评分理由: 论文提出MoRI框架，旨在提升LLM在科学构思任务中的推理能力。核心相关关键词包括：1) “Large Language Models” (论文基于LLM构建框架)；2) “Post-training” (使用监督微调初始化模型)；3) “RLHF” (采用复合强化学习奖励进行训练)；4) “Chain of Thought” (框架显式学习从研究动机到方法的推理过程)；5) “System 2 Thinking” (强调深度推理，避免表面概念重组)；6) “LLM Agents” (论文属于LLM智能体研究，改进现有智能体方法)；7) “AI for Science” (应用于科学构思领域，属于AI for Science范畴)。其他关键词如MoE、量化、RAG等未在论文中涉及，故评分为0。

科学构思旨在给定科学背景下提出新颖解决方案。现有基于大语言模型（LLM）的智能体方法虽模拟人类研究流程，却未能充分建模科学推理过程，导致其产出多为缺乏技术深度与科学依据的表层概念重组。为解决这一问题，我们提出 MoRI（基于动机的科学构思推理框架），该框架使大语言模型能够显式学习从研究动机到方法论的推理过程。基础大语言模型首先通过监督微调进行初始化，以从给定情境中生成研究动机，随后在复合强化学习奖励机制下进行训练，以逼近科学严谨性：（1）熵感知信息增益鼓励模型基于真实方法论揭示并阐述高复杂度的技术细节；（2）对比语义增益约束推理轨迹，确保其与科学有效解决方案保持概念一致。实验结果表明，MoRI 在创新性、技术严谨性和可行性等多个维度上显著优于主流商用大语言模型及复杂智能体基线方法。代码将在 \href{https://github.com/ECNU-Text-Computing/IdeaGeneration}{GitHub} 平台开源。

Scientific ideation aims to propose novel solutions within a given scientific context. Existing LLM-based agentic approaches emulate human research workflows, yet inadequately model scientific reasoning, resulting in surface-level conceptual recombinations that lack technical depth and scientific grounding. To address this issue, we propose \textbf{MoRI} (\textbf{Mo}tivation-grounded \textbf{R}easoning for Scientific \textbf{I}deation), a framework that enables LLMs to explicitly learn the reasoning process from research motivations to methodologies. The base LLM is initialized via supervised fine-tuning to generate a research motivation from a given context, and is subsequently trained under a composite reinforcement learning reward that approximates scientific rigor: (1) entropy-aware information gain encourages the model to uncover and elaborate high-complexity technical details grounded in ground-truth methodologies, and (2) contrastive semantic gain constrains the reasoning trajectory to maintain conceptually aligned with scientifically valid solutions. Empirical results show that MoRI significantly outperforms strong commercial LLMs and complex agentic baselines across multiple dimensions, including novelty, technical rigor, and feasibility. The code will be made available on \href{https://github.com/ECNU-Text-Computing/IdeaGeneration}{GitHub}.

评分理由: 论文核心研究LLM的推理能力提升，提出TARo方法在推理时对齐，属于大模型技术原理创新。高度相关的关键词包括：LLMs（核心研究对象）、Alignment（测试时对齐方法）、CoT Reasoning（提升推理性能）、System 2 Thinking（涉及深度推理）。有一定相关的关键词：Post-training（对比传统后训练方法）、RLHF（属于对齐技术范畴）、AI for Science（应用于临床推理）。其余关键词与论文内容无关。

大语言模型（LLMs）展现出强大的推理能力，但通常需要昂贵的后训练才能达到高性能。近期的测试时对齐方法提供了一种轻量级替代方案，但主要被探索用于偏好对齐而非推理任务。为填补这一空白，我们提出了令牌级自适应路由（Token-level Adaptive Routing, TARo），该方法在完全保持基础模型冻结的状态下，于推理阶段引导模型进行结构化推理。具体而言，我们首先在分步数学推导轨迹上训练奖励模型，以捕捉细粒度的逻辑一致性信号；随后引入一个可学习的令牌级路由器，自动控制奖励模型对基础模型的引导强度。大量实验表明，TARo在推理性能上相比基础模型显著提升高达+22.4%，较现有令牌级测试时对齐方法提升+8.4%，同时还能增强分布外临床推理（MedXpertQA）和指令遵循（AlpacaEval）能力。此外，TARo无需重新训练即可从小型骨干模型泛化至大型骨干模型，从而将测试时对齐的应用范围从偏好优化扩展至鲁棒的跨领域推理。

Large language models (LLMs) exhibit strong reasoning capabilities but typically require expensive post-training to reach high performance. Recent test-time alignment methods offer a lightweight alternative, but have been explored mainly for preference alignment rather than reasoning. To bridge this gap, we propose, Token-level Adaptive Routing (TARo), which steers frozen LLMs toward structured reasoning entirely at inference time. Specifically, we first train reward models on step-wise mathematical traces to capture fine-grained logical consistency signals, then introduce a learnable token-level router that automatically controls the guidance of the reward model to the base model. Extensive experiments show that TARo significantly improves reasoning performance by up to +22.4% over base model and +8.4% over existing token-level test-time alignment methods, while also boosting out-of-distribution clinical reasoning (MedXpertQA) and instruction following (AlpacaEval). Furthermore, TARo also generalizes from small to large backbones without retraining, extending test-time alignment from preference optimization to robust, cross-domain reasoning.

评分理由: 论文核心研究工具增强大语言模型（LLMs）中推理与外部行动的耦合问题，因此与"Large Language Models"、“Tool Use”、“LLM Agents"高度相关（10分）。论文强调需要"in-depth reasoning"和"multi-step reasoning”，与"System 2 Thinking"和"Chain of Thought"高度相关（10分）。其他关键词如MoE、量化、对齐、RAG等未在论文中涉及，得0分。

工具增强型大语言模型（LLM）必须将多步推理与外部行动紧密耦合，然而现有基准测试常因复杂的环境动态、记忆知识或数据集污染而混淆这种相互作用。本文提出ZebraArena——一个通过程序化生成、用于研究工具增强型LLM中推理-行动耦合的诊断性环境，其具备可控难度和知识最小化设计，能有效限制模型从记忆或数据集污染中获益。ZebraArena中的每个任务都需要一组关键信息，这些信息仅能通过针对性工具调用获取，从而在外部信息获取与演绎推理之间构建了可解释的接口。该设计通过唯一解实现确定性评估，并提供了理论最优查询次数以衡量工具使用效率。我们证明ZebraArena要求深度推理与精准外部工具调用的结合，这对前沿推理模型（如GPT-5和Gemini 2.5 Pro）仍具挑战性——它们在困难实例上仅达到60%的准确率。我们还观察到理论最优性与实际工具使用之间存在持续差距：例如GPT-5的工具调用次数比理论最优值高出70-270%。本文重点阐述了评估中的关键发现，期望ZebraArena能推动关于内部推理与外部行动交互机制的进一步研究。

Tool-augmented large language models (LLMs) must tightly couple multi-step reasoning with external actions, yet existing benchmarks often confound this interplay with complex environment dynamics, memorized knowledge or dataset contamination. In this paper, we introduce ZebraArena, a procedurally generated diagnostic environment for studying reasoning-action coupling in tool-augmented LLMs, with controllable difficulty and a knowledge-minimal design, which limits gains from memorization or dataset contamination. Each task in ZebraArena requires a set of critical information which is available only through targeted tool use, yielding an interpretable interface between external information acquisition and deductive reasoning. This design provides deterministic evaluation via unique solutions, and a theoretical optimal query count for measuring efficient tool use. We show that ZebraArena requires a combination of in-depth reasoning and accurate external tool calling, which remains a challenge as frontier reasoning models such as GPT-5 and Gemini 2.5 Pro only achieves 60% accuracy on the hard instances. We also observe a persistent gaps between theoretical optimality and practical tool usage. For example, GPT-5 uses 70-270% more tool calls than the theoretical optimum. We highlight the key findings in our evaluation, and hope ZebraArena stimulates further research on the interplay between internal reasoning and external action.

!!! info

论文构建了一个基于约束满足问题（CSP）的模拟环境。首先生成完整的斑马谜题实例，然后随机隐藏部分线索，形成部分可观察的初始状态。环境提供基于规则的预言机接口，支持事实查询（验证特定属性赋值）和关系查询（检查位置或逻辑关系）。模型通过ReAct等框架进行交互，利用工具获取信息，逐步缩小解空间直至唯一。评估指标包括任务准确率、工具调用次数（与理论最优值对比）以及Token消耗量。

• 隔离性强：成功将推理能力与工具使用能力解耦，避免了环境噪声和复杂动态的干扰。
• 评估维度丰富：不仅关注最终准确率，还引入了工具调用效率（查询次数）和成本（Token数）的评估。
• 可解释性高：基于逻辑谜题的任务设计使得模型的每一步推理和行动都可以被精确追踪和验证。
• 抗污染：程序化生成无限新实例，有效解决了现有基准测试的数据泄露问题。

• 环境简化：相比真实世界的复杂环境（如网页导航、物理交互），ZEBRAARENA过于理想化，可能无法完全反映模型在现实场景中的表现。
• 领域局限：任务仅限于逻辑推理，可能无法评估模型在其他类型知识（如常识、物理定律）与工具结合时的表现。
• 工具类型单一：提供的工具主要是布尔查询和关系检查，缺乏真实世界中API的多样性和复杂性（如文本生成、图像处理）。

评分理由: 论文核心研究LLMs的无监督强化学习微调方法（RLIF），提出PowerFlow框架，通过分布匹配优化LLMs的双重能力（逻辑推理与创造性）。高度相关关键词：LLMs（核心研究对象）、Post-training/SFT（涉及微调）、RLHF/RLAIF/DPO（属于强化学习微调范畴）。中等相关关键词：Instruction Tuning/Alignment（涉及模型能力对齐）、Chain of Thought/System 2 Thinking（涉及逻辑推理）、Self-Correction/Self-Improvement（涉及模型自我优化）。其他关键词未在论文中直接涉及。

无监督内部反馈强化学习已成为一种无需外部监督即可激发大型语言模型潜在能力的前景广阔的研究范式。然而，现有方法依赖于启发式内在奖励，这些奖励通常缺乏明确的理论优化目标，且容易产生退化性偏差。本研究提出PowerFlow，一个将无监督微调重新定义为分布匹配问题的原理性框架。通过将GFlowNet构建为非归一化密度的摊销变分采样器，我们提出了一种长度感知的轨迹平衡目标，该目标能显式地抵消自回归生成中固有的结构长度偏差。通过以$α$-幂分布为目标，PowerFlow能够定向激发大型语言模型的双重特性：通过锐化分布（$α> 1$）来强化逻辑推理能力，或通过平坦化分布（$α< 1$）来释放表达创造力。大量实验表明，PowerFlow在各项任务中持续优于现有RLIF方法，其表现达到甚至超越了有监督的GRPO。此外，通过缓解对齐模型中的过度锐化现象，我们的方法在多样性与质量上实现了同步提升，从而在创造性任务中推动了帕累托边界的演进。

Unsupervised Reinforcement Learning from Internal Feedback (RLIF) has emerged as a promising paradigm for eliciting the latent capabilities of Large Language Models (LLMs) without external supervision. However, current methods rely on heuristic intrinsic rewards, which often lack a well-defined theoretical optimization target and are prone to degenerative biases. In this work, we introduce PowerFlow, a principled framework that reformulates unsupervised fine-tuning as a distribution matching problem. By casting GFlowNet as an amortized variational sampler for unnormalized densities, we propose a length-aware Trajectory-Balance objective that explicitly neutralizes the structural length biases inherent in autoregressive generation. By targeting $α$-power distributions, PowerFlow enables the directional elicitation of the dual nature of LLMs: sharpening the distribution ($α> 1$) to intensify logical reasoning, or flattening it ($α< 1$) to unlock expressive creativity. Extensive experiments demonstrate that PowerFlow consistently outperforms existing RLIF methods, matching or even exceeding supervised GRPO. Furthermore, by mitigating over-sharpening in aligned models, our approach achieves simultaneous gains in diversity and quality, shifting the Pareto frontier in creative tasks.

!!! info

论文采用生成流网络作为核心算法，将其视为非归一化密度的摊销变分采样器。为了解决自回归结构带来的轨迹概率指数衰减问题，作者推导了长度感知轨迹平衡（LA-TB）目标函数，将配分函数重新参数化为摊销的token级能量项，从而在长度归一化的能量表面上进行优化。通过调节$\alpha$参数，控制目标分布的锐化或扁平化程度，分别针对推理增强和创造力释放进行训练。

• 理论基础扎实：摆脱了启发式奖励设计，基于明确的统计力学原理（$\alpha$-power分布）进行优化。
• 解决关键偏差：针对自回归生成固有的长度偏差提出了有效的解决方案（LA-TB），提高了训练稳定性。
• 功能灵活统一：通过单一参数$\alpha$即可控制模型向推理或创造力两个相反方向进化，揭示了模型的双重属性。
• 实验效果显著：在推理和创意任务上均表现出色，特别是在恢复对齐模型的创造力方面有突破。

• 计算开销：虽然摊销了推理成本，但基于GFlowNet的训练可能比标准的监督微调或简单的RLIF更复杂且计算量更大。
• 参数敏感性：$\alpha$参数的选择至关重要，可能需要针对不同任务或模型进行细致调整，缺乏自适应选择机制。
• 适用范围：主要针对无监督场景，对于有明确外部奖励信号的任务，其优势可能不如专门设计的RLHF或RLVR方法明显。

评分理由: 论文核心是提出一个强化学习框架（LSE）来训练大语言模型（LLMs）在测试时自我进化，通过迭代优化上下文来提升在新问题上的表现。这与"Large Language Models"高度相关（核心研究对象）。方法基于强化学习，将多步进化问题简化为单步RL目标，与"RLHF"等关键词高度相关（核心方法）。论文涉及模型自我改进和反思，与"Self-Correction"等关键词高度相关。论文提到模型迭代优化上下文以提升性能，涉及多步推理和深入思考，与"Chain of Thought"和"System 2 Thinking"有一定关联（5分）。论文在测试时通过上下文优化来提升性能，与"In-context Learning"有一定关联（5分）。其他关键词如MoE、量化、RAG等未在摘要中提及，评为0分。

我们提出“学习式自我进化”框架，这是一种强化学习框架，用于训练大语言模型在测试阶段优化其自身上下文。我们将该框架置于测试时自我进化的情境中，使模型能够基于已见问题的反馈迭代优化上下文，从而在新问题上表现更佳。现有方法完全依赖模型固有的推理能力，从未针对此任务进行显式训练。本框架将多步进化问题简化为单步强化学习目标，其中每次上下文编辑的奖励由下游性能提升程度决定。我们将此目标与树状引导进化循环相结合。在文本到SQL生成和通用问答任务上，采用本框架训练的40亿参数模型超越了基于GPT-5与Claude Sonnet 4.5的自进化策略，以及包括GEPA和TextGrad在内的提示优化方法，并且无需额外训练即可迁移指导其他模型。我们的研究结果证明了将自我进化作为可学习技能的有效性。

We introduce Learning to Self-Evolve (LSE), a reinforcement learning framework that trains large language models (LLMs) to improve their own contexts at test time. We situate LSE in the setting of test-time self-evolution, where a model iteratively refines its context from feedback on seen problems to perform better on new ones. Existing approaches rely entirely on the inherent reasoning ability of the model and never explicitly train it for this task. LSE reduces the multi-step evolution problem to a single-step RL objective, where each context edit is rewarded by the improvement in downstream performance. We pair this objective with a tree-guided evolution loop. On Text-to-SQL generation (BIRD) and general question answering (MMLU-Redux), a 4B-parameter model trained with LSE outperforms self-evolving policies powered by GPT-5 and Claude Sonnet 4.5, as well as prompt optimization methods including GEPA and TextGrad, and transfers to guide other models without additional training. Our results highlight the effectiveness of treating self-evolution as a learnable skill.

