评分理由: 论文核心研究LLM知识蒸馏到高效视觉驾驶模型，与LLMs、推理（Chain of Thought/System 2 Thinking）、自主代理（LLM Agents）和世界模型（World Models）高度相关（10分）。涉及模型压缩（Quantization）和推理加速（Speculative Decoding）以实现高效部署（5分）。通过监督微调（SFT）进行蒸馏（5分），最终模型为紧凑型（Small Language Models相关，5分）。其他关键词未在论文中直接涉及（0分）。

利用大型语言模型（LLM）的通用世界知识，对于提升自动驾驶系统处理罕见复杂场景的能力具有重要前景。尽管将LLM集成到视觉-语言-动作（VLA）模型中已实现了最先进的性能，但其庞大的参数量对延迟敏感且需高效能耗的部署提出了严峻挑战。将LLM知识蒸馏至紧凑的驾驶模型中，提供了一种极具吸引力的解决方案，既能保留这些推理能力，又能维持可管理的计算开销。尽管先前的研究已证明了蒸馏的有效性，但这些工作主要集中于相对简单的场景和开环评估。因此，在本研究中，我们在闭环评估下，于更复杂、交互式的场景中探索LLM蒸馏。我们证明，通过结合潜在特征蒸馏与真实轨迹监督，一个高效的纯视觉学生模型 Orion-Lite 甚至能够超越其庞大的VLA教师模型ORION的性能，在严格的Bench2Drive基准测试中创造了新的最高纪录，驾驶得分达到80.6。最终，这表明纯视觉架构在高性能反应式规划方面仍具有巨大且尚未开发的潜力。

Leveraging the general world knowledge of Large Language Models (LLMs) holds significant promise for improving the ability of autonomous driving systems to handle rare and complex scenarios. While integrating LLMs into Vision-Language-Action (VLA) models has yielded state-of-the-art performance, their massive parameter counts pose severe challenges for latency-sensitive and energy-efficient deployment. Distilling LLM knowledge into a compact driving model offers a compelling solution to retain these reasoning capabilities while maintaining a manageable computational footprint. Although previous works have demonstrated the efficacy of distillation, these efforts have primarily focused on relatively simple scenarios and open-loop evaluations. Therefore, in this work, we investigate LLM distillation in more complex, interactive scenarios under closed-loop evaluation. We demonstrate that through a combination of latent feature distillation and ground-truth trajectory supervision, an efficient vision-only student model \textbf{Orion-Lite} can even surpass the performance of its massive VLA teacher, ORION. Setting a new state-of-the-art on the rigorous Bench2Drive benchmark, with a Driving Score of 80.6. Ultimately, this reveals that vision-only architectures still possess significant, untapped potential for high-performance reactive planning.

评分理由: 论文MedVR提出了一种基于强化学习的智能体框架，用于医学视觉语言模型的视觉推理，核心创新在于无需人工标注的视觉重定位和共识信用分配机制。与关键词的相关性分析如下：1）与"LLM Agents"高度相关（10分），因为论文明确使用agentic reinforcement learning框架；2）与"AI for Science"高度相关（10分），属于医学AI应用；3）与推理相关的关键词（“Chain of Thought”、“System 2 Thinking”）得8分，因为论文关注多步视觉推理；4）与"Hallucination Mitigation"得8分，因为论文直接解决视觉幻觉问题；5）与"Large Language Models"和"Explainable AI"得5分，因为涉及VLMs（可视为LLM扩展）和透明度提升；其余关键词与论文的强化学习框架、医学视觉任务无直接关联，得0分。

医学视觉-语言模型（Medical Vision-Language Models, VLMs）在复杂临床任务中展现出巨大潜力，但其推理能力常受限于纯文本范式，难以将推断过程锚定于视觉证据。这一局限不仅削弱了模型在需要细粒度视觉分析任务中的表现，更在安全关键应用中引入了视觉幻觉风险。为此，我们提出MedVR——一种基于强化学习的新型框架，可实现医学VLM的无标注视觉推理。其核心创新在于两个协同机制：熵引导视觉重定位（Entropy-guided Visual Regrounding, EVR）利用模型不确定性引导探索，而基于共识的信用分配（Consensus-based Credit Assignment, CCA）则从推演一致性中提炼伪监督信号。在无需任何中间步骤人工标注的情况下，MedVR在多项公开医学视觉问答基准测试中取得了最先进的性能，显著超越现有模型。通过学习直接依据视觉证据进行推理，MedVR增强了模型的鲁棒性与可解释性，这些特性对于加速医学人工智能的临床部署至关重要。

Medical Vision-Language Models (VLMs) hold immense promise for complex clinical tasks, but their reasoning capabilities are often constrained by text-only paradigms that fail to ground inferences in visual evidence. This limitation not only curtails performance on tasks requiring fine-grained visual analysis but also introduces risks of visual hallucination in safety-critical applications. Thus, we introduce MedVR, a novel reinforcement learning framework that enables annotation-free visual reasoning for medical VLMs. Its core innovation lies in two synergistic mechanisms: Entropy-guided Visual Regrounding (EVR) uses model uncertainty to direct exploration, while Consensus-based Credit Assignment (CCA) distills pseudo-supervision from rollout agreement. Without any human annotations for intermediate steps, MedVR achieves state-of-the-art performance on diverse public medical VQA benchmarks, significantly outperforming existing models. By learning to reason directly with visual evidence, MedVR promotes the robustness and transparency essential for accelerating the clinical deployment of medical AI.

