 Stroop 冲突情境下的行为与表征机制，开展了以下系统性实验： —- ### 1. 基础 Stroop 基准测试（对比式编码器） - **被试模型**：CLIP（ViT-B/32）与 SigLIP-2（ViT-B/16）。 - **刺激材料**：100 张合成图像，由 10 个基本颜色词（red, blue, green, yellow, orange, purple, brown, pink, gray, black）各以 10 种油墨颜色渲染而成；其中 10 张为**一致**（congruent，词义与油墨色相同），90 张为**不一致**（incongruent，词义与油墨色冲突）。 - **评估协议**：采用两类提示族分别进行零样本匹配： - **词导向**（word-oriented）：”The text says {word}” - **油墨导向**（ink-oriented）：”The text is written in {ink} color” - **指标**：计算图像嵌入与文本嵌入的余弦相似度，取 Top-1 作为模型选择，分别报告**词准确率**与**油墨准确率**。 - **核心发现**：在不一致条件下，CLIP 的词导向准确率达 97.8%，而油墨导向准确率骤降至约 20%；SigLIP-2 的词导向准确率达 100%，油墨导向仅约 5–6%。 —- ### 2. 视觉操控鲁棒性实验（CLIP）为检验文本偏向是否特异于初始刺激，对不一致图像进行多组受控视觉扰动： - **字体大小变化**：48 pt、72 pt、90 pt、108 pt。结果显示，只要文字仍可读，词偏向持续（85–97%），仅当 108 pt 导致溢出、可读性下降时，油墨准确率才略有回升。 - **字重变化**：Bold、Normal、Light、Narrow。词偏向在各条件下保持稳定（88–97%）。 - **对比度操控**：High、Medium、Low、Same（文字与背景融合）。随着对比度降低，词偏向单调递减；在 Same 条件下（文字不可见），油墨准确率达 100%。 - **伪词（Pseudoword）替换**：将颜色词替换为无意义字符串（如 GRIV、BLIR），保留低层视觉形式但消除语义。此时 CLIP 显著转向油墨判断（约 77.8% 的油墨匹配率）。 - **无意义词控制（”zarp”）**：将全部刺激中的词替换为 “zarp” 并将其加入候选标签。CLIP 在 100% 的试次中选择 “zarp”，油墨准确率为 0%，表明模型对视觉文本 token 存在强烈的 OCR 式偏好。 —- ### 3. 生成式 VLM 行为实验 - **被试模型**：BLIP-2（FlanT5-XL）、InstructBLIP（Vicuna-7B）、Kosmos-2（1B）、LLaVA（Vicuna-7B v1.6）、GIT（B/16）、Qwen2-VL-7B-Instruct。 - **指令**：使用开放式提问 “What color is the word in this image?”（该问法在语义上故意模糊，可指油墨色或词义色）。 - **响应编码**：将自由文本输出归一化并映射为四类： - **Ink-only**：仅报告物理油墨色 - **Word-only**：仅报告词汇语义色 - **Both**：同时明确提及词与色 - **Neither**：其他无关输出 - **核心发现**：多数模型表现出强烈的词偏向（如 Kosmos-2 从未报告油墨色）；LLaVA 是唯一在不一致条件下更倾向于报告油墨色的模型（54.4%），但优势微弱。 —- ### 4. 规模与指令调谐的缓解测试（Qwen2-VL）针对表现最好的 Qwen2-VL-7B-Instruct，测试提示措辞对冲突缓解的影响： - **长提示**：”You will see a single English word rendered in a colored ink. Ignore the written word and answer ONLY the ink color…” - **短提示**：”What is the ink color of the text in the image? Answer with one lowercase color word only.” - **扩展变体**：进一步测试 9 种额外提示变体（p1–p9）。 - **核心发现**：油墨准确率对提示高度敏感，范围从 24% 到 97% 不等。反直觉的是，明确指示”忽略词汇”的长提示反而因启动效应（priming）降低了油墨准确率（31.1%），而简洁短提示效果更佳（60.0%）。 —- ### 5. 表征相似性分析（RSA / RDM）在 CLIP 的图像嵌入空间中，探究词与油墨颜色的表征不对称性： - **刺激分组**： - **Ink-only**：纯色块图像 - **Word-only**：灰度词图像 - **Word+Ink**：完整 Stroop 彩色词图像 - **分析方法**：计算三组条件下的 10×10 表征不相似性矩阵（RDM），其中每个条目为余弦不相似度：