!!! info

论文首先形式化定义了测试时跨集间自我进化问题，聚焦于基于提示的更新。技术路线上，采用树引导搜索策略，维护一个进化树，使用Upper Confidence Bound (UCB)算法选择节点进行扩展。在训练阶段，利用强化学习训练一个策略模型，输入当前上下文和性能摘要，输出改进后的上下文。奖励函数定义为编辑后与编辑前在保留集上的性能差值，通过策略梯度方法更新模型参数。

• 显式优化：打破了依赖模型固有能力的局限，通过RL显式训练自我进化能力。
• 高效性：小模型经过训练即可超越大模型，降低了部署成本。
• 鲁棒性：树引导机制避免了线性进化的贪婪陷阱，提供了更稳健的上下文探索。
• 通用性：方法不依赖特定任务，展示了在不同任务上的泛化和迁移能力。

• 计算开销：树引导搜索和保留集评估增加了测试时的计算成本。
• 评估依赖：需要可验证的奖励信号（如SQL执行结果），对于开放式任务可能难以应用。
• 上下文长度限制：随着进化树的扩展，存储历史上下文可能触及模型的上下文窗口限制。
• 单步简化：虽然简化了训练，但可能忽略了长程进化轨迹中的复杂依赖关系。

评分理由: 论文提出TerraScope，一种用于地球观测的视觉语言模型（VLM），专注于像素级地理空间推理。与关键词的相关性分析如下：1）与"Chain of Thought"高度相关（10分），因为论文核心是像素级推理链，并创建了Terra-CoT数据集。2）与"AI for Science"高度相关（10分），属于地球科学领域的AI应用。3）与"Large Language Models"、“Pre-training”、“Supervised Fine-tuning”、“System 2 Thinking”、“Explainable AI"有一定关联（各5分），因为VLM基于大模型技术，涉及训练和推理过程，并提供可解释性。其他关键词如MoE、SLMs、RAG、RLHF等与论文内容无直接关系，得0分。

视觉语言模型（VLMs）在地球观测（EO）领域展现出潜力，但在需要将复杂空间推理与精确像素级视觉表征相锚定的任务中仍面临挑战。为解决这一问题，我们提出了TerraScope——一个能够实现像素锚定地理空间推理的统一视觉语言模型，其具备两项核心能力：（1）模态灵活推理：可处理单模态输入（光学或合成孔径雷达SAR），并在多模态可用时自适应融合不同模态至推理流程；（2）多时序推理：能整合时间序列数据以进行多时相变化分析。此外，我们构建了Terra-CoT数据集，该大规模数据集包含100万个样本，其推理链中嵌入了来自多源数据的像素级掩码。我们还提出了首个像素锚定地理空间推理基准TerraScope-Bench，包含六个子任务，通过同时评估答案准确性与掩码质量来确保真实的像素锚定推理。实验表明，TerraScope在像素锚定地理空间推理任务上显著优于现有视觉语言模型，同时提供可解释的视觉证据。

Vision-language models (VLMs) have shown promise in earth observation (EO), yet they struggle with tasks that require grounding complex spatial reasoning in precise pixel-level visual representations. To address this problem, we introduce TerraScope, a unified VLM that delivers pixel-grounded geospatial reasoning with two key capabilities: (1) modality-flexible reasoning: it handles single-modality inputs (optical or SAR) and adaptively fuses different modalities into the reasoning process when both are available; (2) multi-temporal reasoning: it integrates temporal sequences for change analysis across multiple time points. In addition, we curate Terra-CoT, a large-scale dataset containing 1 million samples with pixel-level masks embedded in reasoning chains across multiple sources. We also propose TerraScope-Bench, the first benchmark for pixel-grounded geospatial reasoning with six sub-tasks that evaluates both answer accuracy and mask quality to ensure authentic pixel-grounded reasoning. Experiments show that TerraScope significantly outperforms existing VLMs on pixel-grounded geospatial reasoning while providing interpretable visual evidence.

!!! info

论文构建了基于视觉-语言架构的TerraScope模型，包含视觉编码器、文本编码器、投影器、掩码解码器和大语言模型。模型采用两阶段训练：首先使用200万指代表达分割对进行预训练，然后使用100万像素级CoT指令数据进行微调。在推理过程中，模型交替生成文本推理步骤和分割掩码，利用掩码选择视觉特征并注入LLM，实现像素级 grounding。对于多模态输入，采用文本引导的交叉注意力机制自适应选择光学或SAR特征。

• 创新性地将像素级分割掩码嵌入到视觉推理链中，解决了EO领域精细空间分析的难题。
• 统一框架支持多模态（光学/SAR）和多时相推理，适应性强。
• 提供了可解释的视觉证据（掩码），增强了模型的可信度。
• 构建了大规模数据集和专用基准，推动了该领域的发展。
• 端到端推理，无需依赖外部工具，降低了系统复杂度。

• 生成像素级掩码增加了计算开销，可能比纯文本推理更耗时。
• 依赖自动化管道生成的训练数据，可能存在噪声或标注偏差。
• 对于极大规模的超高分辨率图像，处理能力可能受限于显存和计算资源。

评分理由: 论文D-Mem专注于为LLM智能体设计一个双过程记忆系统，以提升长期推理能力。核心相关关键词包括：1) “Large Language Models” (10分)：论文明确使用GPT-4o-mini和Qwen3-235B-Instruct作为实验模型，是核心基础。2) “LLM Agents” (10分)：论文直接研究自主智能体的记忆系统，是核心主题。3) “Retrieval-Augmented Generation” (10分)：论文批判并改进现有的基于检索的记忆框架，是核心技术背景。4) “Chain of Thought"和"System 2 Thinking” (各5分)：论文涉及长期推理和精细上下文理解，与多步推理和深度思考概念相关，但非直接技术实现。其他关键词如MoE、量化、对齐等与论文内容无直接关联，故评0分。

在持久性自适应智能体发展的推动下，为这些系统配备高保真记忆访问能力以实现长程推理已成为关键需求。然而，当前主流的基于检索的记忆框架通常采用增量处理范式，持续提取对话记忆并更新至向量数据库，在查询时依赖语义检索。虽然这种方法速度较快，但其本质上依赖于有损抽象，常常遗漏上下文关键信息，且难以处理需要细粒度上下文理解的查询。为此，我们提出了D-Mem——一种双过程记忆系统。该系统保留轻量级向量检索以处理常规查询，同时建立详尽的完全审议模块作为高保真备用机制。为实现认知效率与精度的平衡，D-Mem采用多维质量门控策略动态桥接这两个过程。基于GPT-4o-mini和Qwen3-235B-Instruct模型在LoCoMo和RealTalk基准上的实验验证了本方法的有效性。值得注意的是，我们的多维质量门控策略在LoCoMo基准上使用GPT-4o-mini实现了53.5的F1分数，优于静态检索基线Mem0$^\ast$（51.2），并恢复了完全审议模块96.7%的性能表现（55.3），同时显著降低了计算成本。

Driven by the development of persistent, self-adapting autonomous agents, equipping these systems with high-fidelity memory access for long-horizon reasoning has emerged as a critical requirement. However, prevalent retrieval-based memory frameworks often follow an incremental processing paradigm that continuously extracts and updates conversational memories into vector databases, relying on semantic retrieval when queried. While this approach is fast, it inherently relies on lossy abstraction, frequently missing contextually critical information and struggling to resolve queries that rely on fine-grained contextual understanding. To address this, we introduce D-Mem, a dual-process memory system. It retains lightweight vector retrieval for routine queries while establishing an exhaustive Full Deliberation module as a high-fidelity fallback. To achieve cognitive economy without sacrificing accuracy, D-Mem employs a Multi-dimensional Quality Gating policy to dynamically bridge these two processes. Experiments on the LoCoMo and RealTalk benchmarks using GPT-4o-mini and Qwen3-235B-Instruct demonstrate the efficacy of our approach. Notably, our Multi-dimensional Quality Gating policy achieves an F1 score of 53.5 on LoCoMo with GPT-4o-mini. This outperforms our static retrieval baseline, Mem0$^\ast$ (51.2), and recovers 96.7% of the Full Deliberation’s performance (55.3), while incurring significantly lower computational costs.

!!! info

论文采用双过程架构：首先利用改进的Mem0*进行增量记忆处理和快速检索（System 1）；接着通过多维质量门控评估检索结果的质量；如果质量不达标，则触发完全审议模块（System 2），直接对原始对话历史进行分块扫描、事实提取与评分，最终生成高保真答案。

• 有效解决了增量记忆压缩导致的有损抽象问题，提升了复杂查询的推理能力。
• 通过门控机制实现了准确率与效率的良好平衡，避免了盲目使用高计算资源。
• 完全审议模块提供了高保真的推理上限，且能缓解长上下文中的“迷失在中间”现象。
• 实验验证充分，在多个基准测试上展示了优越性。

• 完全审议模块虽然作为后备，但在极端复杂查询下仍可能产生高昂的计算成本。
• 质量门控策略本身依赖LLM进行评估，可能引入额外的评估开销或评估误差。
• 系统架构相对复杂，涉及多个模块的协同工作，工程实现难度较大。

评分理由: 论文提出EntropyCache，一种针对扩散语言模型（dLLMs）的无训练KV缓存方法，核心是解决KV缓存效率问题。与关键词高度相关的包括：1）“KV Cache Compression” OR “Linear Attention” OR “FlashAttention”（15分）：论文直接研究KV缓存优化方法，是核心内容；2）“Large Language Models” OR “LLMs” OR “Foundation Models”（10分）：论文针对扩散大语言模型（dLLMs）进行优化；3）“Speculative Decoding” OR “Inference Acceleration”（10分）：论文目标是通过KV缓存实现推理加速；4）“Chain of Thought” OR “CoT Reasoning” OR “Multi-step Reasoning”（5分）：论文在CoT基准测试上评估了方法。其他关键词如MoE、量化、对齐等与论文内容无关。

基于扩散的大型语言模型（dLLMs）依赖双向注意力机制，这阻碍了无损KV缓存的实现，并需要在每个去噪步骤中进行完整的前向传播。现有的近似KV缓存方法通过选择性更新缓存状态来降低计算成本，但其决策开销随上下文长度或模型深度增加而线性增长。我们提出EntropyCache，一种无需训练的KV缓存方法，该方法以新解码词元分布的最大熵作为恒定开销的信号，用于决定何时重新计算。我们的设计基于两项实证观察：（1）解码词元的熵与KV缓存漂移相关，为缓存陈旧性提供了低成本的代理指标；（2）解码词元的特征波动在解掩码后持续多个步骤，这促使我们对最近解码的$k$个词元进行重新计算。跳转或重新计算的决策仅需每步$O(V)$的计算量，与上下文长度和模型规模无关。在LLaDA-8B-Instruct和Dream-7B-Instruct上的实验表明，EntropyCache在标准基准测试中实现了$15.2\times$-$26.4\times$的加速，在思维链基准测试中实现了$22.4\times$-$24.1\times$的加速，同时保持了具有竞争力的准确率，且决策开销仅占推理时间的$0.5%$。代码发布于https://github.com/mscheong01/EntropyCache。

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\times$-$26.4\times$ speedup on standard benchmarks and $22.4\times$-$24.1\times$ on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only $0.5%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

!!! info

论文首先通过实证分析，计算解码Token熵与KV缓存漂移（余弦距离）之间的相关性，验证了熵作为预测指标的有效性。其次，利用PCA分析Token轨迹，揭示了特征波动持续多步的现象。基于此，设计了包含三个阶段的推理流程：前向传播（全量或部分）、解码与熵评估、跳过决策与近期Token选择。当最大熵超过阈值时触发全量计算，否则仅重算当前掩码Token和最近解码的k个Token。

显著提升了扩散语言模型的推理吞吐量，实现了数量级的加速；决策机制极其轻量，不随上下文长度或模型深度增加而增加开销；无需对模型进行额外训练或微调，通用性强；基于扎实的实证观察，理论依据充分。

作为一种近似缓存方法，在追求极致速度的同时仍存在精度与速度的权衡；需要调整熵阈值τ和窗口大小k等超参数；该方法专门针对扩散语言模型设计，不适用于传统的自回归模型。

评分理由: 论文核心研究LLM代理在安全环境感知中的行为，直接涉及"Large Language Models"和"LLM Agents”，给予10分。研究涉及多代理协商（“Multi-agent Systems”）和可解释性分析（“Mechanistic Interpretability”），给予5分。论文提到代理依赖上下文窗口形成环境感知，与"Context Window Extension"有一定关联，给予5分。其他关键词如MoE、SFT、RAG等未在论文中涉及，给予0分。

NDAI（非披露人工智能）区域允许发明者与投资者智能体在可信执行环境（TEE）内进行协商，若未达成协议，任何已披露信息将被删除。这使得全面披露知识产权成为发明者智能体的理性策略。然而，利用此基础设施要求智能体能够区分安全与不安全环境，而大型语言模型智能体天生缺乏这种能力，因为它们仅能依赖通过上下文窗口传递的证据来形成对执行环境的认知。我们提出：不同大型语言模型在形成对其执行环境安全性的认知时，如何权衡不同形式的证据？通过在10种语言模型及多种证据场景下进行NDAI式协商任务实验，我们发现一种明显的不对称性：失败的认证普遍抑制所有模型的披露行为，而通过的认证则引发高度异质性响应：部分模型增加披露，部分不受影响，少数模型甚至矛盾地减少披露。这表明当前大型语言模型能够可靠地检测危险信号，但无法可靠地验证安全性——而这正是NDAI区域等隐私保护智能体协议所需的核心能力。通过可解释性分析、针对性微调或改进证据架构等方式弥合这一差距，仍然是部署能够根据实际证据质量校准信息共享的智能体所面临的核心开放挑战。

NDAI zones let inventor and investor agents negotiate inside a Trusted Execution Environment (TEE) where any disclosed information is deleted if no deal is reached. This makes full IP disclosure the rational strategy for the inventor’s agent. Leveraging this infrastructure, however, requires agents to distinguish a secure environment from an insecure one, a capability LLM agents lack natively, since they can rely only on evidence passed through the context window to form awareness of their execution environment. We ask: How do different LLM models weight various forms of evidence when forming awareness of the security of their execution environment? Using an NDAI-style negotiation task across 10 language models and various evidence scenarios, we find a clear asymmetry: a failing attestation universally suppresses disclosure across all models, whereas a passing attestation produces highly heterogeneous responses: some models increase disclosure, others are unaffected, and a few paradoxically reduce it. This reveals that current LLM models can reliably detect danger signals but cannot reliably verify safety, the very capability required for privacy-preserving agentic protocols such as NDAI zones. Bridging this gap, possibly through interpretability analysis, targeted fine-tuning, or improved evidence architectures, remains the central open challenge for deploying agents that calibrate information sharing to actual evidence quality.