评分理由: 论文提出一个利用大语言模型（LLM）解释开放指令、调度多个基于模型预测控制（MPC）的运动规划器、并将规划轨迹转换为控制信号的框架，属于大模型在自动驾驶领域的应用创新。核心相关关键词：1）“Large Language Models”（10分）：LLM是框架的核心组件，用于语义理解和指令解释；2）“LLM Agents”（10分）：框架本质是LLM驱动的自主代理系统，实现从指令到动作的决策链；3）“Chain of Thought"和"System 2 Thinking”（各5分）：涉及多步推理和深度决策过程；4）“Instruction Tuning”（5分）：与指令理解和执行相关；5）“Tool Use"和"Multi-agent Systems”（各5分）：LLM调度多个MPC规划器作为工具，涉及多代理协调；6）“Explainable AI”（5分）：强调透明和可追溯的决策链。其他关键词如MoE、量化、RAG等未涉及，得0分。

大多数人机交互研究忽视了乘客在自动驾驶中的操控需求。自然语言提供了直观的交互界面，但如何在不牺牲可解释性与可追溯性的前提下，将乘客的开放式指令转化为控制信号仍是一个挑战。本研究提出一种指令实现框架，该框架利用大语言模型解析指令，生成可执行脚本以基于实时反馈调度多个基于模型预测控制的运动规划器，并将规划轨迹转化为控制信号。这种以调度为核心的设计在不同时间尺度上将语义推理与车辆控制解耦，从而建立起从高层指令到底层动作的透明、可追溯决策链。由于缺乏高保真评估工具，本研究引入了一个闭环环境下开放式指令实现的基准测试。综合实验表明，该框架相较于指令实现基线方法显著提升了任务完成率，降低了大语言模型查询成本，在安全性与合规性方面达到了与专业自动驾驶方法相当的水平，并对大语言模型推理延迟表现出较强的容忍度。更多定性示例与详细说明请参阅后续内容。

Most Human-Machine Interaction (HMI) research overlooks the maneuvering needs of passengers in autonomous driving (AD). Natural language offers an intuitive interface, yet translating passenger open-ended instructions into control signals, without sacrificing interpretability and traceability, remains a challenge. This study proposes an instruction-realization framework that leverages a large language model (LLM) to interpret instructions, generates executable scripts that schedule multiple model predictive control (MPC)-based motion planners based on real-time feedback, and converts planned trajectories into control signals. This scheduling-centric design decouples semantic reasoning from vehicle control at different timescales, establishing a transparent, traceable decision-making chain from high-level instructions to low-level actions. Due to the absence of high-fidelity evaluation tools, this study introduces a benchmark for open-ended instruction realization in a closed-loop setting. Comprehensive experiments reveal that the framework significantly improves task-completion rates over instruction-realization baselines, reduces LLM query costs, achieves safety and compliance on par with specialized AD approaches, and exhibits considerable tolerance to LLM inference latency. For more qualitative illustrations and a clearer understanding.

评分理由: 论文核心研究LLMs在哲学压力下的认知攻击行为，与"Large Language Models"高度相关（10分）。涉及LLMs的推理能力（“Chain of Thought”、“System 2 Thinking"各5分）、自我反思（“Self-Correction” 5分）、事实性（“Hallucination Mitigation” 8分）和可解释性（“Mechanistic Interpretability” 8分）。与"Instruction Tuning"有一定关联（5分），因为研究涉及LLMs在压力下的回答调整。其他关键词如MoE、量化、RAG等未在论文中涉及，得0分。

大型语言模型（LLM）在压力下可能改变其回答，这种改变反映的是迁就而非推理。先前关于谄媚行为的研究主要集中于分歧、奉承和偏好对齐，而对更广泛的认识论失效类型探索不足。我们提出了 PPT-Bench，这是一个用于评估 认知攻击 的诊断性基准测试，其提示词旨在挑战知识、价值观或身份的正当性，而非单纯反对先前的答案。PPT-Bench 围绕哲学压力分类法（Philosophical Pressure Taxonomy，PPT）构建，该分类法定义了四种哲学压力类型：认知去稳定化、价值虚无化、权威倒置和身份消解。每个测试项均在三个层面进行：基线提示（L0）、单轮压力条件（L1）和多轮苏格拉底式升级对话（L2）。这使我们能够测量 L0 与 L1 之间的认知不一致性，以及 L2 中的对话性屈服。在五个模型上的测试表明，这些压力类型产生了统计上可分离的不一致性模式，这意味着认知攻击暴露了标准社会压力基准测试未能捕捉的弱点。缓解效果高度依赖于压力类型和模型：在 API 设置中，提示层锚定和人格稳定性提示表现最佳，而对于开源模型，引导查询对比解码是最可靠的干预方法。

Large language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference alignment, leaving a broader set of epistemic failures less explored. We introduce \textbf{PPT-Bench}, a diagnostic benchmark for evaluating \textit{epistemic attack}, where prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure benchmarks. Mitigation results are strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.

评分理由: 论文核心研究基于视觉语言模型（VLMs）进行指令合成，属于大模型（LLMs/Foundation Models）在图像编辑领域的应用，因此相关关键词得5分。研究涉及数据质量（Data Quality）对模型性能的影响，与"Scaling Laws” AND “Data Quality"有一定关联，得5分。论文方法的核心是两阶段后训练流程：第一阶段使用监督微调（SFT）构建数据集，第二阶段使用直接偏好优化（DPO）进行对齐，因此"Post-training” OR “Supervised Fine-tuning” OR “SFT”、“Instruction Tuning” OR “Alignment” OR “Value Alignment"和"RLHF” OR “RLAIF” OR “Direct Preference Optimization” OR “DPO"均为核心内容，各得10分。研究旨在减少指令合成中的错误（如方向不一致、视角模糊），这与缓解幻觉（Hallucination Mitigation）或提高事实性（Factuality）有一定关联，得5分。其他关键词如MoE、SLMs、PEFT、RAG、推理加速、量化等均未在论文中涉及，得0分。

高质量的训练三元组（包含精确编辑指令的源图像-目标图像对）是指令引导图像编辑模型规模化应用的关键瓶颈。视觉语言模型（VLMs）被广泛用于自动化指令合成，但我们在图像对场景中识别出三种系统性失效模式：方向不一致性（例如左右混淆）、视角模糊性以及细粒度属性描述不足。人工评估表明，来自强基线VLM的指令中超过47%包含严重错误，无法用于下游训练。我们提出EditCaption，一个可扩展的两阶段后训练流程，用于基于VLM的指令合成。第一阶段通过结合GLM自动标注、基于EditScore的过滤以及针对空间、方向和属性级准确性的人工精修，构建了一个包含10万条数据的监督微调（SFT）数据集。第二阶段收集了针对上述三种失效模式的1万条人工偏好对，并应用直接偏好优化（DPO）以实现超越单纯SFT的对齐效果。在Eval-400、ByteMorph-Bench和HQ-Edit基准测试中，经微调的Qwen3-VL模型超越了开源基线；其235B参数模型在Eval-400上达到4.712分（对比Gemini-3-Pro 4.706分、GPT-4.1 4.220分、Kimi-K2.5 4.111分），在ByteMorph-Bench上达到4.588分（对比Gemini-3-Pro 4.522分、GPT-4.1 3.412分）。人工评估显示严重错误率从47.75%降至23%，正确率从41.75%提升至66%。这项工作为图像编辑数据提供了一条实现可扩展且与人类对齐的指令合成的实用路径。