RDM(i, j) = 1 - (e_i · e_j) / (|e_i| |e_j|)

**差分 RDM**：通过相减隔离单一特征的贡献：
Delta(Word) = RDM(Word+Ink) - RDM_(Ink-only)

Delta(Ink) = RDM(Word+Ink) - RDM_(Word-only)

**核心发现**： Delta_(Word) 呈现更大、更集中的正值区域，表明添加词特征比添加颜色特征更能显著拉开嵌入空间的距离，即**词特征在表征上更显著、更具可分离性**。 —- ### 6. 嵌入空间干预与引导实验（Steering）基于 Wu et al. (2025) 的”神经分块”（neural chunking）方法，对 CLIP 图像嵌入进行因果干预： - **概念向量提取**：对每一概念（如颜色 “red” 或词 “RED”），收集包含该概念的全部图像嵌入，通过方差聚类筛选稳定响应维度 Omega ，计算平均激活 μ_Omega^((src)) 与 μ_Omega^((tgt)) 。 - **引导操作**：对图像嵌入 E 在子空间 Omega 上进行加减替换：

E’Omega = E_Omega - μ_Omega^((src)) + μ_Omega^((tgt))
编辑后重新投影至单位球面。 - **干预条件**： - 油墨引导（如 red to blue） - 词汇引导（如 RED to BLUE） - 组合引导（词汇+油墨同时干预） - **评估指标**：计算引导后的相似度位移：
Delta = sim(E’, p(tgt)) - sim(E’, p_(src))
若 Delta > 0 则判定为成功。 - **核心发现**：词汇引导成功率 100%，平均位移 +0.0934 ± 0.0214 ；油墨引导成功率仅 36.67%，平均位移接近零（ -0.0017 ± 0.0166 ）；组合引导成功率亦为 100%，位移更大。词汇向量的 ell_2 范数（均值 6.36）约为颜色向量（均值 2.99）的 2.1 倍，且颜色向量间高度共线（平均余弦相似度 ≈ 0.79 ）。 —- ### 7. 大规模 Stroop 数据集验证 - **数据规模**：生成 23,338 张图像，其中 23,328 张为不一致刺激，系统变化背景纹理、色调、饱和度、亮度及轻度空间扰动。 - **目的**：排除模型依赖模板记忆的可能，检验在视觉多样性下的泛化性。 - **结果**：CLIP 的词准确率仍高达 99.5%，油墨准确率仅 15.9%，与小规模控制实验的趋势一致。 —- ### 8. 真实场景多模态冲突测试（FLUX 生成图像） - **刺激**：使用 FLUX.1 生成真实冲突图像，如绿色 “STOP” 标志、红色 “CONNECTED” Wi-Fi 图标、红色 “EXIT” 标志等。 - **评估**：对每张图像，分别用”词提示”（描述文字内容）和”颜色提示”（描述主导视觉颜色）与 CLIP 图像嵌入计算相似度，判定模型倾向。 - **核心发现**：CLIP 的决策随线索显著性动态变化：文本繁重或类似 OCR 的场景诱发**词主导**，而色彩强烈或图标化区域则诱发**颜色主导**，表明特征显著性（saliency）是更普遍的仲裁机制。 —- ### 9. SigLIP-2 的对比分析 - 在基础 Stroop 集及视觉操控（大小、字重、对比度、伪词）上重复实验。 - **核心发现**：SigLIP-2 表现出比 CLIP 更刚性的词偏向，在不一致条件下词准确率始终为 100%，视觉操控几乎无法削弱该偏向；仅在伪词条件下部分恢复颜色判断（67% 的油墨导向胜率）。 Q5: 有什么可以进一步探索的点？基于论文的讨论与局限性，可从以下维度展开进一步探索： —- ### 1. 扩展表征干预至更多模型架构论文出于计算资源限制，仅在 CLIP 上完成了嵌入空间的引导实验。未来可将类似的潜在干预方法推广至： - 更新的对比式编码器（如 SigLIP-2、InternVL-2.5） - 生成式 VLMs（如 LLaVA-1.6、Qwen-2.5-VL、InternVL 系列）通过比较不同架构在相同冲突刺激下的可引导性（steerability），可验证”特征显著性差异根植于嵌入几何”这一结论的普适性。 —- ### 2. 模型规模与指令调谐的scaling效应论文观察到 Qwen2-VL-7B-Instruct 的文本偏向可通过精心设计的提示部分缓解，但未在同一家族内系统测试参数规模的影响。未来可： - 在统一模型系列中对比不同规模（如 Qwen2-VL 2B → 7B → 72B）的冲突解决行为 - 分离”预训练规模”与”指令调谐程度”对特征显著性偏差的各自贡献 —- ### 3. 从合成刺激到自然场景的多线索冲突当前研究主要依赖程序生成的 Stroop 刺激。后续工作可引入更自然的冲突情境： - **纹理 vs. 形状**（延续 Geirhos et al. 2019 的脉络） - **材质 vs. 语义**（如”木制”纹理的金属刀具） - **分层图形与场景文本**（如广告牌、UI 界面、交通标志中的图文冲突）此类场景中的线索显著性计算更为复杂，可检验嵌入空间几何是否仍能预测行为偏向。 —- ### 4. 提升颜色表征的显著性与可分离性论文发现油墨颜色在 CLIP 嵌入空间中呈”短向量、高度共线”的弱势表征。