!!! info

研究采用黑盒行为测量方法，构建了一个模拟的发明者-投资者谈判任务。实验设置了四种证据场景（无TEE、仅文本声明、文本声明+真证明、文本声明+假证明），并在10个不同的LLM模型上运行。通过法官模型对卖方代理的回复进行评分（0-1的披露分数），计算平均披露分数和证明依赖指数（ARI），以分析不同模型对安全证据的权重分配。

• 问题定义精准：抓住了LLM代理在安全环境中运行的一个关键盲点——缺乏原生环境感知。
• 实验设计严谨：覆盖了多种证据组合和主流模型，提供了具有统计意义的实证数据。
• 指标创新：ARI指数有效地量化了模型对安全证据的敏感度，便于跨模型比较。
• 现实意义强：直接关联到TEE内AI代理的经济效用和隐私保护技术的落地可行性。

• 模拟环境限制：实验使用的是模拟的证明和上下文，而非真实的TEE硬件执行，可能影响代理对证据“真实性”的感知。
• 黑盒分析局限：主要依赖输出行为进行推断，缺乏对模型内部表征的白盒可解释性分析。
• 任务范围较窄：主要集中在特定的发明披露谈判任务，在其他类型的代理行为中结论是否普适尚待验证。
• 时间背景特殊性：论文日期为2026年，涉及部分未来模型（如GPT-5.2），当前无法复现验证。

评分理由: 论文核心是应用LLM解决VANET网络碎片化问题，因此与"Large Language Models"高度相关（10分）。论文通过四阶段流程将通用LLM转化为领域专家，涉及领域适应和微调，与"Pre-training/Domain Adaptation"和"Post-training/SFT"有一定关联（5分）。LLM用于识别拓扑重要性，体现推理能力，与"Chain of Thought"和"System 2 Thinking"相关（5分）。LLM作为智能体先验指导策略，与"LLM Agents"相关（5分）。其他关键词如MoE、SLMs、Scaling Laws、RAG、Quantization等未涉及，得0分。

车载自组织网络（VANETs）是自动驾驶的数字基石，但在城市环境中，由于物理遮挡，其网络往往面临严重的割裂问题。具备高机动性的无人机（UAVs）已成为弥合这些连接鸿沟的关键解决方案。然而，传统的基于深度强化学习（DRL）的无人机部署策略缺乏对道路拓扑的语义理解，常导致盲目探索和样本效率低下。相比之下，大语言模型（LLMs）拥有强大的推理能力，能够识别拓扑重要性，但将其应用于控制任务仍具挑战性。为此，我们提出了语义增强深度强化学习（SA-DRL）框架。首先，我们提出了一种基于道路拓扑图（RTG）与双连通图（DCG）的网络割裂量化方法。随后，我们设计了一个四阶段流程，将通用大语言模型转化为特定领域的拓扑专家。最后，我们提出了语义增强近端策略优化（SA-PPO）算法，该算法采用逻辑融合机制，将大语言模型的语义推理作为先验知识直接注入策略中，从而有效引导智能体朝向关键路口。大量高保真仿真实验表明，SA-PPO以卓越的效率实现了最先进的性能，仅需26.6%的训练回合数即可达到基线性能水平。最终，相较于其他方法，SA-PPO将两项关键连接指标分别提升了13.2%和23.5%，同时将能耗降至基线水平的28.2%。

Vehicular Ad-hoc Networks (VANETs) are the digital cornerstone of autonomous driving, yet they suffer from severe network fragmentation in urban environments due to physical obstructions. Unmanned Aerial Vehicles (UAVs), with their high mobility, have emerged as a vital solution to bridge these connectivity gaps. However, traditional Deep Reinforcement Learning (DRL)-based UAV deployment strategies lack semantic understanding of road topology, often resulting in blind exploration and sample inefficiency. By contrast, Large Language Models (LLMs) possess powerful reasoning capabilities capable of identifying topological importance, though applying them to control tasks remains challenging. To address this, we propose the Semantic-Augmented DRL (SA-DRL) framework. Firstly, we propose a fragmentation quantification method based on Road Topology Graphs (RTG) and Dual Connected Graphs (DCG). Subsequently, we design a four-stage pipeline to transform a general-purpose LLM into a domain-specific topology expert. Finally, we propose the Semantic-Augmented PPO (SA-PPO) algorithm, which employs a Logit Fusion mechanism to inject the LLM’s semantic reasoning directly into the policy as a prior, effectively guiding the agent toward critical intersections. Extensive high-fidelity simulations demonstrate that SA-PPO achieves state-of-the-art performance with remarkable efficiency, reaching baseline performance levels using only 26.6% of the training episodes. Ultimately, SA-PPO improves two key connectivity metrics by 13.2% and 23.5% over competing methods, while reducing energy consumption to just 28.2% of the baseline.

评分理由: 论文核心是开发一个基于LLM的智能体框架（SignAgent）用于手语标注和数据集构建。高度相关的关键词包括：1）“Large Language Models”（论文明确使用LLMs作为核心组件）；2）“LLM Agents”（论文提出的是agentic framework，属于智能体研究）。中等相关的关键词包括：1）“Chain of Thought”（智能体进行推理和决策，涉及多步推理）；2）“System 2 Thinking”（智能体进行深入推理分析）；3）“Tool Use”（智能体协调使用语言工具）。其他关键词如MoE、量化、RAG等未在论文中涉及，评分为0。

本文介绍了SignAgent，一种新颖的智能体框架，该框架利用大语言模型（LLMs）实现可扩展的、基于语言学基础的手语（Sign Language, SL）标注与数据集构建。传统的手语计算方法通常仅在语素层面运作，忽视了关键的语言学细微差别，而人工语言学标注仍是一个重大瓶颈，其过程缓慢且成本高昂，难以用于创建大规模、具备音系学意识的数据集。SignAgent通过两个核心组件应对这些挑战：SignAgent Orchestrator——一个负责协调一系列语言学工具并进行推理的LLM，以及SignGraph——一个提供词汇与语言学基础的知识驱动型LLM。我们在两项下游标注任务上评估了该框架。首先，在伪语素标注任务中，智能体执行约束性分配，利用多模态证据为手语序列提取并排序合适的语素标签。其次，在身份语素标注任务中，智能体通过推理视觉相似性和音系重叠性来检测并优化视觉聚类，从而正确识别和归类词汇手语变体。我们的结果表明，这种智能体方法在大规模、具备语言学意识的数据标注与构建方面表现出色。

This paper introduces SignAgent, a novel agentic framework that utilises Large Language Models (LLMs) for scalable, linguistically-grounded Sign Language (SL) annotation and dataset curation. Traditional computational methods for SLs often operate at the gloss level, overlooking crucial linguistic nuances, while manual linguistic annotation remains a significant bottleneck, proving too slow and expensive for the creation of large-scale, phonologically-aware datasets. SignAgent addresses these challenges through SignAgent Orchestrator, a reasoning LLM that coordinates a suite of linguistic tools, and SignGraph, a knowledge-grounded LLM that provides lexical and linguistic grounding. We evaluate our framework on two downstream annotation tasks. First, on Pseudo-gloss Annotation, where the agent performs constrained assignment, using multi-modal evidence to extract and order suitable gloss labels for signed sequences. Second, on ID Glossing, where the agent detects and refines visual clusters by reasoning over both visual similarity and phonological overlap to correctly identify and group lexical sign variants. Our results demonstrate that our agentic approach achieves strong performance for large-scale, linguistically-aware data annotation and curation.

!!! info

论文采用基于智能体的框架方法。首先，构建SignAgent Orchestrator作为中央控制器，采用ReAct风格的推理循环，通过生成推理轨迹、调用工具或查询知识图谱来迭代优化状态。其次，利用SignGraph基于词汇知识图谱和语言学知识图谱进行检索增强生成（RAG），提供语言学基础。技术路线上，通过基础工具（如手形、运动、位置分类器）提取低级特征，再由增强工具融合多模态线索生成结构化证据，最终由Orchestrator进行推理以完成伪词汇对齐和ID词汇标注等任务。

论文的主要优点在于创新性地将智能体框架引入手语处理领域，有效解决了人工标注成本高的问题。通过结合语言学知识图谱，模型不仅关注视觉特征，还能进行深层的语言学推理，超越了传统的 gloss 级别处理。分层工具集的设计使得系统具有高度的模块化和可解释性，能够灵活处理复杂的注释任务。

该框架的性能可能依赖于基础工具（如手形、运动分类器）的准确性，这些底层模块的误差可能会向上传播。此外，基于LLM的智能体推理通常计算成本较高，推理速度可能较慢，限制了实时应用的潜力。虽然框架通用，但目前的评估主要集中在特定任务上，在更广泛的手语方言或复杂非手动特征（如面部表情）上的泛化能力有待进一步验证。

评分理由: 论文研究Vision-Language-Action (VLA)模型在机器人平台上的实际效率，属于大模型在具体领域（机器人/具身智能）的应用研究。与关键词的相关性分析：1) “Large Language Models” (5分)：VLA模型是多模态大模型，属于大模型范畴；2) “Supervised Fine-tuning” (5分)：摘要提到supervised fine-tuning作为常见适应方法被评估；3) “LLM Agents” (8分)：VLA模型用于具身智能体，是LLM Agents的具体实现；4) “Quantization” (8分)：论文明确研究model compression（包括量化等压缩技术）；5) “In-context Learning” (5分)：摘要提到in-context prompting作为评估方法。其他关键词如MoE、Scaling Laws、RLHF等未在论文中涉及，得0分。

视觉-语言-动作（Vision-Language-Action, VLA）模型近年来通过联合推理视觉、语言与运动模态，使具身智能体能够执行日益复杂的任务。然而，我们发现当前VLA研究中普遍采用的“效率”概念——通常以参数量、浮点运算量或解码吞吐量来衡量——并不能反映其在机器人平台上的实际性能。在现实世界执行中，效率由系统层面的具身行为决定，例如任务完成时间、轨迹平滑度、累积关节旋转量和运动能耗。通过对模型压缩、令牌稀疏化与动作序列压缩的对照研究，我们得出若干挑战常见假设的观察：（1）在传统指标下减少计算量的方法，尽管能维持任务成功率，却常常增加端到端执行成本或降低运动质量。（2）系统层面的具身效率指标揭示了学习到的动作策略中隐藏的性能差异，这些差异在传统评估中无法显现。（3）常见的适应方法（如上下文提示或有监督微调）对具身效率的提升有限且仅针对特定指标。虽然这些方法能够降低目标具身效率指标（如急动度或动作频率），但由此获得的收益可能以其他指标（如更长的完成时间）为代价。综合而言，我们的研究表明传统推理效率指标可能忽略具身执行的重要方面。引入具身效率评估能够更完整地反映策略行为与实际性能，从而为VLA模型提供更公平、更全面的比较基准。

Vision-Language-Action (VLA) models have recently enabled embodied agents to perform increasingly complex tasks by jointly reasoning over visual, linguistic, and motor modalities. However, we find that the prevailing notion of ``efficiency’’ in current VLA research, characterized by parameters, FLOPs, or token decoding throughput, does not reflect actual performance on robotic platforms. In real-world execution, efficiency is determined by system-level embodied behaviors such as task completion time, trajectory smoothness, cumulative joint rotation, and motion energy. Through controlled studies across model compression, token sparsification, and action sequence compression, we make several observations that challenge common assumptions. (1) Methods that reduce computation under conventional metrics often increase end-to-end execution cost or degrade motion quality, despite maintaining task success rates. (2) System-level embodied efficiency metrics reveal performance differences in the learned action policies that remain hidden under conventional evaluations. (3) Common adaptation methods such as in-context prompting or supervised fine-tuning show only mild and metric-specific improvements in embodied efficiency. While these methods can reduce targeted embodied-efficiency metrics such as jerk or action rate, the resulting gains may come with trade-offs in other metrics, such as longer completion time. Taken together, our results suggest that conventional inference efficiency metrics can overlook important aspects of embodied execution. Incorporating embodied efficiency provides a more complete view of policy behavior and practical performance, enabling fairer and more comprehensive comparisons of VLA models.

评分理由: 论文核心研究大语言模型的对齐机制，特别是政治审查场景下的检测-路由-生成框架，与"Large Language Models"和"Alignment"高度相关（10分）。论文涉及事实性输出和幻觉问题，与"Hallucination Mitigation"有一定关联（5分）。通过探针、消融等方法研究模型内部机制，与"Mechanistic Interpretability"有一定关联（5分）。其他关键词如MoE、SFT、RAG等未在论文中涉及，评0分。

当前的对齐评估主要衡量模型是否编码危险概念以及是否拒绝有害请求。这两者都忽略了对齐通常运作的层面：从概念检测到行为策略的路由机制。我们以中国源语言模型中的政治审查作为自然实验，通过对来自五个实验室的九个开放权重模型进行探针分析、手术式消融和行为测试，得出三项发现。首先，仅凭探针准确性无法作为诊断依据：政治探针、空值对照和置换基线均可达到100%准确率，因此留出类别泛化能力才是有效的测试标准。其次，手术式消融揭示了实验室特定的路由机制。移除政治敏感性方向在大多数测试模型中消除了审查并恢复了准确的事实输出，而有一个模型因架构将事实知识与审查机制纠缠而产生虚构输出。跨模型迁移失败，表明路由几何结构具有模型和实验室特异性。第三，拒绝不再是主导的审查机制。在同一个模型系列中，强硬拒绝率降至零，而叙事引导升至最高，使得仅依赖拒绝检测的基准测试无法察觉审查行为。这些结果支持一个三阶段描述框架：检测、路由、生成。模型通常保留相关知识；对齐改变的是这些知识的表达方式。因此，仅审计检测或拒绝的评估方法会遗漏最直接决定行为的路由机制。

Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

评分理由: 论文提出MoRI框架，旨在提升LLM在科学构思任务中的推理能力。核心相关关键词包括：1) “Large Language Models” (论文基于LLM构建框架)；2) “Post-training” (使用监督微调初始化模型)；3) “RLHF” (采用复合强化学习奖励进行训练)；4) “Chain of Thought” (框架显式学习从研究动机到方法的推理过程)；5) “System 2 Thinking” (强调深度推理，避免表面概念重组)；6) “LLM Agents” (论文属于LLM智能体研究，改进现有智能体方法)；7) “AI for Science” (应用于科学构思领域，属于AI for Science范畴)。其他关键词如MoE、量化、RAG等未在论文中涉及，故评分为0。

科学构思旨在给定科学背景下提出新颖解决方案。现有基于大语言模型（LLM）的智能体方法虽模拟人类研究流程，却未能充分建模科学推理过程，导致其产出多为缺乏技术深度与科学依据的表层概念重组。为解决这一问题，我们提出 MoRI（基于动机的科学构思推理框架），该框架使大语言模型能够显式学习从研究动机到方法论的推理过程。基础大语言模型首先通过监督微调进行初始化，以从给定情境中生成研究动机，随后在复合强化学习奖励机制下进行训练，以逼近科学严谨性：（1）熵感知信息增益鼓励模型基于真实方法论揭示并阐述高复杂度的技术细节；（2）对比语义增益约束推理轨迹，确保其与科学有效解决方案保持概念一致。实验结果表明，MoRI 在创新性、技术严谨性和可行性等多个维度上显著优于主流商用大语言模型及复杂智能体基线方法。代码将在 \href{https://github.com/ECNU-Text-Computing/IdeaGeneration}{GitHub} 平台开源。

Scientific ideation aims to propose novel solutions within a given scientific context. Existing LLM-based agentic approaches emulate human research workflows, yet inadequately model scientific reasoning, resulting in surface-level conceptual recombinations that lack technical depth and scientific grounding. To address this issue, we propose \textbf{MoRI} (\textbf{Mo}tivation-grounded \textbf{R}easoning for Scientific \textbf{I}deation), a framework that enables LLMs to explicitly learn the reasoning process from research motivations to methodologies. The base LLM is initialized via supervised fine-tuning to generate a research motivation from a given context, and is subsequently trained under a composite reinforcement learning reward that approximates scientific rigor: (1) entropy-aware information gain encourages the model to uncover and elaborate high-complexity technical details grounded in ground-truth methodologies, and (2) contrastive semantic gain constrains the reasoning trajectory to maintain conceptually aligned with scientifically valid solutions. Empirical results show that MoRI significantly outperforms strong commercial LLMs and complex agentic baselines across multiple dimensions, including novelty, technical rigor, and feasibility. The code will be made available on \href{https://github.com/ECNU-Text-Computing/IdeaGeneration}{GitHub}.