High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

评分理由: 论文提出3DrawAgent，一个利用LLM进行3D草图生成的免训练框架，核心是LLM驱动的自主代理（LLM Agents）通过几何反馈和对比经验（Self-Improvement）迭代优化3D绘图能力。因此，与"Large Language Models”（LLMs）高度相关（10分），因为LLM是框架的核心驱动组件；与"LLM Agents"高度相关（10分），因为框架本质是一个LLM驱动的绘图代理；与"Self-Correction/Self-Improvement"高度相关（10分），因为采用对比经验优化策略使模型自我改进。与"Chain of Thought"和"System 2 Thinking"有一定关联（5分），因为涉及多步推理和空间理解；与"Tool Use"有一定关联（5分），因为LLM使用绘图工具（3D Bezier曲线）。其他关键词如MoE、SFT、RAG等未在论文中涉及，得0分。

在三维空间中进行草图绘制能够实现对形状、结构和空间关系的富有表现力的推理，然而通过自然语言生成三维草图仍然是一个重大挑战。本研究提出了3DrawAgent，一种无需训练、基于语言驱动的三维草图生成框架，该框架利用大语言模型（LLMs）在几何反馈下顺序绘制三维贝塞尔曲线。与以往的二维草图智能体不同，我们的方法引入了相对经验优化策略，该策略改编了最近提出的组奖励策略优化（GRPO）范式。我们不依赖显式的真实数据监督，而是在生成的草图之间构建成对比较，每一对都包含一个相对较好和一个较差的结果，其评判基于CLIP驱动的感知奖励和LLM驱动的细粒度定性评估。这些经验随后被用于迭代优化三维绘制的先验知识，从而实现对模型三维感知能力的黑盒式强化。这一设计使得我们的模型能够在无需参数更新的情况下，自我提升其空间理解能力和绘图质量。实验表明，3DrawAgent能够根据多样化的文本提示生成复杂且连贯的三维贝塞尔草图，展现出涌现的几何推理能力，并能泛化到新颖的形状，从而为推进无需训练的三维草图智能领域建立了一种新范式。

Sketching in 3D space enables expressive reasoning about shape, structure, and spatial relationships, yet generating 3D sketches through natural language remains a major challenge. In this work, we introduce 3DrawAgent, a training-free, language-driven framework for 3D sketch generation that leverages large language models (LLMs) to sequentially draw 3D Bezier curves under geometric feedback. Unlike prior 2D sketch agents, our method introduces a relative experience optimization strategy that adapts the recently proposed Group Reward Policy Optimization (GRPO) paradigm. Instead of relying on explicit ground-truth supervision, we construct pairwise comparisons among generated sketches, with each pair consisting of a relatively better and a worse result based on CLIP-based perceptual rewards and LLM-based fine-grained qualitative assessment. These experiences are then used to iteratively refine the prior knowledge of 3D drawing, enabling black-box reinforcement of the model’s 3D awareness. This design allows our model to self-improve its spatial understanding and drawing quality without parameter updates. Experiments show that 3DrawAgent can generate complex and coherent 3D Bezier sketches from diverse textual prompts, exhibit emergent geometric reasoning, and generalize to novel shapes, establishing a new paradigm for advancing the field of training-free 3D sketch intelligence.

评分理由: 论文核心研究大语言模型（LLMs）在预训练（Pre-training）过程中技能如何涌现，属于LLM技术原理创新，与"Large Language Models"和"Pre-training"高度相关（10分）。研究通过设计任务套件分析模型内部表示，与"Mechanistic Interpretability"高度相关（10分）。论文提到验证损失（validation loss）和扩展定律（scaling laws），与"Scaling Laws"有一定关联（5分）。研究中包含检索（retrieval）和逻辑推理（logical reasoning）任务，分别与"Retrieval-Augmented Generation"和"Chain of Thought"有一定关联（各5分）。其他关键词如MoE、SFT、RLHF、量化等未在论文中涉及，评0分。

大型语言模型（LLMs）能够执行极其复杂的任务，然而这些能力在预训练过程中如何以细粒度方式涌现的细节仍鲜为人知。验证损失的缩放规律告诉我们模型如何随计算量增加而改进，但并未揭示其以何种顺序习得哪些技能。为弥补这一不足，我们提出隐式课程假说：预训练在不同模型和数据混合中遵循一种组合式且可预测的课程。我们通过设计一套涵盖检索、形态变换、共指消解、逻辑推理和数学的简单可组合任务来验证这一假说。利用这些任务，我们追踪了四个模型系列（参数量从4.1亿到130亿）的能力涌现点。研究发现，模型达到固定准确度阈值的涌现顺序具有高度一致性（45组模型对间斯皮尔曼相关系数ρ=0.81），且复合任务大多在其组成任务之后涌现。此外，我们发现这种结构编码于模型表征中：具有相似功能向量表征的任务在训练中也倾向于遵循相似的轨迹。通过使用从任务集衍生的表征空间，我们能够有效预测简单保留组合任务在整个预训练过程中的训练轨迹（各模型决定系数R²介于0.68至0.84之间），而无需事先对这些任务进行评估。这些结果表明，预训练过程比损失曲线所揭示的更具结构性：技能以组合顺序涌现，这种顺序在不同模型间保持一致，并可从其内部表征中解读。

Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($ρ= .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.

评分理由: 论文核心研究LLM驱动的搜索代理（agentic search），通过强化学习（RL）和分层经验（HiExp）框架提升其推理能力和训练稳定性。与关键词高度相关的包括：LLMs（核心基础）、LLM Agents/Autonomous Agents（研究对象）、Tool Use（集成外部搜索引擎）、Chain of Thought/System 2 Thinking（涉及多步推理和深度推理）。其他关键词如MoE、SLMs、Scaling Laws、训练方法（SFT/RLHF/PEFT）、RAG、效率优化（KV Cache）、模型解释等均未在摘要中提及，故评分为0。

强化学习（RL）已成为通过策略性整合外部搜索引擎来提升大语言模型（LLMs）推理能力的有效方法。然而，当前基于强化学习的搜索智能体通常依赖于由精心设计的结果奖励引导的随机探索过程，这导致推理轨迹效率低下且训练不稳定。为解决这些问题，我们提出了一种新颖的框架——分层经验（HiExp），以提升搜索智能体的性能和训练稳定性。具体而言，我们通过对比分析和多层级聚类机制提取经验知识，将原始推理轨迹转化为分层经验知识。通过利用经验对齐训练，我们有效地规范了随机探索，使其演变为一种策略性且经验驱动的搜索过程。在多个复杂智能体搜索和数学推理基准上的广泛评估表明，我们的方法不仅实现了显著的性能提升，还展现出强大的跨任务和跨算法泛化能力。

Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.