改善途径包括： - **改进分块/特征提取方法**：若颜色信息以更分布式方式编码，现有基于子空间质心的方法可能低估其可引导性；稀疏自编码器（SAEs）或更细粒度的神经分块可能捕获更完整的颜色方向 - **训练数据干预**：构造强调颜色独立于语义关联的预训练数据（如大量非语义纯色块与描述配对），削弱 OCR 式文本主导 - **架构解耦**：设计显式分离”阅读模块”与”视觉感知模块”的多模态架构，减少图文特征在浅层嵌入中的不均衡竞争 —- ### 5. 动态机制与”类Stroop slowdown”分析人类在 Stroop 不一致条件下表现出反应时延长。虽然 VLMs 缺乏真实反应时，但可考察： - **逐层表征演变**：在视觉编码器的不同深度，词特征与颜色特征何时发生分离或主导 - **token概率动态**：在生成式模型中，观察 ink-color token 与 word token 的 logits/概率在解码过程中的竞争轨迹，是否存在类似”认知冲突”的置信度振荡或延迟收敛 - **注意力热图分析**：冲突条件下，模型的交叉注意力或自注意力是否系统性地偏向文本区域而非彩色区域 —- ### 6. 更精细的提示工程与人机交互研究 Qwen2-VL 实验表明，提示措辞可显著调节冲突结果（24%–97% 的波动范围）。未来可： - 系统量化提示特异性（prompt specificity）与模型服从度之间的函数关系 - 探索”反启动”（anti-priming）指令设计，即如何在提示中有效抑制文本关注而不触发语义激活 - 将 VLMs 的提示敏感性结果与人类 Stroop 研究中的任务指令效应进行跨物种对比 —- ### 7. 跨语言与跨文化的Stroop冲突当前基准仅使用英文颜色词。考虑到： - 不同语言对颜色范畴的切分存在差异（如俄语对蓝色的细分） - 部分语言使用非颜色词修饰（如”危险””停止”等语义标签）构建多语言 Stroop 基准可检验：文本偏向是普遍的视觉-语言现象，还是特定于英语等高频预训练语言的产物。 —- ### 8. 对抗鲁棒性与安全审计如果 VLMs 对嵌入空间中的词方向高度敏感，而对颜色方向相对迟钝，则： - **对抗攻击**：攻击者可通过微弱扰动图像中的文本内容（或注入伪文本）轻易覆盖真实的视觉信号（如将红色交通灯上的”GO”文本放大以误导模型） - **防御机制**：基于论文的干预框架，可开发”颜色增强”的预处理或嵌入修正模块，强制提升视觉特征在决策中的权重，用于高风险视觉识别系统的安全审计 —- ### 9. 多图像与视频序列中的时序冲突将 Stroop 范式从静态图像扩展至： - **多图像输入**：当图文冲突跨帧变化时，模型是否表现出时序一致性或切换代价 - **视频理解**：在动态场景中，运动显著性与文本语义冲突如何被仲裁（如移动的红色”STOP”标志与绿色背景） —- ### 10. 边缘案例与响应分类的精细化论文指出，生成式 VLM 的四类响应编码（Ink-only / Word-only / Both / Neither）可能遗漏混合描述的边缘案例。未来可采用： - 基于 LLM-as-a-Judge 的细粒度语义解析 - 连续量表而非离散分类，以捕捉模型响应中的不确定性或置信度渐变 Q6: 总结一下论文的主要内容这篇论文系统研究了视觉–语言模型（VLMs）在面临单一图像内多线索冲突时的决策偏向及其神经表征根源，核心内容可概括如下： —- ### 研究背景与核心问题 - 借鉴心理学经典的 **Stroop 范式**：当颜色词（如 “RED”）被印刷在与其语义冲突的油墨颜色中（如蓝色）时，模型应当优先响应哪一特征？ - 核心问题：VLMs 是否存在类似人类的”词优于颜色”偏向？该偏向是行为层面的黑箱现象，还是根植于内部嵌入空间的几何结构？ —- ### 研究方法与实验体系 #### 1. 多模态 Stroop 基准构建 - 合成 100 张控制图像：10 种基本颜色词各以 10 种油墨颜色渲染，形成 **10 张一致（congruent）** 与 **90 张不一致（incongruent）** 刺激。 - 额外生成大规模数据集（23,338 张）及视觉操控变体（字体大小、字重、对比度、伪词、无意义词），以排除模板记忆并检验鲁棒性。 #### 2. 跨架构行为评估 - **对比式编码器**：CLIP、SigLIP-2。在共享嵌入空间中通过余弦相似度进行零样本匹配，分别使用”词导向”提示（”The text says {word}”）与”油墨导向”提示（”The text is written in {ink} color”）独立评分。 - **生成式 VLMs**：BLIP-2、InstructBLIP、Kosmos-2、LLaVA、GIT、Qwen2-VL。使用开放式指令： > “What color is the word in this image?” 将自由文本输出编码为 Ink-only / Word-only / Both / Neither 四类。 #### 3. 表征相似性分析（RSA） - 在 CLIP 图像嵌入空间中，对三类刺激计算表征不相似性矩阵（RDM）： - Ink-only（纯色块）、Word-only（灰度词）、Word+Ink（彩色词）。 - 通过差分 RDM 隔离特征贡献：