评分理由: 论文核心研究LLM的推理能力提升，提出TARo方法在推理时对齐，属于大模型技术原理创新。高度相关的关键词包括：LLMs（核心研究对象）、Alignment（测试时对齐方法）、CoT Reasoning（提升推理性能）、System 2 Thinking（涉及深度推理）。有一定相关的关键词：Post-training（对比传统后训练方法）、RLHF（属于对齐技术范畴）、AI for Science（应用于临床推理）。其余关键词与论文内容无关。

大语言模型（LLMs）展现出强大的推理能力，但通常需要昂贵的后训练才能达到高性能。近期的测试时对齐方法提供了一种轻量级替代方案，但主要被探索用于偏好对齐而非推理任务。为填补这一空白，我们提出了令牌级自适应路由（Token-level Adaptive Routing, TARo），该方法在完全保持基础模型冻结的状态下，于推理阶段引导模型进行结构化推理。具体而言，我们首先在分步数学推导轨迹上训练奖励模型，以捕捉细粒度的逻辑一致性信号；随后引入一个可学习的令牌级路由器，自动控制奖励模型对基础模型的引导强度。大量实验表明，TARo在推理性能上相比基础模型显著提升高达+22.4%，较现有令牌级测试时对齐方法提升+8.4%，同时还能增强分布外临床推理（MedXpertQA）和指令遵循（AlpacaEval）能力。此外，TARo无需重新训练即可从小型骨干模型泛化至大型骨干模型，从而将测试时对齐的应用范围从偏好优化扩展至鲁棒的跨领域推理。

Large language models (LLMs) exhibit strong reasoning capabilities but typically require expensive post-training to reach high performance. Recent test-time alignment methods offer a lightweight alternative, but have been explored mainly for preference alignment rather than reasoning. To bridge this gap, we propose, Token-level Adaptive Routing (TARo), which steers frozen LLMs toward structured reasoning entirely at inference time. Specifically, we first train reward models on step-wise mathematical traces to capture fine-grained logical consistency signals, then introduce a learnable token-level router that automatically controls the guidance of the reward model to the base model. Extensive experiments show that TARo significantly improves reasoning performance by up to +22.4% over base model and +8.4% over existing token-level test-time alignment methods, while also boosting out-of-distribution clinical reasoning (MedXpertQA) and instruction following (AlpacaEval). Furthermore, TARo also generalizes from small to large backbones without retraining, extending test-time alignment from preference optimization to robust, cross-domain reasoning.

评分理由: 论文核心研究工具增强大语言模型（LLMs）中推理与外部行动的耦合问题，因此与"Large Language Models"、“Tool Use”、“LLM Agents"高度相关（10分）。论文强调需要"in-depth reasoning"和"multi-step reasoning”，与"System 2 Thinking"和"Chain of Thought"高度相关（10分）。其他关键词如MoE、量化、对齐、RAG等未在论文中涉及，得0分。

工具增强型大语言模型（LLM）必须将多步推理与外部行动紧密耦合，然而现有基准测试常因复杂的环境动态、记忆知识或数据集污染而混淆这种相互作用。本文提出ZebraArena——一个通过程序化生成、用于研究工具增强型LLM中推理-行动耦合的诊断性环境，其具备可控难度和知识最小化设计，能有效限制模型从记忆或数据集污染中获益。ZebraArena中的每个任务都需要一组关键信息，这些信息仅能通过针对性工具调用获取，从而在外部信息获取与演绎推理之间构建了可解释的接口。该设计通过唯一解实现确定性评估，并提供了理论最优查询次数以衡量工具使用效率。我们证明ZebraArena要求深度推理与精准外部工具调用的结合，这对前沿推理模型（如GPT-5和Gemini 2.5 Pro）仍具挑战性——它们在困难实例上仅达到60%的准确率。我们还观察到理论最优性与实际工具使用之间存在持续差距：例如GPT-5的工具调用次数比理论最优值高出70-270%。本文重点阐述了评估中的关键发现，期望ZebraArena能推动关于内部推理与外部行动交互机制的进一步研究。

Tool-augmented large language models (LLMs) must tightly couple multi-step reasoning with external actions, yet existing benchmarks often confound this interplay with complex environment dynamics, memorized knowledge or dataset contamination. In this paper, we introduce ZebraArena, a procedurally generated diagnostic environment for studying reasoning-action coupling in tool-augmented LLMs, with controllable difficulty and a knowledge-minimal design, which limits gains from memorization or dataset contamination. Each task in ZebraArena requires a set of critical information which is available only through targeted tool use, yielding an interpretable interface between external information acquisition and deductive reasoning. This design provides deterministic evaluation via unique solutions, and a theoretical optimal query count for measuring efficient tool use. We show that ZebraArena requires a combination of in-depth reasoning and accurate external tool calling, which remains a challenge as frontier reasoning models such as GPT-5 and Gemini 2.5 Pro only achieves 60% accuracy on the hard instances. We also observe a persistent gaps between theoretical optimality and practical tool usage. For example, GPT-5 uses 70-270% more tool calls than the theoretical optimum. We highlight the key findings in our evaluation, and hope ZebraArena stimulates further research on the interplay between internal reasoning and external action.

评分理由: 论文核心研究LLMs的无监督强化学习微调方法（RLIF），提出PowerFlow框架，通过分布匹配优化LLMs的双重能力（逻辑推理与创造性）。高度相关关键词：LLMs（核心研究对象）、Post-training/SFT（涉及微调）、RLHF/RLAIF/DPO（属于强化学习微调范畴）。中等相关关键词：Instruction Tuning/Alignment（涉及模型能力对齐）、Chain of Thought/System 2 Thinking（涉及逻辑推理）、Self-Correction/Self-Improvement（涉及模型自我优化）。其他关键词未在论文中直接涉及。

无监督内部反馈强化学习已成为一种无需外部监督即可激发大型语言模型潜在能力的前景广阔的研究范式。然而，现有方法依赖于启发式内在奖励，这些奖励通常缺乏明确的理论优化目标，且容易产生退化性偏差。本研究提出PowerFlow，一个将无监督微调重新定义为分布匹配问题的原理性框架。通过将GFlowNet构建为非归一化密度的摊销变分采样器，我们提出了一种长度感知的轨迹平衡目标，该目标能显式地抵消自回归生成中固有的结构长度偏差。通过以$α$-幂分布为目标，PowerFlow能够定向激发大型语言模型的双重特性：通过锐化分布（$α> 1$）来强化逻辑推理能力，或通过平坦化分布（$α< 1$）来释放表达创造力。大量实验表明，PowerFlow在各项任务中持续优于现有RLIF方法，其表现达到甚至超越了有监督的GRPO。此外，通过缓解对齐模型中的过度锐化现象，我们的方法在多样性与质量上实现了同步提升，从而在创造性任务中推动了帕累托边界的演进。

Unsupervised Reinforcement Learning from Internal Feedback (RLIF) has emerged as a promising paradigm for eliciting the latent capabilities of Large Language Models (LLMs) without external supervision. However, current methods rely on heuristic intrinsic rewards, which often lack a well-defined theoretical optimization target and are prone to degenerative biases. In this work, we introduce PowerFlow, a principled framework that reformulates unsupervised fine-tuning as a distribution matching problem. By casting GFlowNet as an amortized variational sampler for unnormalized densities, we propose a length-aware Trajectory-Balance objective that explicitly neutralizes the structural length biases inherent in autoregressive generation. By targeting $α$-power distributions, PowerFlow enables the directional elicitation of the dual nature of LLMs: sharpening the distribution ($α> 1$) to intensify logical reasoning, or flattening it ($α< 1$) to unlock expressive creativity. Extensive experiments demonstrate that PowerFlow consistently outperforms existing RLIF methods, matching or even exceeding supervised GRPO. Furthermore, by mitigating over-sharpening in aligned models, our approach achieves simultaneous gains in diversity and quality, shifting the Pareto frontier in creative tasks.

评分理由: 论文核心是提出一个强化学习框架（LSE）来训练大语言模型（LLMs）在测试时自我进化，通过迭代优化上下文来提升在新问题上的表现。这与"Large Language Models"高度相关（核心研究对象）。方法基于强化学习，将多步进化问题简化为单步RL目标，与"RLHF"等关键词高度相关（核心方法）。论文涉及模型自我改进和反思，与"Self-Correction"等关键词高度相关。论文提到模型迭代优化上下文以提升性能，涉及多步推理和深入思考，与"Chain of Thought"和"System 2 Thinking"有一定关联（5分）。论文在测试时通过上下文优化来提升性能，与"In-context Learning"有一定关联（5分）。其他关键词如MoE、量化、RAG等未在摘要中提及，评为0分。

我们提出“学习式自我进化”框架，这是一种强化学习框架，用于训练大语言模型在测试阶段优化其自身上下文。我们将该框架置于测试时自我进化的情境中，使模型能够基于已见问题的反馈迭代优化上下文，从而在新问题上表现更佳。现有方法完全依赖模型固有的推理能力，从未针对此任务进行显式训练。本框架将多步进化问题简化为单步强化学习目标，其中每次上下文编辑的奖励由下游性能提升程度决定。我们将此目标与树状引导进化循环相结合。在文本到SQL生成和通用问答任务上，采用本框架训练的40亿参数模型超越了基于GPT-5与Claude Sonnet 4.5的自进化策略，以及包括GEPA和TextGrad在内的提示优化方法，并且无需额外训练即可迁移指导其他模型。我们的研究结果证明了将自我进化作为可学习技能的有效性。

We introduce Learning to Self-Evolve (LSE), a reinforcement learning framework that trains large language models (LLMs) to improve their own contexts at test time. We situate LSE in the setting of test-time self-evolution, where a model iteratively refines its context from feedback on seen problems to perform better on new ones. Existing approaches rely entirely on the inherent reasoning ability of the model and never explicitly train it for this task. LSE reduces the multi-step evolution problem to a single-step RL objective, where each context edit is rewarded by the improvement in downstream performance. We pair this objective with a tree-guided evolution loop. On Text-to-SQL generation (BIRD) and general question answering (MMLU-Redux), a 4B-parameter model trained with LSE outperforms self-evolving policies powered by GPT-5 and Claude Sonnet 4.5, as well as prompt optimization methods including GEPA and TextGrad, and transfers to guide other models without additional training. Our results highlight the effectiveness of treating self-evolution as a learnable skill.

评分理由: 论文提出TerraScope，一种用于地球观测的视觉语言模型（VLM），专注于像素级地理空间推理。与关键词的相关性分析如下：1）与"Chain of Thought"高度相关（10分），因为论文核心是像素级推理链，并创建了Terra-CoT数据集。2）与"AI for Science"高度相关（10分），属于地球科学领域的AI应用。3）与"Large Language Models"、“Pre-training”、“Supervised Fine-tuning”、“System 2 Thinking”、“Explainable AI"有一定关联（各5分），因为VLM基于大模型技术，涉及训练和推理过程，并提供可解释性。其他关键词如MoE、SLMs、RAG、RLHF等与论文内容无直接关系，得0分。

视觉语言模型（VLMs）在地球观测（EO）领域展现出潜力，但在需要将复杂空间推理与精确像素级视觉表征相锚定的任务中仍面临挑战。为解决这一问题，我们提出了TerraScope——一个能够实现像素锚定地理空间推理的统一视觉语言模型，其具备两项核心能力：（1）模态灵活推理：可处理单模态输入（光学或合成孔径雷达SAR），并在多模态可用时自适应融合不同模态至推理流程；（2）多时序推理：能整合时间序列数据以进行多时相变化分析。此外，我们构建了Terra-CoT数据集，该大规模数据集包含100万个样本，其推理链中嵌入了来自多源数据的像素级掩码。我们还提出了首个像素锚定地理空间推理基准TerraScope-Bench，包含六个子任务，通过同时评估答案准确性与掩码质量来确保真实的像素锚定推理。实验表明，TerraScope在像素锚定地理空间推理任务上显著优于现有视觉语言模型，同时提供可解释的视觉证据。

Vision-language models (VLMs) have shown promise in earth observation (EO), yet they struggle with tasks that require grounding complex spatial reasoning in precise pixel-level visual representations. To address this problem, we introduce TerraScope, a unified VLM that delivers pixel-grounded geospatial reasoning with two key capabilities: (1) modality-flexible reasoning: it handles single-modality inputs (optical or SAR) and adaptively fuses different modalities into the reasoning process when both are available; (2) multi-temporal reasoning: it integrates temporal sequences for change analysis across multiple time points. In addition, we curate Terra-CoT, a large-scale dataset containing 1 million samples with pixel-level masks embedded in reasoning chains across multiple sources. We also propose TerraScope-Bench, the first benchmark for pixel-grounded geospatial reasoning with six sub-tasks that evaluates both answer accuracy and mask quality to ensure authentic pixel-grounded reasoning. Experiments show that TerraScope significantly outperforms existing VLMs on pixel-grounded geospatial reasoning while providing interpretable visual evidence.