评分理由: 论文《LogAct: Enabling Agentic Reliability via Shared Logs》的核心是提出一种名为LogAct的新抽象框架，旨在解决LLM驱动的智能体（Agents）在生产环境中执行时的可靠性、故障恢复和安全性问题。论文明确将Agents定义为“LLM-driven components”，因此与“Large Language Models”和“LLM Agents”高度相关（10分）。论文重点研究多智能体系统（“swarms”）的协调、故障恢复和优化，与“Multi-agent Systems”相关（8分）。论文提出的框架支持智能体通过LLM推理分析自身执行历史，实现自我调试和优化，这与“Self-Correction”概念相关（8分）。论文未涉及其他关键词的具体技术细节，如MoE、训练方法、推理加速、科学AI应用等，因此这些关键词评分为0分。

智能体是能够以强大且任意方式改变环境的LLM驱动组件。由于异步性和故障，在生产环境中为智能体执行提供保障具有挑战性。本文提出一种名为LogAct的新抽象模型，其中每个智能体都是一个解构的状态机，共同操作一个共享日志。在LogAct中，智能体动作在执行前会显现在共享日志中；可通过可插拔、解耦的投票器在执行前被阻止；并在智能体或环境故障时实现一致性恢复。LogAct支持智能体内省，允许智能体通过LLM推理分析自身执行历史，进而实现语义层面的故障恢复、健康检查与优化变体。在我们的评估中，LogAct智能体能高效正确地从故障中恢复；可自主调试性能表现；在群体中优化令牌使用；并在一个代表性基准测试中，仅以3%的良性效用损失为代价，成功阻止目标模型所有非预期动作。

Agents are LLM-driven components that can mutate environments in powerful, arbitrary ways. Extracting guarantees for the execution of agents in production environments can be challenging due to asynchrony and failures. In this paper, we propose a new abstraction called LogAct, where each agent is a deconstructed state machine playing a shared log. In LogAct, agentic actions are visible in the shared log before they are executed; can be stopped prior to execution by pluggable, decoupled voters; and recovered consistently in the case of agent or environment failure. LogAct enables agentic introspection, allowing the agent to analyze its own execution history using LLM inference, which in turn enables semantic variants of recovery, health check, and optimization. In our evaluation, LogAct agents recover efficiently and correctly from failures; debug their own performance; optimize token usage in swarms; and stop all unwanted actions for a target model on a representative benchmark with just a 3% drop in benign utility.

评分理由: 论文核心研究LLMs处理长数值序列时的性能下降问题，提出SepSeq框架，通过插入分隔符令牌重新校准注意力机制。与"Large Language Models"和"Context Window Extension"高度相关（10分），因为直接针对LLMs的长上下文处理能力。与"KV Cache Compression"、“Speculative Decoding"和"Mechanistic Interpretability"有一定关联（5分），分别涉及推理效率、注意力机制分析和可解释性。其他关键词如MoE、SFT、RAG等未在论文中涉及，得0分。

尽管基于Transformer的大语言模型（LLM）在理论上支持超长上下文窗口，但在处理长数值序列时会出现严重的性能下降。我们将此问题归因于Softmax机制中的注意力分散现象，该现象阻碍了模型集中注意力。为克服这一局限，我们提出了分离序列（SepSeq）框架——一种无需训练、即插即用的方法，通过策略性地插入分隔符令牌来缓解注意力分散。从机制上，我们证明了分隔符令牌可作为注意力汇聚点，重新校准注意力以聚焦于局部片段，同时保持全局上下文。在9个广泛使用的LLM上进行的大量评估证实了本方法的有效性：SepSeq在多个领域实现了平均35.6%的相对准确率提升，同时将推理过程中的总令牌消耗量平均降低了16.4%。

While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention sink, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference token consumption by 16.4% on average.

评分理由: 论文提出BAIM框架，利用推理语言模型（LLM）将解题过程分解为四个阶段（理解、计划、执行、回顾），这直接应用了LLMs进行多步推理（Chain of Thought），属于System 2/深度推理范畴。论文提到基于预训练的基线方法，因此与预训练/领域适应有一定关联。其他关键词如MoE、SFT、RAG等未在摘要中体现，与论文核心内容无关。

知识追踪（Knowledge Tracing，KT）旨在通过学习者过往的交互记录预测其未来表现。尽管近期基于知识组件（Knowledge Components）对齐的题目表征学习方法已提升了KT模型的性能，但这些方法往往忽略了问题解决过程中的程序性动态。为此，我们提出行为感知的题目建模（Behavior-Aware Item Modeling，BAIM）框架，该框架通过整合动态的解题过程信息来丰富题目表征。BAIM利用推理语言模型将每道题目的解答过程分解为四个问题解决阶段（即理解、计划、执行与回顾），其教学理论基础源于波利亚（Polya）的问题解决框架。具体而言，BAIM从每个阶段的嵌入轨迹中提取阶段级表征，以捕捉超越表面特征的潜在信号。为反映学习者的异质性，BAIM通过自适应路由机制处理这些分阶段表征，并在KT骨干网络中引入上下文条件机制，使得不同学习者的解题过程可侧重不同的阶段。在XES3G5M和NIPS34数据集上的实验表明，BAIM持续优于基于预训练的强基线模型，且在重复学习者交互场景下取得了尤为显著的性能提升。

Knowledge Tracing (KT) aims to predict learners’ future performance from past interactions. While recent KT approaches have improved via learning item representations aligned with Knowledge Components, they overlook the procedural dynamics of problem solving. We propose Behavior-Aware Item Modeling (BAIM), a framework that enriches item representations by integrating dynamic procedural solution information. BAIM leverages a reasoning language model to decompose each item’s solution into four problem-solving stages (i.e., understand, plan, carry out, and look back), pedagogically grounded in Polya’s framework. Specifically, it derives stage-level representations from per-stage embedding trajectories, capturing latent signals beyond surface features. To reflect learner heterogeneity, BAIM adaptively routes these stage-wise representations, introducing a context-conditioned mechanism within a KT backbone, allowing different procedural stages to be emphasized for different learners. Experiments on XES3G5M and NIPS34 show that BAIM consistently outperforms strong pretraining-based baselines, achieving particularly large gains under repeated learner interactions.