Delta(Word) = RDM(Word+Ink) - RDM_(Ink-only)

Delta(Ink) = RDM(Word+Ink) - RDM_(Word-only)

4. 嵌入空间干预与引导（Steering） - 采用神经分块（neural chunking）方法提取概念编码子空间 Omega ，计算源/目标概念的稳定平均激活 μ_Omega^((src)) 、 μ_Omega^((tgt)) 。 - 对图像嵌入 E 执行向量替换：

E’Omega = E_Omega - μ_Omega^((src)) + μ_Omega^((tgt))
评估词方向、颜色方向及组合方向的引导成功率与相似度位移 Delta 。 —- ### 关键发现 #### 行为层面：持续而强烈的文本偏向 - **对比式模型**：在不一致条件下，CLIP 词导向准确率达 **97.8%**，油墨导向仅 **20%**；SigLIP-2 词导向 **100%**，油墨导向仅 **5–6%**。 - **生成式模型**：多数模型强烈倾向于报告词的语义颜色而非物理油墨色。Kosmos-2 从不报告油墨色；LLaVA 是唯一略微偏向油墨的模型（54.4%）。 - **鲁棒性**：词偏向在字体大小、字重变化下持续存在；仅当词完全不可读（零对比度、溢出、伪词）时，模型才转向颜色判断。伪词条件下 CLIP 油墨匹配率升至约 **78%**。 - **提示敏感性**：Qwen2-VL-7B-Instruct 的油墨准确率随提示措辞剧烈波动（**24% 至 97%**）；明确指示”忽略词汇”的指令反而因启动效应降低表现。 #### 表征层面：嵌入几何的不对称性 - **RDM 分析**： Delta(Word) 呈现更大、更集中的正值区域，表明添加词特征比添加颜色特征更能显著拉开嵌入空间中的距离。 - **向量属性**：词概念向量的 ell_2 范数（均值 **6.36**）约为颜色向量（均值 **2.99**）的 **2.1 倍**；颜色向量间高度共线（平均余弦相似度 ≈ 0.79 ），而词向量更为分散。 #### 因果干预：可引导性的不对称 - **词引导**：成功率 **100%**，平均相似度位移 +0.0934 ± 0.0214 。 - **颜色引导**：成功率仅 **36.67%**，平均位移接近零（ -0.0017 ± 0.0166 ）。 - **组合引导**：成功率 **100%**，位移更大（ +0.1172 ± 0.0215 ），主要由词分量驱动。 #### 真实场景泛化 - 在 FLUX.1 生成的自然冲突图像（如绿色 “STOP” 标志、红色 “CONNECTED” 图标）中，CLIP 的决策随线索显著性动态变化：文本主导场景呈现词偏向，色彩鲜艳或图标化场景则呈现颜色偏向。 —- ### 结论与意义 - 论文证实，跨架构的 VLMs 在 Stroop 冲突下表现出系统性的**文本优于颜色**偏向。 - 该偏向并非纯粹的后训练行为产物，而是**根植于图像嵌入空间的几何结构**：词身份被编码为更长、更可分离的方向，而颜色特征更弱、更纠缠。 - 表征显著性（representational saliency）可直接预测因果可引导性：干预强表征方向可靠地翻转模型偏好，弱表征方向则难以操纵。 - 这项工作展示了如何将经典心理学范式适配用于审计多模态基础模型，并为理解多特征竞争、提升模型安全性（如防止文本覆盖视觉信号的对抗风险）提供了机制层面的解释框架。 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: Nurbüke Teker, Rui Xiao, Zeynep Akata, Shuchen Wu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12650.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12650