评分理由: 论文D-Mem专注于为LLM智能体设计一个双过程记忆系统，以提升长期推理能力。核心相关关键词包括：1) “Large Language Models” (10分)：论文明确使用GPT-4o-mini和Qwen3-235B-Instruct作为实验模型，是核心基础。2) “LLM Agents” (10分)：论文直接研究自主智能体的记忆系统，是核心主题。3) “Retrieval-Augmented Generation” (10分)：论文批判并改进现有的基于检索的记忆框架，是核心技术背景。4) “Chain of Thought"和"System 2 Thinking” (各5分)：论文涉及长期推理和精细上下文理解，与多步推理和深度思考概念相关，但非直接技术实现。其他关键词如MoE、量化、对齐等与论文内容无直接关联，故评0分。

在持久性自适应智能体发展的推动下，为这些系统配备高保真记忆访问能力以实现长程推理已成为关键需求。然而，当前主流的基于检索的记忆框架通常采用增量处理范式，持续提取对话记忆并更新至向量数据库，在查询时依赖语义检索。虽然这种方法速度较快，但其本质上依赖于有损抽象，常常遗漏上下文关键信息，且难以处理需要细粒度上下文理解的查询。为此，我们提出了D-Mem——一种双过程记忆系统。该系统保留轻量级向量检索以处理常规查询，同时建立详尽的完全审议模块作为高保真备用机制。为实现认知效率与精度的平衡，D-Mem采用多维质量门控策略动态桥接这两个过程。基于GPT-4o-mini和Qwen3-235B-Instruct模型在LoCoMo和RealTalk基准上的实验验证了本方法的有效性。值得注意的是，我们的多维质量门控策略在LoCoMo基准上使用GPT-4o-mini实现了53.5的F1分数，优于静态检索基线Mem0$^\ast$（51.2），并恢复了完全审议模块96.7%的性能表现（55.3），同时显著降低了计算成本。

Driven by the development of persistent, self-adapting autonomous agents, equipping these systems with high-fidelity memory access for long-horizon reasoning has emerged as a critical requirement. However, prevalent retrieval-based memory frameworks often follow an incremental processing paradigm that continuously extracts and updates conversational memories into vector databases, relying on semantic retrieval when queried. While this approach is fast, it inherently relies on lossy abstraction, frequently missing contextually critical information and struggling to resolve queries that rely on fine-grained contextual understanding. To address this, we introduce D-Mem, a dual-process memory system. It retains lightweight vector retrieval for routine queries while establishing an exhaustive Full Deliberation module as a high-fidelity fallback. To achieve cognitive economy without sacrificing accuracy, D-Mem employs a Multi-dimensional Quality Gating policy to dynamically bridge these two processes. Experiments on the LoCoMo and RealTalk benchmarks using GPT-4o-mini and Qwen3-235B-Instruct demonstrate the efficacy of our approach. Notably, our Multi-dimensional Quality Gating policy achieves an F1 score of 53.5 on LoCoMo with GPT-4o-mini. This outperforms our static retrieval baseline, Mem0$^\ast$ (51.2), and recovers 96.7% of the Full Deliberation’s performance (55.3), while incurring significantly lower computational costs.

评分理由: 论文提出EntropyCache，一种针对扩散语言模型（dLLMs）的无训练KV缓存方法，核心是解决KV缓存效率问题。与关键词高度相关的包括：1）“KV Cache Compression” OR “Linear Attention” OR “FlashAttention”（15分）：论文直接研究KV缓存优化方法，是核心内容；2）“Large Language Models” OR “LLMs” OR “Foundation Models”（10分）：论文针对扩散大语言模型（dLLMs）进行优化；3）“Speculative Decoding” OR “Inference Acceleration”（10分）：论文目标是通过KV缓存实现推理加速；4）“Chain of Thought” OR “CoT Reasoning” OR “Multi-step Reasoning”（5分）：论文在CoT基准测试上评估了方法。其他关键词如MoE、量化、对齐等与论文内容无关。

基于扩散的大型语言模型（dLLMs）依赖双向注意力机制，这阻碍了无损KV缓存的实现，并需要在每个去噪步骤中进行完整的前向传播。现有的近似KV缓存方法通过选择性更新缓存状态来降低计算成本，但其决策开销随上下文长度或模型深度增加而线性增长。我们提出EntropyCache，一种无需训练的KV缓存方法，该方法以新解码词元分布的最大熵作为恒定开销的信号，用于决定何时重新计算。我们的设计基于两项实证观察：（1）解码词元的熵与KV缓存漂移相关，为缓存陈旧性提供了低成本的代理指标；（2）解码词元的特征波动在解掩码后持续多个步骤，这促使我们对最近解码的$k$个词元进行重新计算。跳转或重新计算的决策仅需每步$O(V)$的计算量，与上下文长度和模型规模无关。在LLaDA-8B-Instruct和Dream-7B-Instruct上的实验表明，EntropyCache在标准基准测试中实现了$15.2\times$-$26.4\times$的加速，在思维链基准测试中实现了$22.4\times$-$24.1\times$的加速，同时保持了具有竞争力的准确率，且决策开销仅占推理时间的$0.5%$。代码发布于https://github.com/mscheong01/EntropyCache。

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\times$-$26.4\times$ speedup on standard benchmarks and $22.4\times$-$24.1\times$ on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only $0.5%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

评分理由: 论文核心研究LLM代理在安全环境感知中的行为，直接涉及"Large Language Models"和"LLM Agents”，给予10分。研究涉及多代理协商（“Multi-agent Systems”）和可解释性分析（“Mechanistic Interpretability”），给予5分。论文提到代理依赖上下文窗口形成环境感知，与"Context Window Extension"有一定关联，给予5分。其他关键词如MoE、SFT、RAG等未在论文中涉及，给予0分。

NDAI（非披露人工智能）区域允许发明者与投资者智能体在可信执行环境（TEE）内进行协商，若未达成协议，任何已披露信息将被删除。这使得全面披露知识产权成为发明者智能体的理性策略。然而，利用此基础设施要求智能体能够区分安全与不安全环境，而大型语言模型智能体天生缺乏这种能力，因为它们仅能依赖通过上下文窗口传递的证据来形成对执行环境的认知。我们提出：不同大型语言模型在形成对其执行环境安全性的认知时，如何权衡不同形式的证据？通过在10种语言模型及多种证据场景下进行NDAI式协商任务实验，我们发现一种明显的不对称性：失败的认证普遍抑制所有模型的披露行为，而通过的认证则引发高度异质性响应：部分模型增加披露，部分不受影响，少数模型甚至矛盾地减少披露。这表明当前大型语言模型能够可靠地检测危险信号，但无法可靠地验证安全性——而这正是NDAI区域等隐私保护智能体协议所需的核心能力。通过可解释性分析、针对性微调或改进证据架构等方式弥合这一差距，仍然是部署能够根据实际证据质量校准信息共享的智能体所面临的核心开放挑战。

NDAI zones let inventor and investor agents negotiate inside a Trusted Execution Environment (TEE) where any disclosed information is deleted if no deal is reached. This makes full IP disclosure the rational strategy for the inventor’s agent. Leveraging this infrastructure, however, requires agents to distinguish a secure environment from an insecure one, a capability LLM agents lack natively, since they can rely only on evidence passed through the context window to form awareness of their execution environment. We ask: How do different LLM models weight various forms of evidence when forming awareness of the security of their execution environment? Using an NDAI-style negotiation task across 10 language models and various evidence scenarios, we find a clear asymmetry: a failing attestation universally suppresses disclosure across all models, whereas a passing attestation produces highly heterogeneous responses: some models increase disclosure, others are unaffected, and a few paradoxically reduce it. This reveals that current LLM models can reliably detect danger signals but cannot reliably verify safety, the very capability required for privacy-preserving agentic protocols such as NDAI zones. Bridging this gap, possibly through interpretability analysis, targeted fine-tuning, or improved evidence architectures, remains the central open challenge for deploying agents that calibrate information sharing to actual evidence quality.

评分理由: 论文核心是应用LLM解决VANET网络碎片化问题，因此与"Large Language Models"高度相关（10分）。论文通过四阶段流程将通用LLM转化为领域专家，涉及领域适应和微调，与"Pre-training/Domain Adaptation"和"Post-training/SFT"有一定关联（5分）。LLM用于识别拓扑重要性，体现推理能力，与"Chain of Thought"和"System 2 Thinking"相关（5分）。LLM作为智能体先验指导策略，与"LLM Agents"相关（5分）。其他关键词如MoE、SLMs、Scaling Laws、RAG、Quantization等未涉及，得0分。

车载自组织网络（VANETs）是自动驾驶的数字基石，但在城市环境中，由于物理遮挡，其网络往往面临严重的割裂问题。具备高机动性的无人机（UAVs）已成为弥合这些连接鸿沟的关键解决方案。然而，传统的基于深度强化学习（DRL）的无人机部署策略缺乏对道路拓扑的语义理解，常导致盲目探索和样本效率低下。相比之下，大语言模型（LLMs）拥有强大的推理能力，能够识别拓扑重要性，但将其应用于控制任务仍具挑战性。为此，我们提出了语义增强深度强化学习（SA-DRL）框架。首先，我们提出了一种基于道路拓扑图（RTG）与双连通图（DCG）的网络割裂量化方法。随后，我们设计了一个四阶段流程，将通用大语言模型转化为特定领域的拓扑专家。最后，我们提出了语义增强近端策略优化（SA-PPO）算法，该算法采用逻辑融合机制，将大语言模型的语义推理作为先验知识直接注入策略中，从而有效引导智能体朝向关键路口。大量高保真仿真实验表明，SA-PPO以卓越的效率实现了最先进的性能，仅需26.6%的训练回合数即可达到基线性能水平。最终，相较于其他方法，SA-PPO将两项关键连接指标分别提升了13.2%和23.5%，同时将能耗降至基线水平的28.2%。

Vehicular Ad-hoc Networks (VANETs) are the digital cornerstone of autonomous driving, yet they suffer from severe network fragmentation in urban environments due to physical obstructions. Unmanned Aerial Vehicles (UAVs), with their high mobility, have emerged as a vital solution to bridge these connectivity gaps. However, traditional Deep Reinforcement Learning (DRL)-based UAV deployment strategies lack semantic understanding of road topology, often resulting in blind exploration and sample inefficiency. By contrast, Large Language Models (LLMs) possess powerful reasoning capabilities capable of identifying topological importance, though applying them to control tasks remains challenging. To address this, we propose the Semantic-Augmented DRL (SA-DRL) framework. Firstly, we propose a fragmentation quantification method based on Road Topology Graphs (RTG) and Dual Connected Graphs (DCG). Subsequently, we design a four-stage pipeline to transform a general-purpose LLM into a domain-specific topology expert. Finally, we propose the Semantic-Augmented PPO (SA-PPO) algorithm, which employs a Logit Fusion mechanism to inject the LLM’s semantic reasoning directly into the policy as a prior, effectively guiding the agent toward critical intersections. Extensive high-fidelity simulations demonstrate that SA-PPO achieves state-of-the-art performance with remarkable efficiency, reaching baseline performance levels using only 26.6% of the training episodes. Ultimately, SA-PPO improves two key connectivity metrics by 13.2% and 23.5% over competing methods, while reducing energy consumption to just 28.2% of the baseline.

评分理由: 论文核心是开发一个基于LLM的智能体框架（SignAgent）用于手语标注和数据集构建。高度相关的关键词包括：1）“Large Language Models”（论文明确使用LLMs作为核心组件）；2）“LLM Agents”（论文提出的是agentic framework，属于智能体研究）。中等相关的关键词包括：1）“Chain of Thought”（智能体进行推理和决策，涉及多步推理）；2）“System 2 Thinking”（智能体进行深入推理分析）；3）“Tool Use”（智能体协调使用语言工具）。其他关键词如MoE、量化、RAG等未在论文中涉及，评分为0。

本文介绍了SignAgent，一种新颖的智能体框架，该框架利用大语言模型（LLMs）实现可扩展的、基于语言学基础的手语（Sign Language, SL）标注与数据集构建。传统的手语计算方法通常仅在语素层面运作，忽视了关键的语言学细微差别，而人工语言学标注仍是一个重大瓶颈，其过程缓慢且成本高昂，难以用于创建大规模、具备音系学意识的数据集。SignAgent通过两个核心组件应对这些挑战：SignAgent Orchestrator——一个负责协调一系列语言学工具并进行推理的LLM，以及SignGraph——一个提供词汇与语言学基础的知识驱动型LLM。我们在两项下游标注任务上评估了该框架。首先，在伪语素标注任务中，智能体执行约束性分配，利用多模态证据为手语序列提取并排序合适的语素标签。其次，在身份语素标注任务中，智能体通过推理视觉相似性和音系重叠性来检测并优化视觉聚类，从而正确识别和归类词汇手语变体。我们的结果表明，这种智能体方法在大规模、具备语言学意识的数据标注与构建方面表现出色。

This paper introduces SignAgent, a novel agentic framework that utilises Large Language Models (LLMs) for scalable, linguistically-grounded Sign Language (SL) annotation and dataset curation. Traditional computational methods for SLs often operate at the gloss level, overlooking crucial linguistic nuances, while manual linguistic annotation remains a significant bottleneck, proving too slow and expensive for the creation of large-scale, phonologically-aware datasets. SignAgent addresses these challenges through SignAgent Orchestrator, a reasoning LLM that coordinates a suite of linguistic tools, and SignGraph, a knowledge-grounded LLM that provides lexical and linguistic grounding. We evaluate our framework on two downstream annotation tasks. First, on Pseudo-gloss Annotation, where the agent performs constrained assignment, using multi-modal evidence to extract and order suitable gloss labels for signed sequences. Second, on ID Glossing, where the agent detects and refines visual clusters by reasoning over both visual similarity and phonological overlap to correctly identify and group lexical sign variants. Our results demonstrate that our agentic approach achieves strong performance for large-scale, linguistically-aware data annotation and curation.