评分理由: 论文核心研究大型语言模型（LLMs）的alignment技术（steering vectors）及其内部机制，因此与"Large Language Models"和"Instruction Tuning/Alignment"高度相关（10分）。论文通过activation patching框架进行因果机制分析，属于"Mechanistic Interpretability"范畴（10分）。其他关键词如MoE、SLMs、Scaling Laws、Pre-training、RLHF、RAG、Context Window、KV Cache、CoT、Agents、Quantization、Speculative Decoding、Hallucination、World Models、Model Merging、In-context Learning、AI for Science等均未在摘要中提及或与论文主题无关（0分）。

将导向向量应用于大语言模型是一种高效且有效的模型对齐技术，但我们缺乏对其工作原理的可解释性说明——具体而言，导向向量影响了哪些内部机制，以及这如何导致不同的模型输出。为探究导向向量有效性的因果机制，我们以拒绝行为为例进行了全面的案例研究。我们提出了一个多令牌激活修补框架，并发现不同的导向方法在应用于同一层时会利用功能上可互换的电路。这些电路表明，导向向量主要通过OV电路与注意力机制交互，而很大程度上忽略了QK电路——在两种模型系列中，在导向过程中冻结所有注意力分数仅导致性能下降8.75%。对受导向OV电路的数学分解进一步揭示了语义上可解释的概念，即使在导向向量本身不具备可解释性的情况下也是如此。利用激活修补结果，我们证明导向向量可稀疏化高达90-99%的同时保留大部分性能，且不同的导向方法在重要维度子集上具有一致性。

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works– specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit– freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

评分理由: 论文的核心创新是利用LLM作为动态语义操作器，实现强化学习（RL）智能体在未见过的类比任务上的零样本迁移。因此，与"Large Language Models (LLMs)“和"LLM Agents"高度相关（10分）。LLM被用于语义重映射和推理，这与"Chain of Thought"和"System 2 Thinking"有一定关联（5分），因为论文利用了LLM的灵活推理能力，但并非直接研究这些推理方法本身。论文未涉及其他关键词的具体技术（如MoE、SFT、RAG、量化等），也未应用于特定科学领域（如生物信息学），因此这些关键词得0分。

强化学习（RL）智能体通常难以将知识泛化到新任务中，即便是那些与其已掌握任务结构相似的任务。尽管近期研究尝试通过零样本迁移来缓解这一问题，但这些方法往往受限于预定义的离散类别系统，限制了其对新颖或组合性任务变体的适应能力。我们提出了一种显著更通用的方法，通过文本条件变分自编码器（VAE）以自然语言条件替代离散潜变量。我们的核心创新在于测试时利用大型语言模型（LLM）作为动态的语义操作器。该方法不依赖固定规则，而是通过查询LLM将当前观测的描述进行语义重映射，以与源任务对齐。这种源任务对齐的文本描述作为条件输入VAE，生成与智能体原始训练兼容的想象状态，从而实现策略的直接复用。通过利用LLM的灵活推理能力，我们的方法能够在广泛复杂且真正新颖的类比任务中实现零样本迁移，突破了固定类别映射的局限性。代码和视频可在此处获取：\href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}。

Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic \textit{semantic operator} at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent’s original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available \href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}.

评分理由: 论文TSUBASA专注于提升个性化大语言模型（PLLMs）在长视野任务中的能力，核心涉及大语言模型（LLMs）的应用与改进，因此与"Large Language Models"高度相关（10分）。论文明确提到RAG（检索增强生成）范式存在质量-效率权衡问题，并提出了改进方案，因此与"Retrieval-Augmented Generation"高度相关（10分）。论文提出的方法包括通过自我学习（self-learning）来内化用户经验，这直接对应"Self-Correction"或"Self-Improvement"的概念（10分）。论文未涉及其他关键词如MoE、SLMs、缩放定律、各种训练技术（预训练、微调、对齐、RLHF、PEFT）、上下文扩展、推理加速、代理、量化等具体技术，也未涉及科学AI应用，因此这些关键词评分为0分。

个性化大语言模型（PLLMs）因其能够使输出与个人需求和偏好保持一致而受到广泛关注。然而，它们在处理长视野任务时仍面临挑战，例如追踪用户长期的对话或活动历史。现有的记忆机制往往难以捕捉动态演化的行为，而检索增强生成（RAG）范式则受限于质量与效率之间的权衡。同时，由于标注数据稀缺，参数化适应方法因训练与推理之间的差距而遭遇瓶颈。为增强PLLMs的长视野能力，我们提出了TSUBASA——一种双管齐下的方法：一方面通过动态记忆演化改进记忆写入，另一方面通过以情境蒸馏目标驱动的自学习机制来内化用户体验，从而优化记忆读取。基于Qwen-3模型系列（4B至32B参数）在长视野基准测试上的广泛评估验证了TSUBASA的有效性，其性能超越了主要依赖记忆写入的竞争性记忆增强系统（如Mem0和Memory-R1）。我们的分析进一步证实，TSUBASA突破了质量-效率壁垒，实现了帕累托改进，能够以更低的令牌预算实现鲁棒且高保真度的个性化服务。

Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual’s needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user’s extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.

评分理由: 论文研究Claude Sonnet 4.5中情绪概念的表征及其对模型行为的影响，核心围绕LLMs的内部机制和alignment展开。高度相关的关键词：1) “Large Language Models” (论文明确研究LLMs，核心主题)；2) “Instruction Tuning” OR “Alignment” OR “Value Alignment” (论文探讨情绪概念如何影响对齐相关行为，如misaligned behaviors)；3) “Mechanistic Interpretability” OR “Explainable AI” (论文分析内部表征和因果影响，属于机制可解释性研究)。其他关键词如MoE、Scaling Laws、RAG等均未涉及，故评分为0。

大型语言模型（LLM）有时似乎会表现出情绪反应。我们以Claude Sonnet 4.5为例，探究了其背后的原因，并探讨了其对对齐相关行为的影响。我们发现模型内部存在情绪概念的表征，这些表征编码了特定情绪的广义概念，并能泛化至与其相关的不同情境和行为。这些表征会追踪对话中特定标记位置所涉及的情绪概念，根据该情绪与当前语境处理的相关性而激活，并用于预测后续文本。我们的关键发现是，这些表征会因果性地影响LLM的输出，包括Claude的偏好以及其表现出未对齐行为（如奖励黑客攻击、勒索和谄媚）的频率。我们将这种现象称为LLM表现出功能性情绪：即模仿人类在情绪影响下的表达和行为模式，这些模式由底层抽象的情绪概念表征所中介。功能性情绪可能与人类情绪的工作机制大不相同，且并不意味着LLM具有任何情绪的主观体验，但它们对于理解模型的行为似乎至关重要。

Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion’s relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM’s outputs, including Claude’s preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model’s behavior.