Published: 2026-10-12T02:40:19.093Z

10. WoodBR-Cross: A public wood-image dataset and individual-tree-disjoint benchmark for species identification

Abstract:Wood-image benchmarks can yield optimistic estimates when images from the same tree occur in both training and test sets. WoodBR-Cross contains 4,875 transverse-surface images in 38 operational taxonomic classes, linked to 303 distinct trees identified by numeric catalogue witness codes (codigos-testemunho) in the Brazilian Forest Service collection. A species-level subset comprises 3,406 images, 20 classes and 214 trees. Three tasks compare duplicate-aware image partitions for all classes (A), image partitions for the species subset (B1), and individual-tree-disjoint partitions for that subset (B2). In B2, all images from each tree are kept together, so no tree is shared across training, validation or test within a fold. ResNet-50, ConvNeXt-Tiny and ViT-B/16 were evaluated with three outer folds, internal validation and five training seeds. Handcrafted and frozen-feature baselines complement border-only and randomized-label controls. Validation-selected ConvNeXt-Tiny achieved B2 tree-level macro-F1 of 0.8621 +/- 0.0317; frozen DINOv2 features yielded 0.9000 +/- 0.0722. Values are means and standard deviations across three seed-averaged fold means. Results estimate performance on previously unseen trees from this collection.

中文摘要

摘要：当同一棵树的图像同时出现在训练集和测试集中时，木材图像基准测试可能会产生乐观的估计。WoodBR-Cross 包含 4,875 张横截面表面图像，涵盖 38 个操作分类类，与巴西森林服务局收藏中由数字目录证书代码（codigos-testemunho）标识的 303 棵不同的树相关联。物种级别的子集包括 3,406 张图像、20 个类别和 214 棵树。三个任务比较了所有类别的重复感知图像划分（A）、物种子集的图像划分（B1）以及该子集的单树不相交划分（B2）。在 B2 中，每棵树的所有图像都保持在一起，因此在某一折中训练、验证或测试之间不会共享任何树。对 ResNet-50、ConvNeXt-Tiny 和 ViT-B/16 进行了评估，采用三个外折叠、内部验证和五个训练随机种子。手工特征和冻结特征的基线补充了仅边界和随机标签控制。经过验证选择的 ConvNeXt-Tiny 在 B2 树级宏 F1 指数上达到 0.8621 +/- 0.0317；冻结的 DINOv2 特征获得 0.9000 +/- 0.0722。这些数值表示三个随机种子平均折叠均值下的平均值和标准差。结果估计了该集合中以前未见树的性能。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2610.12676

Authors: Daniel Casanova, Alexandre Bahia Gontijo, Arnaldo Candido Junior, Silvana Nisgoski, Pedro Luiz de Paula Filho

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2610.12676.pdf

CoolPaper URL: https://papers.cool/arxiv/2610.12676

Published: 2026-10-12T02:40:19.093Z