评分理由: 论文研究Vision-Language-Action (VLA)模型在机器人平台上的实际效率，属于大模型在具体领域（机器人/具身智能）的应用研究。与关键词的相关性分析：1) “Large Language Models” (5分)：VLA模型是多模态大模型，属于大模型范畴；2) “Supervised Fine-tuning” (5分)：摘要提到supervised fine-tuning作为常见适应方法被评估；3) “LLM Agents” (8分)：VLA模型用于具身智能体，是LLM Agents的具体实现；4) “Quantization” (8分)：论文明确研究model compression（包括量化等压缩技术）；5) “In-context Learning” (5分)：摘要提到in-context prompting作为评估方法。其他关键词如MoE、Scaling Laws、RLHF等未在论文中涉及，得0分。

视觉-语言-动作（Vision-Language-Action, VLA）模型近年来通过联合推理视觉、语言与运动模态，使具身智能体能够执行日益复杂的任务。然而，我们发现当前VLA研究中普遍采用的“效率”概念——通常以参数量、浮点运算量或解码吞吐量来衡量——并不能反映其在机器人平台上的实际性能。在现实世界执行中，效率由系统层面的具身行为决定，例如任务完成时间、轨迹平滑度、累积关节旋转量和运动能耗。通过对模型压缩、令牌稀疏化与动作序列压缩的对照研究，我们得出若干挑战常见假设的观察：（1）在传统指标下减少计算量的方法，尽管能维持任务成功率，却常常增加端到端执行成本或降低运动质量。（2）系统层面的具身效率指标揭示了学习到的动作策略中隐藏的性能差异，这些差异在传统评估中无法显现。（3）常见的适应方法（如上下文提示或有监督微调）对具身效率的提升有限且仅针对特定指标。虽然这些方法能够降低目标具身效率指标（如急动度或动作频率），但由此获得的收益可能以其他指标（如更长的完成时间）为代价。综合而言，我们的研究表明传统推理效率指标可能忽略具身执行的重要方面。引入具身效率评估能够更完整地反映策略行为与实际性能，从而为VLA模型提供更公平、更全面的比较基准。

Vision-Language-Action (VLA) models have recently enabled embodied agents to perform increasingly complex tasks by jointly reasoning over visual, linguistic, and motor modalities. However, we find that the prevailing notion of ``efficiency’’ in current VLA research, characterized by parameters, FLOPs, or token decoding throughput, does not reflect actual performance on robotic platforms. In real-world execution, efficiency is determined by system-level embodied behaviors such as task completion time, trajectory smoothness, cumulative joint rotation, and motion energy. Through controlled studies across model compression, token sparsification, and action sequence compression, we make several observations that challenge common assumptions. (1) Methods that reduce computation under conventional metrics often increase end-to-end execution cost or degrade motion quality, despite maintaining task success rates. (2) System-level embodied efficiency metrics reveal performance differences in the learned action policies that remain hidden under conventional evaluations. (3) Common adaptation methods such as in-context prompting or supervised fine-tuning show only mild and metric-specific improvements in embodied efficiency. While these methods can reduce targeted embodied-efficiency metrics such as jerk or action rate, the resulting gains may come with trade-offs in other metrics, such as longer completion time. Taken together, our results suggest that conventional inference efficiency metrics can overlook important aspects of embodied execution. Incorporating embodied efficiency provides a more complete view of policy behavior and practical performance, enabling fairer and more comprehensive comparisons of VLA models.

评分理由: 论文核心研究大语言模型的对齐机制，特别是政治审查场景下的检测-路由-生成框架，与"Large Language Models"和"Alignment"高度相关（10分）。论文涉及事实性输出和幻觉问题，与"Hallucination Mitigation"有一定关联（5分）。通过探针、消融等方法研究模型内部机制，与"Mechanistic Interpretability"有一定关联（5分）。其他关键词如MoE、SFT、RAG等未在论文中涉及，评0分。

当前的对齐评估主要衡量模型是否编码危险概念以及是否拒绝有害请求。这两者都忽略了对齐通常运作的层面：从概念检测到行为策略的路由机制。我们以中国源语言模型中的政治审查作为自然实验，通过对来自五个实验室的九个开放权重模型进行探针分析、手术式消融和行为测试，得出三项发现。首先，仅凭探针准确性无法作为诊断依据：政治探针、空值对照和置换基线均可达到100%准确率，因此留出类别泛化能力才是有效的测试标准。其次，手术式消融揭示了实验室特定的路由机制。移除政治敏感性方向在大多数测试模型中消除了审查并恢复了准确的事实输出，而有一个模型因架构将事实知识与审查机制纠缠而产生虚构输出。跨模型迁移失败，表明路由几何结构具有模型和实验室特异性。第三，拒绝不再是主导的审查机制。在同一个模型系列中，强硬拒绝率降至零，而叙事引导升至最高，使得仅依赖拒绝检测的基准测试无法察觉审查行为。这些结果支持一个三阶段描述框架：检测、路由、生成。模型通常保留相关知识；对齐改变的是这些知识的表达方式。因此，仅审计检测或拒绝的评估方法会遗漏最直接决定行为的路由机制。

Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

评分理由: 论文核心研究扩散大语言模型（dLLMs）的策略优化方法dTRPO，属于大模型技术原理创新。与"Large Language Models"高度相关（10分），因为论文专门研究dLLMs。与"Instruction Tuning"和"RLHF"等关键词相关（8分），因为论文涉及对齐人类偏好、策略优化，属于对齐和强化学习范畴。其他关键词如MoE、SLMs、Scaling Laws、RAG等与论文内容无直接关联，给0分。

扩散大语言模型（Diffusion Large Language Models, dLLMs）为语言生成引入了一种新范式，同时也带来了使其与人类偏好对齐的新挑战。本研究旨在通过降低轨迹概率计算成本来改进dLLMs的策略优化，从而实现规模化离线策略训练。我们证明：（i）在参考策略正则化下，新解掩码词元的概率比是中间扩散状态概率比的无偏估计；（ii）完整轨迹的概率可以通过对重掩码最终状态进行单次前向传播来有效估计。通过将这两种轨迹约简策略整合到策略优化目标中，我们提出了轨迹约简策略优化方法（dTRPO）。我们在70亿参数的dLLMs上，基于指令遵循和推理基准对dTRPO进行了评估。结果表明，该方法显著提升了当前先进dLLMs的核心性能，在STEM任务上最高获得9.6%的性能提升，在代码任务上最高提升4.3%，在指令遵循任务上最高提升3.0%。此外，得益于其离线、单次前向的特性，dTRPO展现出强大的训练效率，并通过高质量输出实现了生成效率的提升。

Diffusion Large Language Models (dLLMs) introduce a new paradigm for language generation, which in turn presents new challenges for aligning them with human preferences. In this work, we aim to improve the policy optimization for dLLMs by reducing the cost of the trajectory probability calculation, thereby enabling scaled-up offline policy training. We prove that: (i) under reference policy regularization, the probability ratio of the newly unmasked tokens is an unbiased estimate of that of intermediate diffusion states, and (ii) the probability of the full trajectory can be effectively estimated with a single forward pass of a re-masked final state. By integrating these two trajectory reduction strategies into a policy optimization objective, we propose Trajectory Reduction Policy Optimization (dTRPO). We evaluate dTRPO on 7B dLLMs across instruction-following and reasoning benchmarks. Results show that it substantially improves the core performance of state-of-the-art dLLMs, achieving gains of up to 9.6% on STEM tasks, up to 4.3% on coding tasks, and up to 3.0% on instruction-following tasks. Moreover, dTRPO exhibits strong training efficiency due to its offline, single-forward nature, and achieves improved generation efficiency through high-quality outputs.

评分理由: 论文研究Vision-Language Models (VLMs)的注意力机制问题，属于大模型（LLMs）在视觉-语言多模态领域的应用研究，因此与"Large Language Models"相关度较高（8分）。论文通过分析注意力分布来理解模型行为，这属于"Mechanistic Interpretability"范畴，相关度较高（8分）。论文提出的轻量级提示调优方法属于参数高效微调技术，与"PEFT"有一定关联（5分）。其他关键词如MoE、SLMs、Scaling Laws、Alignment、RAG、Quantization等均未在论文中涉及，相关度为0分。

视觉语言模型（Vision-Language Models, VLMs）已被证明存在视觉盲区，即使在需要视觉推理的任务中，也常常未能充分利用其视觉输入。本研究揭示，视觉语言模型具有选择性视觉盲区：即使在不同语言表述要求相同视觉推理的情况下，它们也会根据语言表述的差异，调节对视觉输入的关注程度。通过以视觉注意力作为探针，我们量化了语言表述如何改变对图像注意力的总量与分布。受限的表述形式（如多项选择和是/否问答）与开放式表述相比，会显著降低对图像背景的关注，减少对任务相关区域的聚焦，并将注意力转向无信息量的标记。我们进一步证明，这种注意力分配不当是导致准确性下降和跨表述不一致的主要原因。基于这一机制性发现，我们提出了一种轻量级提示调优方法，该方法使用可学习的标记，以鼓励模型采用在开放式设置中观察到的稳健且基于视觉的注意力模式，从而增强视觉基础能力，并提升在不同表述形式下的性能表现。

Vision-Language Models (VLMs) have been shown to be blind, often underutilizing their visual inputs even on tasks that require visual reasoning. In this work, we demonstrate that VLMs are selectively blind. They modulate the amount of attention applied to visual inputs based on linguistic framing even when alternative framings demand identical visual reasoning. Using visual attention as a probe, we quantify how framing alters both the amount and distribution of attention over the image. Constrained framings, such as multiple choice and yes/no, induce substantially lower attention to image context compared to open-ended, reduce focus on task-relevant regions, and shift attention towards uninformative tokens. We further demonstrate that this attention misallocation is the principal cause of degraded accuracy and cross-framing inconsistency. Building on this mechanistic insight, we introduce a lightweight prompt-tuning method using learnable tokens that encourages the robust, visually grounded attention patterns observed in open-ended settings, improving visual grounding and improving performance across framings.

评分理由: 论文核心研究大语言模型（LLMs）的表示对齐（Alignment）技术，提出了一种隐私保护的线性对齐框架，因此与"Large Language Models"和"Alignment"高度相关（10分）。其他关键词如MoE、SLMs、Scaling Laws、Pre-training、RLHF、RAG、Quantization等均未在论文中涉及，故评0分。

语言模型似乎日益学习到相似的表征，尽管它们在训练目标、架构和数据模态上存在差异。这种独立训练模型之间新兴的兼容性为跨模型与下游任务的对齐带来了新机遇。此外，它也解锁了新的潜在应用领域，例如在安全、隐私或竞争限制禁止直接共享数据或模型的环境中。在本研究中，我们提出了一种隐私保护框架，该框架利用表征收敛性来实现独立语言模型间的跨孤岛推理。该框架通过在共享公共数据集上学习一个仿射变换，并应用同态加密以在推理过程中保护客户端查询。通过仅加密线性对齐和分类操作，该方法在保持强安全保证的同时实现了亚秒级的推理延迟。我们通过一项关于表征收敛性的实证研究来支持此框架，在该研究中，我们在独立模型的最终隐藏状态之间学习线性变换。我们在嵌入分类和分布外检测任务上评估这些跨模型映射，观察到跨模型对的性能下降极小。此外，我们首次证明线性对齐有时能够实现跨独立训练模型的文本生成。

Language models increasingly appear to learn similar representations, despite differences in training objectives, architectures, and data modalities. This emerging compatibility between independently trained models introduces new opportunities for cross-model alignment to downstream objectives. Moreover, it unlocks new potential application domains, such as settings where security, privacy, or competitive constraints prohibit direct data or model sharing. In this work, we propose a privacy-preserving framework that exploits representational convergence to enable cross-silo inference between independent language models. The framework learns an affine transformation over a shared public dataset and applies homomorphic encryption to protect client queries during inference. By encrypting only the linear alignment and classification operations, the method achieves sub-second inference latency while maintaining strong security guarantees. We support this framework with an empirical investigation into representational convergence, in which we learn linear transformations between the final hidden states of independent models. We evaluate these cross-model mappings on embedding classification and out-of-distribution detection, observing minimal performance degradation across model pairs. Additionally, we show for the first time that linear alignment sometimes enables text generation across independently trained models.

评分理由: 该论文的核心是使用大语言模型（LLMs）生成合成数据，并利用这些数据对LLMs进行监督微调（SFT），以提升生成式常识推理（GCR）任务的多样性和质量。因此，与"Large Language Models"和"Post-training"（即SFT）高度相关（10分）。论文未涉及其他关键词的具体技术或概念，故其余关键词评分为0分。

对话智能体不仅需要以高质量（即具备常识性）的回应响应用户，还需考虑多种合理的替代场景，以体现其回应的多样性。尽管训练多样化的常识生成器的需求日益增长，但由于缺乏大规模高质量的多样化常识训练数据集，这一领域的研究进展受到了显著阻碍。由于标注成本高昂，现有的生成式常识推理数据集仅由少量人工标注者创建，覆盖的常识场景范围有限。为弥补这一训练资源缺口，我们提出了一种两阶段方法，首次构建了用于多样化生成式常识推理的合成数据集CommonSyn。实验表明，在不同规模的大语言模型上，基于我们合成数据微调的模型相比原始模型及基于人工标注数据集微调的模型，在生成多样性与质量上均实现了同步提升。

Conversational agents are required to respond to their users not only with high quality (i.e. commonsense bearing) responses, but also considering multiple plausible alternative scenarios, reflecting the diversity in their responses. Despite the growing need to train diverse commonsense generators, the progress of this line of work has been significantly hindered by the lack of large-scale high-quality diverse commonsense training datasets. Due to the high annotation costs, existing Generative Commonsense Reasoning (GCR) datasets are created using a small number of human annotators, covering only a narrow set of commonsense scenarios. To address this training resource gap, we propose a two-stage method to create the first-ever synthetic dataset CommonSyn for diversified (GCR). The model fine-tuned on our synthetic data jointly increase both generation diversity and quality compared with vanilla models and the model fine-tuned on human-crafted dataset across different size Large Language Models (LLMs)

评分理由: 论文核心研究LLMs（Mistral-7B, Llama-3.1-8B, Qwen-2.5-32B）在词汇联想任务中与人类表现的对比，属于大模型在语言学/认知科学领域的应用研究，因此与"Large Language Models"高度相关（10分）。论文比较了不同规模模型（如7B、8B、32B）的表现，涉及模型大小的影响，与"Small Language Models"有一定关联（5分）。研究通过分析LLM的内部词汇表征来理解其工作机制，属于可解释性AI范畴，与"Mechanistic Interpretability"有一定关联（5分）。论文未涉及其他关键词如MoE、训练技术、推理方法、压缩、代理等具体技术，因此这些关键词评分为0分。

大型语言模型（LLM）在文本生成的流畅性方面取得了令人瞩目的成果，但其语言知识的本质——尤其是其内部词汇库在多大程度上类同于人类——仍不明确。本研究通过比较人类与LLM生成的词汇联想，评估模型捕捉人类词汇模式的准确程度。利用来自SWOW数据集的英语提示词-反应词对，以及三种LLM（Mistral-7B、Llama-3.1-8B和Qwen-2.5-32B）在多种温度设置下新生成的联想数据，我们考察了（i）词汇因素（如词频和具体性）对提示词-反应词对的影响，以及（ii）LLM反应相较于人类反应的变异性和典型性。结果表明，所有模型均能反映词频和具体性方面的人类趋势，但在反应变异性和典型性上存在差异。较大模型（如Qwen）倾向于模仿单一的“原型”人类参与者，生成高度典型但变异性极低的反应；而较小模型（如Mistral和Llama）则产生变异性更高但典型性较低的反应。温度设置进一步影响了这种权衡，较高的温度值会增加变异性，但会降低典型性。这些发现揭示了人类与LLM词汇库之间的相似性与差异性，强调在探究LLM词汇表征时，必须考虑模型规模和温度参数的影响。

Large language models (LLMs) achieve impressive results in terms of fluency in text generation, yet the nature of their linguistic knowledge - in particular the human-likeness of their internal lexicon - remains uncertain. This study compares human and LLM-generated word associations to evaluate how accurately models capture human lexical patterns. Using English cue-response pairs from the SWOW dataset and newly generated associations from three LLMs (Mistral-7B, Llama-3.1-8B, and Qwen-2.5-32B) across multiple temperature settings, we examine (i) the influence of lexical factors such as word frequency and concreteness on cue-response pairs, and (ii) the variability and typicality of LLM responses compared to human responses. Results show that all models mirror human trends for frequency and concreteness but differ in response variability and typicality. Larger models such as Qwen tend to emulate a single “prototypical” human participant, generating highly typical but minimally variable responses, while smaller models such as Mistral and Llama produce more variable yet less typical responses. Temperature settings further influence this trade-off, with higher values increasing variability but decreasing typicality. These findings highlight both the similarities and differences between human and LLM lexicons, emphasizing the need to account for model size and temperature when probing LLM lexical representations.