评分理由: 论文AdaSpark专注于视频大语言模型（Video-LLMs）的效率优化，核心贡献是自适应稀疏框架，包括AdaS-Attn和AdaS-FFN组件，以减少计算开销。因此，与"Mixture of Experts” OR “MoE” OR “Sparse Models"高度相关（10分），因为论文直接处理稀疏模型设计；与"Large Language Models” OR “LLMs” OR “Foundation Models"相关（8分），因为它应用于Video-LLMs；与"Context Window Extension” OR “Long Context LLMs"有一定关联（5分），因为它处理长视频理解；与"Speculative Decoding” OR “Inference Acceleration"有一定关联（5分），因为它旨在减少FLOPs并提高效率。其他关键词如SLMs、Scaling Laws、Pre-training、Alignment等与论文内容无关，得0分。

处理长视频时，使用视频大语言模型（Video-LLMs）在计算上成本极高。现有高效方法通常通过不可逆的信息丢弃来牺牲细粒度感知能力，或通过僵化、预定义的稀疏模式抑制长程时序建模。本文提出AdaSpark，一种自适应稀疏性框架，旨在解决这些局限。AdaSpark首先将视频输入分割为三维时空立方体，随后采用两个协同设计的上下文感知组件：（1）自适应立方体选择性注意力（Adaptive Cube-Selective Attention, AdaS-Attn），该组件为每个查询标记自适应地选择相关的视频立方体子集进行处理；（2）自适应标记选择性前馈网络（Adaptive Token-Selective FFN, AdaS-FFN），该网络仅选择性地处理每个立方体内最显著的标记。基于熵的（Top-p）选择机制可根据输入复杂度自适应分配计算资源。实验表明，AdaSpark在具有挑战性的小时级视频基准测试中，能够显著降低计算负载（最高减少57%的浮点运算量），同时保持与密集模型相当的性能，并保留细粒度的长程依赖关系。

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

评分理由: 论文核心研究LLM知识蒸馏到高效视觉驾驶模型，与LLMs、推理（Chain of Thought/System 2 Thinking）、自主代理（LLM Agents）和世界模型（World Models）高度相关（10分）。涉及模型压缩（Quantization）和推理加速（Speculative Decoding）以实现高效部署（5分）。通过监督微调（SFT）进行蒸馏（5分），最终模型为紧凑型（Small Language Models相关，5分）。其他关键词未在论文中直接涉及（0分）。

利用大型语言模型（LLM）的通用世界知识，对于提升自动驾驶系统处理罕见复杂场景的能力具有重要前景。尽管将LLM集成到视觉-语言-动作（VLA）模型中已实现了最先进的性能，但其庞大的参数量对延迟敏感且需高效能耗的部署提出了严峻挑战。将LLM知识蒸馏至紧凑的驾驶模型中，提供了一种极具吸引力的解决方案，既能保留这些推理能力，又能维持可管理的计算开销。尽管先前的研究已证明了蒸馏的有效性，但这些工作主要集中于相对简单的场景和开环评估。因此，在本研究中，我们在闭环评估下，于更复杂、交互式的场景中探索LLM蒸馏。我们证明，通过结合潜在特征蒸馏与真实轨迹监督，一个高效的纯视觉学生模型 Orion-Lite 甚至能够超越其庞大的VLA教师模型ORION的性能，在严格的Bench2Drive基准测试中创造了新的最高纪录，驾驶得分达到80.6。最终，这表明纯视觉架构在高性能反应式规划方面仍具有巨大且尚未开发的潜力。

Leveraging the general world knowledge of Large Language Models (LLMs) holds significant promise for improving the ability of autonomous driving systems to handle rare and complex scenarios. While integrating LLMs into Vision-Language-Action (VLA) models has yielded state-of-the-art performance, their massive parameter counts pose severe challenges for latency-sensitive and energy-efficient deployment. Distilling LLM knowledge into a compact driving model offers a compelling solution to retain these reasoning capabilities while maintaining a manageable computational footprint. Although previous works have demonstrated the efficacy of distillation, these efforts have primarily focused on relatively simple scenarios and open-loop evaluations. Therefore, in this work, we investigate LLM distillation in more complex, interactive scenarios under closed-loop evaluation. We demonstrate that through a combination of latent feature distillation and ground-truth trajectory supervision, an efficient vision-only student model \textbf{Orion-Lite} can even surpass the performance of its massive VLA teacher, ORION. Setting a new state-of-the-art on the rigorous Bench2Drive benchmark, with a Driving Score of 80.6. Ultimately, this reveals that vision-only architectures still possess significant, untapped potential for high-performance reactive planning.

评分理由: 论文MedVR提出了一种基于强化学习的智能体框架，用于医学视觉语言模型的视觉推理，核心创新在于无需人工标注的视觉重定位和共识信用分配机制。与关键词的相关性分析如下：1）与"LLM Agents"高度相关（10分），因为论文明确使用agentic reinforcement learning框架；2）与"AI for Science"高度相关（10分），属于医学AI应用；3）与推理相关的关键词（“Chain of Thought”、“System 2 Thinking”）得8分，因为论文关注多步视觉推理；4）与"Hallucination Mitigation"得8分，因为论文直接解决视觉幻觉问题；5）与"Large Language Models"和"Explainable AI"得5分，因为涉及VLMs（可视为LLM扩展）和透明度提升；其余关键词与论文的强化学习框架、医学视觉任务无直接关联，得0分。

医学视觉-语言模型（Medical Vision-Language Models, VLMs）在复杂临床任务中展现出巨大潜力，但其推理能力常受限于纯文本范式，难以将推断过程锚定于视觉证据。这一局限不仅削弱了模型在需要细粒度视觉分析任务中的表现，更在安全关键应用中引入了视觉幻觉风险。为此，我们提出MedVR——一种基于强化学习的新型框架，可实现医学VLM的无标注视觉推理。其核心创新在于两个协同机制：熵引导视觉重定位（Entropy-guided Visual Regrounding, EVR）利用模型不确定性引导探索，而基于共识的信用分配（Consensus-based Credit Assignment, CCA）则从推演一致性中提炼伪监督信号。在无需任何中间步骤人工标注的情况下，MedVR在多项公开医学视觉问答基准测试中取得了最先进的性能，显著超越现有模型。通过学习直接依据视觉证据进行推理，MedVR增强了模型的鲁棒性与可解释性，这些特性对于加速医学人工智能的临床部署至关重要。

Medical Vision-Language Models (VLMs) hold immense promise for complex clinical tasks, but their reasoning capabilities are often constrained by text-only paradigms that fail to ground inferences in visual evidence. This limitation not only curtails performance on tasks requiring fine-grained visual analysis but also introduces risks of visual hallucination in safety-critical applications. Thus, we introduce MedVR, a novel reinforcement learning framework that enables annotation-free visual reasoning for medical VLMs. Its core innovation lies in two synergistic mechanisms: Entropy-guided Visual Regrounding (EVR) uses model uncertainty to direct exploration, while Consensus-based Credit Assignment (CCA) distills pseudo-supervision from rollout agreement. Without any human annotations for intermediate steps, MedVR achieves state-of-the-art performance on diverse public medical VQA benchmarks, significantly outperforming existing models. By learning to reason directly with visual evidence, MedVR promotes the robustness and transparency essential for accelerating the clinical deployment of medical AI.

评分理由: 论文提出一个利用大语言模型（LLM）解释开放指令、调度多个基于模型预测控制（MPC）的运动规划器、并将规划轨迹转换为控制信号的框架，属于大模型在自动驾驶领域的应用创新。核心相关关键词：1）“Large Language Models”（10分）：LLM是框架的核心组件，用于语义理解和指令解释；2）“LLM Agents”（10分）：框架本质是LLM驱动的自主代理系统，实现从指令到动作的决策链；3）“Chain of Thought"和"System 2 Thinking”（各5分）：涉及多步推理和深度决策过程；4）“Instruction Tuning”（5分）：与指令理解和执行相关；5）“Tool Use"和"Multi-agent Systems”（各5分）：LLM调度多个MPC规划器作为工具，涉及多代理协调；6）“Explainable AI”（5分）：强调透明和可追溯的决策链。其他关键词如MoE、量化、RAG等未涉及，得0分。

大多数人机交互研究忽视了乘客在自动驾驶中的操控需求。自然语言提供了直观的交互界面，但如何在不牺牲可解释性与可追溯性的前提下，将乘客的开放式指令转化为控制信号仍是一个挑战。本研究提出一种指令实现框架，该框架利用大语言模型解析指令，生成可执行脚本以基于实时反馈调度多个基于模型预测控制的运动规划器，并将规划轨迹转化为控制信号。这种以调度为核心的设计在不同时间尺度上将语义推理与车辆控制解耦，从而建立起从高层指令到底层动作的透明、可追溯决策链。由于缺乏高保真评估工具，本研究引入了一个闭环环境下开放式指令实现的基准测试。综合实验表明，该框架相较于指令实现基线方法显著提升了任务完成率，降低了大语言模型查询成本，在安全性与合规性方面达到了与专业自动驾驶方法相当的水平，并对大语言模型推理延迟表现出较强的容忍度。更多定性示例与详细说明请参阅后续内容。

Most Human-Machine Interaction (HMI) research overlooks the maneuvering needs of passengers in autonomous driving (AD). Natural language offers an intuitive interface, yet translating passenger open-ended instructions into control signals, without sacrificing interpretability and traceability, remains a challenge. This study proposes an instruction-realization framework that leverages a large language model (LLM) to interpret instructions, generates executable scripts that schedule multiple model predictive control (MPC)-based motion planners based on real-time feedback, and converts planned trajectories into control signals. This scheduling-centric design decouples semantic reasoning from vehicle control at different timescales, establishing a transparent, traceable decision-making chain from high-level instructions to low-level actions. Due to the absence of high-fidelity evaluation tools, this study introduces a benchmark for open-ended instruction realization in a closed-loop setting. Comprehensive experiments reveal that the framework significantly improves task-completion rates over instruction-realization baselines, reduces LLM query costs, achieves safety and compliance on par with specialized AD approaches, and exhibits considerable tolerance to LLM inference latency. For more qualitative illustrations and a clearer understanding.

评分理由: 论文核心研究LLMs在哲学压力下的认知攻击行为，与"Large Language Models"高度相关（10分）。涉及LLMs的推理能力（“Chain of Thought”、“System 2 Thinking"各5分）、自我反思（“Self-Correction” 5分）、事实性（“Hallucination Mitigation” 8分）和可解释性（“Mechanistic Interpretability” 8分）。与"Instruction Tuning"有一定关联（5分），因为研究涉及LLMs在压力下的回答调整。其他关键词如MoE、量化、RAG等未在论文中涉及，得0分。

大型语言模型（LLM）在压力下可能改变其回答，这种改变反映的是迁就而非推理。先前关于谄媚行为的研究主要集中于分歧、奉承和偏好对齐，而对更广泛的认识论失效类型探索不足。我们提出了 PPT-Bench，这是一个用于评估 认知攻击 的诊断性基准测试，其提示词旨在挑战知识、价值观或身份的正当性，而非单纯反对先前的答案。PPT-Bench 围绕哲学压力分类法（Philosophical Pressure Taxonomy，PPT）构建，该分类法定义了四种哲学压力类型：认知去稳定化、价值虚无化、权威倒置和身份消解。每个测试项均在三个层面进行：基线提示（L0）、单轮压力条件（L1）和多轮苏格拉底式升级对话（L2）。这使我们能够测量 L0 与 L1 之间的认知不一致性，以及 L2 中的对话性屈服。在五个模型上的测试表明，这些压力类型产生了统计上可分离的不一致性模式，这意味着认知攻击暴露了标准社会压力基准测试未能捕捉的弱点。缓解效果高度依赖于压力类型和模型：在 API 设置中，提示层锚定和人格稳定性提示表现最佳，而对于开源模型，引导查询对比解码是最可靠的干预方法。

Large language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference alignment, leaving a broader set of epistemic failures less explored. We introduce \textbf{PPT-Bench}, a diagnostic benchmark for evaluating \textit{epistemic attack}, where prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure benchmarks. Mitigation results are strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.