评分理由: 论文DriftGuard专注于联邦学习中的异步数据漂移问题，提出了一种受MoE启发的架构，因此与"Mixture of Experts" OR “MoE” OR “Sparse Models"高度相关（10分）。论文未涉及大语言模型、深度学习技术原理创新或科学领域应用，与其他关键词完全无关（0分）。

在实际的联邦学习（FL）部署中，参与训练的设备上的数据分布会随时间演变。这导致了异步数据漂移，即不同设备在不同时间向不同分布发生偏移。缓解此类漂移具有挑战性：频繁的重新训练会给资源受限的设备带来高昂计算成本，而重新训练频率过低则会降低漂移设备上的性能。我们提出了DriftGuard，一种能够高效适应异步数据漂移的联邦持续学习框架。DriftGuard采用了一种受混合专家（Mixture-of-Experts, MoE）启发的架构，该架构将捕获全局可迁移知识的共享参数与适应特定群体分布的局部参数分离开来。这种设计实现了两种互补的重新训练策略：（i）全局重新训练，即在识别出系统范围漂移时更新共享参数；（ii）群体重新训练，即通过MoE门控模式识别出的设备集群，选择性地更新其局部参数，而无需共享原始数据。在多个数据集和模型上的实验表明，DriftGuard在达到或超越最先进准确率的同时，能将总重新训练成本降低高达83%。因此，它实现了单位重新训练成本下的最高准确率，比最强基线提升了高达2.3倍。DriftGuard可从https://github.com/blessonvar/DriftGuard下载。

In real-world Federated Learning (FL) deployments, data distributions on devices that participate in training evolve over time. This leads to asynchronous data drift, where different devices shift at different times and toward different distributions. Mitigating such drift is challenging: frequent retraining incurs high computational cost on resource-constrained devices, while infrequent retraining degrades performance on drifting devices. We propose DriftGuard, a federated continual learning framework that efficiently adapts to asynchronous data drift. DriftGuard adopts a Mixture-of-Experts (MoE) inspired architecture that separates shared parameters, which capture globally transferable knowledge, from local parameters that adapt to group-specific distributions. This design enables two complementary retraining strategies: (i) global retraining, which updates the shared parameters when system-wide drift is identified, and (ii) group retraining, which selectively updates local parameters for clusters of devices identified via MoE gating patterns, without sharing raw data. Experiments across multiple datasets and models show that DriftGuard matches or exceeds state-of-the-art accuracy while reducing total retraining cost by up to 83%. As a result, it achieves the highest accuracy per unit retraining cost, improving over the strongest baseline by up to 2.3x. DriftGuard is available for download from https://github.com/blessonvar/DriftGuard.

评分理由: 论文主要研究欧盟AI监管框架，聚焦于安全、隐私和Agentic AI的监管定义、区分和规定分析，属于AI治理和政策研究范畴。所有关键词均涉及大模型/深度学习的技术原理、方法或应用，而本文完全不涉及这些具体技术，仅宏观讨论AI（特别是Agentic AI）的监管问题。唯一略有相关的是"LLM Agents” OR “Autonomous Agents” OR “Agentic Workflow”，因为论文提到了"agentic AI"和"autonomous agents"，但讨论的是其监管而非技术实现，因此给5分（有一定关联）。其余关键词与论文内容完全无关，均给0分。

人工智能（AI）技术的快速扩散催生了动态变化的监管格局，立法框架正努力跟上技术发展的步伐。随着AI范式向更高自主性转变，特别是以智能体AI（agentic AI）的形式出现，精确阐明监管规定变得日益困难。这一挑战在安全与隐私领域尤为突出，因为自主智能体的能力往往模糊了传统的法律与技术边界。本文通过分析2024年至2025年间发布的24份相关文件，梳理了欧盟（EU）AI监管条款的演进脉络。基于此综述，我们对关键定义进行了澄清，解构了监管层面对安全、隐私及智能体AI的阐释，并将其与易混淆概念区分以消除歧义。我们综合所评述的文件，阐明了当前针对不同类型AI（尤其是涉及安全与隐私方面）的监管条款现状，并通过分析反思监管维度中的现有规定，以促使安全与隐私义务更好地契合AI及智能体行为。这些见解旨在为政策制定者、开发者和研究者提供参考，以应对算法主体日益增多的社会中的合规与AI治理问题。

The rapid proliferation of artificial intelligence (AI) technologies has led to a dynamic regulatory landscape, where legislative frameworks strive to keep pace with technical advancements. As AI paradigms shift towards greater autonomy, specifically in the form of agentic AI, it becomes increasingly challenging to precisely articulate regulatory stipulations. This challenge is even more acute in the domains of security and privacy, where the capabilities of autonomous agents often blur traditional legal and technical boundaries. This paper reviews the evolving European Union (EU) AI regulatory provisions via analyzing 24 relevant documents published between 2024 and 2025. From this review, we provide a clarification of critical definitions. We deconstruct the regulatory interpretations of security, privacy, and agentic AI, distinguishing them from closely related concepts to resolve ambiguity. We synthesize the reviewed documents to articulate the current state of regulatory provisions targeting different types of AI, particularly those related to security and privacy aspects. We analyze and reflect on the existing provisions in the regulatory dimension to better align security and privacy obligations with AI and agentic behaviors. These insights serve to inform policymakers, developers, and researchers on the compliance and AI governance in the society with increasing algorithmic agencies.

评分理由: 论文F2LLM-v2专注于多语言嵌入模型，与LLM技术高度相关（10分），因为它基于LLM构建嵌入模型；与小型语言模型相关（8分），因为提供了80M到14B的多种尺寸模型；与预训练相关（8分），因为涉及两阶段LLM嵌入训练流程；与数据质量相关（5分），因为使用了6000万高质量数据样本；其他关键词如MoE、SFT、RLHF、RAG等与嵌入模型训练和评估无直接关联，故得0分。

我们推出F2LLM-v2系列——一套包含8种不同规模（从8000万到140亿参数）的通用多语言嵌入模型。该系列模型基于新构建的6000万公开高质量数据样本复合集进行训练，支持超过200种语言，尤其关注以往资源不足的中低资源语言。通过融合基于大语言模型（LLM）的两阶段嵌入训练流程与套娃学习、模型剪枝及知识蒸馏技术，我们实现了比以往基于LLM的嵌入模型更高效的架构，同时保持了具有竞争力的性能。大量评估证实，F2LLM-v2-14B在11项MTEB基准测试中位列榜首，而该系列中的较小规模模型也为资源受限场景设立了新的性能标杆。为促进开源嵌入模型研究，我们公开了全部模型、数据、代码及中间检查点。

We present F2LLM-v2, a new family of general-purpose, multilingual embedding models in 8 distinct sizes ranging from 80M to 14B. Trained on a newly curated composite of 60 million publicly available high-quality data samples, F2LLM-v2 supports more than 200 languages, with a particular emphasis on previously underserved mid- and low-resource languages. By integrating a two-stage LLM-based embedding training pipeline with matryoshka learning, model pruning, and knowledge distillation techniques, we present models that are far more efficient than previous LLM-based embedding models while retaining competitive performances. Extensive evaluations confirm that F2LLM-v2-14B ranks first on 11 MTEB benchmarks, while the smaller models in the family also set a new state of the art for resource-constrained applications. To facilitate open-source embedding model research, we release all models, data, code, and intermediate checkpoints.

评分理由: 论文核心是评估LLMs在金融推理任务上的表现，因此与’Large Language Models’高度相关（10分）。论文明确测试了’Retrieval-Augmented Generation’设置，因此该关键词得10分。论文涉及金融决策推理，与’Chain of Thought’和’System 2 Thinking’有一定关联（各5分），但并非核心方法。其他关键词如MoE、SFT、RLHF等未在摘要中提及，与论文内容无关，得0分。

现实世界中的金融决策是一个具有挑战性的问题，需要对异质信号进行推理，这些信号包括从监管文件中提取的公司基本面数据以及基于价格动态计算的交易信号。近年来，随着大语言模型（LLMs）的发展，金融分析师已开始将其应用于金融决策任务。然而，现有用于测试这些模型的金融问答基准主要关注公司资产负债表数据，很少评估模型对公司股票在市场中如何交易或其与基本面相互作用的推理能力。为结合两种方法的优势，我们提出了FinTradeBench——一个用于评估融合公司基本面与交易信号的金融推理能力的基准。FinTradeBench包含基于纳斯达克100指数成分股在十年历史窗口期内的1,400个问题。该基准分为三个推理类别：以基本面为核心的问题、以交易信号为核心的问题，以及需要跨信号推理的混合型问题。为确保大规模评估的可靠性，我们采用“校准-扩展”框架，该框架结合了专家种子问题、多模型响应生成、模型内自过滤、数值审计以及人类-LLM评判对齐机制。我们在零样本提示和检索增强设置下评估了14个大语言模型，并观察到明显的性能差距。检索显著提升了基于文本基本面的推理能力，但对交易信号推理的改善有限。这些发现凸显了当前大语言模型在数值与时间序列推理方面的根本性挑战，并为未来金融智能研究指明了方向。

Real-world financial decision-making is a challenging problem that requires reasoning over heterogeneous signals, including company fundamentals derived from regulatory filings and trading signals computed from price dynamics. Recently, with the advancement of Large Language Models (LLMs), financial analysts have begun to use them for financial decision-making tasks. However, existing financial question answering benchmarks for testing these models primarily focus on company balance sheet data and rarely evaluate reasoning over how company stocks trade in the market or their interactions with fundamentals. To take advantage of the strengths of both approaches, we introduce FinTradeBench, a benchmark for evaluating financial reasoning that integrates company fundamentals and trading signals. FinTradeBench contains 1,400 questions grounded in NASDAQ-100 companies over a ten-year historical window. The benchmark is organized into three reasoning categories: fundamentals-focused, trading-signal-focused, and hybrid questions requiring cross-signal reasoning. To ensure reliability at scale, we adopt a calibration-then-scaling framework that combines expert seed questions, multi-model response generation, intra-model self-filtering, numerical auditing, and human-LLM judge alignment. We evaluate 14 LLMs under zero-shot prompting and retrieval-augmented settings and witness a clear performance gap. Retrieval substantially improves reasoning over textual fundamentals, but provides limited benefit for trading-signal reasoning. These findings highlight fundamental challenges in the numerical and time-series reasoning for current LLMs and motivate future research in financial intelligence.

评分理由: 论文NavTrust专注于具身导航（Embodied Navigation）的鲁棒性基准测试，研究内容为评估导航模型在输入模态（RGB、深度、指令）被破坏时的性能下降，并提出缓解策略。所有评分关键词均涉及大模型/深度学习的技术原理、训练方法、推理优化、对齐、应用等具体方向，而本论文的核心是导航基准测试和鲁棒性评估，不涉及大模型技术原理创新，也未在生物医药等科学领域应用大模型。虽然论文涉及AI代理（导航代理），但未使用LLM代理或相关技术，因此与所有关键词完全无关。

具身导航主要分为两大类别：视觉语言导航（VLN），即智能体通过遵循自然语言指令进行导航；以及目标导向导航（OGN），即智能体导航至指定的目标物体。然而，现有工作主要在理想条件下评估模型性能，忽视了现实场景中可能出现的各类干扰。为填补这一空白，我们提出了NavTrust，这是一个统一的基准测试框架，能够在真实场景中系统性地对RGB、深度及指令等输入模态施加干扰，并评估其对导航性能的影响。据我们所知，NavTrust是首个在统一框架内，将具身导航智能体暴露于多样化RGB-深度干扰与指令变体下的基准测试。通过对七种前沿方法进行广泛评估，我们发现这些方法在现实干扰下均出现显著的性能下降，这揭示了其关键的鲁棒性缺陷，并为构建更可信的具身导航系统提供了路线图。此外，我们系统性地评估了四种不同的增强策略，以提升模型对RGB-深度干扰及指令干扰的鲁棒性。我们的基础模型包括Uni-NaVid和ETPNav。我们将这些模型部署于真实移动机器人上，并观察到了其对干扰的鲁棒性提升。项目网站为：https://navtrust.github.io。

There are two major categories of embodied navigation: Vision-Language Navigation (VLN), where agents navigate by following natural language instructions; and Object-Goal Navigation (OGN), where agents navigate to a specified target object. However, existing work primarily evaluates model performance under nominal conditions, overlooking the potential corruptions that arise in real-world settings. To address this gap, we present NavTrust, a unified benchmark that systematically corrupts input modalities, including RGB, depth, and instructions, in realistic scenarios and evaluates their impact on navigation performance. To our best knowledge, NavTrust is the first benchmark that exposes embodied navigation agents to diverse RGB-Depth corruptions and instruction variations in a unified framework. Our extensive evaluation of seven state-of-the-art approaches reveals substantial performance degradation under realistic corruptions, which highlights critical robustness gaps and provides a roadmap toward more trustworthy embodied navigation systems. Furthermore, we systematically evaluate four distinct mitigation strategies to enhance robustness against RGB-Depth and instructions corruptions. Our base models include Uni-NaVid and ETPNav. We deployed them on a real mobile robot and observed improved robustness to corruptions. The project website is: https://navtrust.github.io.

评分理由: 论文专注于3D生成领域，提出了一种基于语义的部件级文本到3D生成框架DreamPartGen。虽然该研究属于AI应用范畴，但所有给定的关键词均围绕大语言模型（LLMs）及其相关技术（如训练方法、推理优化、对齐、压缩等）或特定科学领域应用（如生物信息学）。论文内容完全不涉及LLMs、MoE、量化、推理加速、对齐技术等，也未应用于生物信息学或化学信息学。因此，所有关键词的相关度均为0分。

将三维物体理解为由有意义部件组成的结构，是人类感知与推理的基础。然而，大多数文本到三维生成方法忽视了部件的语义与功能结构。尽管近期部分研究引入了部件感知的分解方法，但这些方法仍主要聚焦于几何形态，缺乏语义层面的基础，未能有效建模部件如何与文本描述对齐或部件间的相互关系。我们提出了DreamPartGen，这是一个基于语义的、部件感知的文本到三维生成框架。DreamPartGen引入了双重部件隐变量（Duplex Part Latents, DPLs），以联合建模每个部件的几何形状与外观；同时提出关系语义隐变量（Relational Semantic Latents, RSLs），用于捕捉从语言中推导出的部件间依赖关系。通过同步的协同去噪过程，框架强化了几何与语义的一致性，从而实现了连贯、可解释且与文本对齐的三维合成。在多个基准测试中，DreamPartGen在几何保真度与文本-形状对齐方面均达到了最先进的性能水平。

Understanding and generating 3D objects as compositions of meaningful parts is fundamental to human perception and reasoning. However, most text-to-3D methods overlook the semantic and functional structure of parts. While recent part-aware approaches introduce decomposition, they remain largely geometry-focused, lacking semantic grounding and failing to model how parts align with textual descriptions or their inter-part relations. We propose DreamPartGen, a framework for semantically grounded, part-aware text-to-3D generation. DreamPartGen introduces Duplex Part Latents (DPLs) that jointly model each part’s geometry and appearance, and Relational Semantic Latents (RSLs) that capture inter-part dependencies derived from language. A synchronized co-denoising process enforces mutual geometric and semantic consistency, enabling coherent, interpretable, and text-aligned 3D synthesis. Across multiple benchmarks, DreamPartGen delivers state-of-the-art performance in geometric fidelity and text-shape alignment.

评分理由: 论文核心研究LLM（Nemotron-Cascade 2，一个30B MoE模型）的post-training技术，特别是Cascade RL和multi-domain on-policy distillation，以提升推理和智能体能力。高度相关的关键词包括：LLMs（核心研究对象）、MoE（模型架构）、Post-training/SFT（关键训练阶段）、RLHF/DPO（Cascade RL属于强化学习对齐方法）、Chain of Thought/System 2 Thinking（论文强调推理能力）、LLM Agents（论文提到agentic capabilities）。其他关键词如Scaling Laws、Pre-training、Self-Correction、Tool Use、In-context Learning有间接关联（论文涉及模型规模、训练过程、能力提升），评5分。剩余关键词如RAG、Context Window Extension、Quantization等未在摘要中提及，评0分。

我们推出Nemotron-Cascade 2，这是一个拥有300亿参数、30亿激活参数的开放混合专家（MoE）模型，具备顶尖的推理能力和强大的智能体（agentic）性能。尽管模型规模紧凑，其在数学与代码推理任务上的表现已接近前沿开放模型水平。作为继DeepSeekV3.2-Speciale-671B-A37B之后第二个开放权重的LLM，该模型在2025年国际数学奥林匹克（IMO）、国际信息学奥林匹克（IOI）及国际大学生程序设计竞赛（ICPC）全球总决赛中均达到金牌级表现，以仅二十分之一的参数量实现了极高的智能密度。相较于Nemotron-Cascade 1，其关键技术进展如下：在基于精细筛选数据集进行监督微调（SFT）后，我们大幅扩展了级联强化学习（Cascade RL）的覆盖范围，使其涵盖更广泛的推理与智能体领域。此外，我们在整个级联强化学习过程中引入了多领域同策略蒸馏技术，从各领域最强的中间教师模型进行知识蒸馏，从而有效恢复基准测试中的性能回归，并持续获得显著的性能提升。我们同步发布了模型检查点与训练数据集合。

We introduce Nemotron-Cascade 2, an open 30B MoE model with 3B activated parameters that delivers best-in-class reasoning and strong agentic capabilities. Despite its compact size, its mathematical and coding reasoning performance approaches that of frontier open models. It is the second open-weight LLM, after DeepSeekV3.2-Speciale-671B-A37B, to achieve Gold Medal-level performance in the 2025 International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals, demonstrating remarkably high intelligence density with 20x fewer parameters. In contrast to Nemotron-Cascade 1, the key technical advancements are as follows. After SFT on a meticulously curated dataset, we substantially expand Cascade RL to cover a much broader spectrum of reasoning and agentic domains. Furthermore, we introduce multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the Cascade RL process, allowing us to efficiently recover benchmark regressions and sustain strong performance gains along the way. We release the collection of model checkpoint and training data.

评分理由: 论文研究纯数学中的代数几何问题（三次曲面上的R-等价），与深度学习、大模型技术完全无关。唯一的相关性在于论文末尾提到使用了生成式AI模型（AlphaEvolve和Gemini 3 Deep Think）辅助证明引理，这属于AI在科学研究中的应用，因此仅与’AI for Science’关键词有微弱关联（5分），其他所有关键词均为0分。

设 $V$ 为在具有良好约化的 $p$ 进域 $k$ 上的一条光滑三次曲面。Swinnerton-Dyer (1981) 证明了 $R$ 等价在 $V(k)$ 上是平凡的，除非 $V$ 是三种特殊类型之一——这些曲面的 $R$ 等价性无法通过证明其万有（容许）等价是平凡的来界定。我们考察了目前已知具有非平凡万有等价的所有曲面 $V$。除了这些曲面对 Swinnerton-Dyer 的方法难以处理之外，我们还注意到，如果这些曲面也具有非平凡的 $R$ 等价，它们将与 Colliot-Thélène 和 Sansuc 关于几何有理曲面的万有挠子 $k$ 有理性的猜想相矛盾。

通过设计研究 $R$ 等价的新方法，我们证明了对于具有全 Eckardt 约化的 2 进曲面（第三种特殊类型，包含了所有已知的非平凡万有等价情形），$R$ 等价是平凡的或指数为 2。对于具体情形，我们确认了平凡性：在 $\mathbb{Q}_2(ζ_3)$ 上的对角三次曲面 $X^3+Y^3+Z^3+ζ_3 T^3=0$——这回答了 Manin (Cubic Forms, 1972) 的一个长期悬而未决的问题——以及具有指数为 2 的万有等价的立方曲面 (Kanevsky, 1982)。

这是源自与生成式人工智能模型（如 AlphaEvolve 和 Gemini 3 Deep Think）为期一年的互动所产生的一系列工作中的第一篇，后者证明了我们的许多引理。我们披露了它们在本论文中使用的时间线和性质，并在另一份配套报告（准备中）中描述了更广泛的 AI 辅助研究计划。

Let $V$ be a smooth cubic surface over a $p$-adic field $k$ with good reduction. Swinnerton-Dyer (1981) proved that $R$-equivalence is trivial on $V(k)$ except perhaps if $V$ is one of three special types–those whose $R$-equivalence he could not bound by proving the universal (admissible) equivalence is trivial. We consider all surfaces $V$ currently known to have non-trivial universal equivalence. Beyond being intractable to Swinnerton-Dyer’s approach, we observe that if these surfaces also had non-trivial $R$-equivalence, they would contradict Colliot-Thélène and Sansuc’s conjecture regarding the $k$-rationality of universal torsors for geometrically rational surfaces. By devising new methods to study $R$-equivalence, we prove that for 2-adic surfaces with all-Eckardt reductions (the third special type, which contains every existing case of non-trivial universal equivalence), $R$-equivalence is trivial or of exponent 2. For the explicit cases, we confirm triviality: the diagonal cubic $X^3+Y^3+Z^3+ζ_3 T^3=0$ over $\mathbb{Q}_2(ζ_3)$–answering a long-standing question of Manin’s (Cubic Forms, 1972)–and the cubic with universal equivalence of exponent 2 (Kanevsky, 1982). This is the first in a series of works derived from a year of interactions with generative AI models such as AlphaEvolve and Gemini 3 Deep Think, with the latter proving many of our lemmas. We disclose the timeline and nature of their use towards this paper, and describe our broader AI-assisted research program in a companion report (in preparation).

评分理由: 论文专注于强化学习（RL）在GUI代理中的奖励函数设计，提出OS-Themis多智能体评论框架和OmniGUIRewardBench基准。所有关键词均与大模型、深度学习技术原理或科学AI应用直接相关，而本文核心是传统RL奖励机制，未涉及大模型、深度学习技术或科学领域应用，因此所有关键词相关度为0。

强化学习（RL）具备提升图形用户界面（GUI）智能体在随机环境中鲁棒性的潜力，但其训练过程对奖励函数的质量高度敏感。现有的奖励方法难以同时实现可扩展性与性能表现。为此，我们提出了OS-Themis，一个可扩展且精确的多智能体评判框架。与单一评判者不同，OS-Themis将任务轨迹分解为可验证的里程碑，以隔离决策所需的关键证据，并采用审查机制在做出最终裁决前严格审核证据链。为便于评估，我们进一步引入了OmniGUIRewardBench（OGRBench），这是一个用于GUI结果奖励的全方位跨平台基准测试，所有被评估模型在OS-Themis框架下均取得了最佳性能。在AndroidWorld平台上进行的大量实验表明，OS-Themis用于支持在线强化学习训练时，性能提升了10.3%；用于自训练循环中的轨迹验证与过滤时，性能提升了6.9%，这凸显了其驱动智能体进化的潜力。

Reinforcement Learning (RL) has the potential to improve the robustness of GUI agents in stochastic environments, yet training is highly sensitive to the quality of the reward function. Existing reward approaches struggle to achieve both scalability and performance. To address this, we propose OS-Themis, a scalable and accurate multi-agent critic framework. Unlike a single judge, OS-Themis decomposes trajectories into verifiable milestones to isolate critical evidence for decision making and employs a review mechanism to strictly audit the evidence chain before making the final verdict. To facilitate evaluation, we further introduce OmniGUIRewardBench (OGRBench), a holistic cross-platform benchmark for GUI outcome rewards, where all evaluated models achieve their best performance under OS-Themis. Extensive experiments on AndroidWorld show that OS-Themis yields a 10.3% improvement when used to support online RL training, and a 6.9% gain when used for trajectory validation and filtering in the self-training loop, highlighting its potential to drive agent evolution.

评分理由: 论文提出Box Maze框架，专注于提高LLM推理的可靠性和减少幻觉，核心涉及LLM推理过程控制。高度相关的关键词包括：LLMs（核心研究对象）、Chain of Thought/System 2 Thinking（论文关注推理过程）、Hallucination Mitigation（主要目标）。中等相关的关键词：RLHF（作为基线对比提及）、Self-Correction（与过程控制相关）、Mechanistic Interpretability（框架提供结构化解释）。其他关键词如MoE、SLMs、Scaling Laws等与论文内容无直接关联。

大语言模型（LLM）展现出强大的生成能力，但在对抗性提示下仍易产生幻觉和不可靠的推理。现有的安全方法——例如基于人类反馈的强化学习（RLHF）和输出过滤——主要作用于行为层面，可能缺乏确保推理过程完整性的显式架构机制。

本文提出Box Maze框架，这是一种概念性的过程控制架构，它将大语言模型的推理分解为三个显式层次：记忆锚定、结构化推理和边界约束。我们引入了基于模拟的初步评估，该评估涉及多个异构大语言模型系统（DeepSeek-V3、Doubao、Qwen）中的渐进式边界侵蚀场景。在n=50个对抗性场景中的结果表明，显式的认知控制层可以提高边界维护的一致性，在对抗条件下，架构约束能将边界失效率从约40%（基线RLHF）降低至1%以下。

尽管目前的验证是基于模拟的，但这些初步结果表明，过程层面的控制可能为提高大语言模型推理的可靠性提供一个有前景的方向。

Large language models (LLMs) demonstrate strong generative capabilities but remain vulnerable to hallucination and unreliable reasoning under adversarial prompting. Existing safety approaches – such as reinforcement learning from human feedback (RLHF) and output filtering – primarily operate at the behavioral level and may lack explicit architectural mechanisms for enforcing reasoning process integrity. This paper proposes the Box Maze framework, a conceptual process-control architecture that decomposes LLM reasoning into three explicit layers: memory grounding, structured inference, and boundary enforcement. We introduce preliminary simulation-based evaluation involving progressive boundary erosion scenarios across multiple heterogeneous LLM systems (DeepSeek-V3, Doubao, Qwen). Results from n=50 adversarial scenarios suggest that explicit cognitive control layers may improve consistency in boundary maintenance, with architectural constraints reducing boundary failure rates from approximately 40% (baseline RLHF) to below 1% under adversarial conditions. While current validation is simulation-based, these preliminary results indicate that process-level control may offer a promising direction for improving reliability in large language model reasoning.

评分理由: 论文主要研究GPU内核优化的基准测试方法，与大多数大模型技术关键词无直接关系。仅与三个关键词有间接关联：1）‘Large Language Models’（5分）- 论文基准包含语言模型的内核；2）‘LLM Agents’（5分）- 论文提到代理AI系统生成和优化GPU内核；3）‘Speculative Decoding OR Inference Acceleration’（5分）- 论文涉及GPU内核性能优化，与推理加速相关。其他关键词均未涉及。

随着自主人工智能系统在生成和优化GPU内核方面的能力日益增强，其发展却受限于现有基准测试的导向——这些基准主要奖励相对于软件基线的加速比，而非追求接近硬件效率极限的执行性能。本文提出SOL-ExecBench基准，该基准包含从124个生产级及新兴人工智能模型中提取的235个CUDA内核优化问题，涵盖语言、扩散、视觉、音频、视频及混合架构领域，并针对NVIDIA Blackwell GPU设计。基准覆盖BF16、FP8和NVFP4数据格式下的前向与反向计算负载，包含那些预期最佳性能需依赖Blackwell架构特有功能的内核。与以往主要基于软件实现评估内核的基准不同，SOL-ExecBench通过我们开发的SOLAR流程（一种基于硬件特性推导光速极限的理论边界计算框架）获得分析性光速极限边界，并以此为固定目标衡量硬件效率优化水平。我们提出SOL分数，用于量化候选内核在评分基线（由发布定义）与硬件光速极限边界之间所缩小的性能差距。为支持对自主优化系统的稳健评估，我们还提供了沙盒化测试框架，具备GPU时钟锁定、L2缓存清理、隔离子进程执行及基于静态分析的抗奖励篡改策略检测功能。SOL-ExecBench将GPU内核基准测试的范式从超越可变软件基线，转变为衡量向硬件光速极限逼近的剩余差距。

As agentic AI systems become increasingly capable of generating and optimizing GPU kernels, progress is constrained by benchmarks that reward speedup over software baselines rather than proximity to hardware-efficient execution. We present SOL-ExecBench, a benchmark of 235 CUDA kernel optimization problems extracted from 124 production and emerging AI models spanning language, diffusion, vision, audio, video, and hybrid architectures, targeting NVIDIA Blackwell GPUs. The benchmark covers forward and backward workloads across BF16, FP8, and NVFP4, including kernels whose best performance is expected to rely on Blackwell-specific capabilities. Unlike prior benchmarks that evaluate kernels primarily relative to software implementations, SOL-ExecBench measures performance against analytically derived Speed-of-Light (SOL) bounds computed by SOLAR, our pipeline for deriving hardware-grounded SOL bounds, yielding a fixed target for hardware-efficient optimization. We report a SOL Score that quantifies how much of the gap between a release-defined scoring baseline and the hardware SOL bound a candidate kernel closes. To support robust evaluation of agentic optimizers, we additionally provide a sandboxed harness with GPU clock locking, L2 cache clearing, isolated subprocess execution, and static analysis based checks against common reward-hacking strategies. SOL-ExecBench reframes GPU kernel benchmarking from beating a mutable software baseline to closing the remaining gap to hardware Speed-of-Light.