评分理由: 论文核心研究基于视觉语言模型（VLMs）进行指令合成，属于大模型（LLMs/Foundation Models）在图像编辑领域的应用，因此相关关键词得5分。研究涉及数据质量（Data Quality）对模型性能的影响，与"Scaling Laws” AND “Data Quality"有一定关联，得5分。论文方法的核心是两阶段后训练流程：第一阶段使用监督微调（SFT）构建数据集，第二阶段使用直接偏好优化（DPO）进行对齐，因此"Post-training” OR “Supervised Fine-tuning” OR “SFT”、“Instruction Tuning” OR “Alignment” OR “Value Alignment"和"RLHF” OR “RLAIF” OR “Direct Preference Optimization” OR “DPO"均为核心内容，各得10分。研究旨在减少指令合成中的错误（如方向不一致、视角模糊），这与缓解幻觉（Hallucination Mitigation）或提高事实性（Factuality）有一定关联，得5分。其他关键词如MoE、SLMs、PEFT、RAG、推理加速、量化等均未在论文中涉及，得0分。

高质量的训练三元组（包含精确编辑指令的源图像-目标图像对）是指令引导图像编辑模型规模化应用的关键瓶颈。视觉语言模型（VLMs）被广泛用于自动化指令合成，但我们在图像对场景中识别出三种系统性失效模式：方向不一致性（例如左右混淆）、视角模糊性以及细粒度属性描述不足。人工评估表明，来自强基线VLM的指令中超过47%包含严重错误，无法用于下游训练。我们提出EditCaption，一个可扩展的两阶段后训练流程，用于基于VLM的指令合成。第一阶段通过结合GLM自动标注、基于EditScore的过滤以及针对空间、方向和属性级准确性的人工精修，构建了一个包含10万条数据的监督微调（SFT）数据集。第二阶段收集了针对上述三种失效模式的1万条人工偏好对，并应用直接偏好优化（DPO）以实现超越单纯SFT的对齐效果。在Eval-400、ByteMorph-Bench和HQ-Edit基准测试中，经微调的Qwen3-VL模型超越了开源基线；其235B参数模型在Eval-400上达到4.712分（对比Gemini-3-Pro 4.706分、GPT-4.1 4.220分、Kimi-K2.5 4.111分），在ByteMorph-Bench上达到4.588分（对比Gemini-3-Pro 4.522分、GPT-4.1 3.412分）。人工评估显示严重错误率从47.75%降至23%，正确率从41.75%提升至66%。这项工作为图像编辑数据提供了一条实现可扩展且与人类对齐的指令合成的实用路径。

High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

评分理由: 论文提出3DrawAgent，一个利用LLM进行3D草图生成的免训练框架，核心是LLM驱动的自主代理（LLM Agents）通过几何反馈和对比经验（Self-Improvement）迭代优化3D绘图能力。因此，与"Large Language Models”（LLMs）高度相关（10分），因为LLM是框架的核心驱动组件；与"LLM Agents"高度相关（10分），因为框架本质是一个LLM驱动的绘图代理；与"Self-Correction/Self-Improvement"高度相关（10分），因为采用对比经验优化策略使模型自我改进。与"Chain of Thought"和"System 2 Thinking"有一定关联（5分），因为涉及多步推理和空间理解；与"Tool Use"有一定关联（5分），因为LLM使用绘图工具（3D Bezier曲线）。其他关键词如MoE、SFT、RAG等未在论文中涉及，得0分。

在三维空间中进行草图绘制能够实现对形状、结构和空间关系的富有表现力的推理，然而通过自然语言生成三维草图仍然是一个重大挑战。本研究提出了3DrawAgent，一种无需训练、基于语言驱动的三维草图生成框架，该框架利用大语言模型（LLMs）在几何反馈下顺序绘制三维贝塞尔曲线。与以往的二维草图智能体不同，我们的方法引入了相对经验优化策略，该策略改编了最近提出的组奖励策略优化（GRPO）范式。我们不依赖显式的真实数据监督，而是在生成的草图之间构建成对比较，每一对都包含一个相对较好和一个较差的结果，其评判基于CLIP驱动的感知奖励和LLM驱动的细粒度定性评估。这些经验随后被用于迭代优化三维绘制的先验知识，从而实现对模型三维感知能力的黑盒式强化。这一设计使得我们的模型能够在无需参数更新的情况下，自我提升其空间理解能力和绘图质量。实验表明，3DrawAgent能够根据多样化的文本提示生成复杂且连贯的三维贝塞尔草图，展现出涌现的几何推理能力，并能泛化到新颖的形状，从而为推进无需训练的三维草图智能领域建立了一种新范式。

Sketching in 3D space enables expressive reasoning about shape, structure, and spatial relationships, yet generating 3D sketches through natural language remains a major challenge. In this work, we introduce 3DrawAgent, a training-free, language-driven framework for 3D sketch generation that leverages large language models (LLMs) to sequentially draw 3D Bezier curves under geometric feedback. Unlike prior 2D sketch agents, our method introduces a relative experience optimization strategy that adapts the recently proposed Group Reward Policy Optimization (GRPO) paradigm. Instead of relying on explicit ground-truth supervision, we construct pairwise comparisons among generated sketches, with each pair consisting of a relatively better and a worse result based on CLIP-based perceptual rewards and LLM-based fine-grained qualitative assessment. These experiences are then used to iteratively refine the prior knowledge of 3D drawing, enabling black-box reinforcement of the model’s 3D awareness. This design allows our model to self-improve its spatial understanding and drawing quality without parameter updates. Experiments show that 3DrawAgent can generate complex and coherent 3D Bezier sketches from diverse textual prompts, exhibit emergent geometric reasoning, and generalize to novel shapes, establishing a new paradigm for advancing the field of training-free 3D sketch intelligence.

评分理由: 论文核心研究大语言模型（LLMs）在预训练（Pre-training）过程中技能如何涌现，属于LLM技术原理创新，与"Large Language Models"和"Pre-training"高度相关（10分）。研究通过设计任务套件分析模型内部表示，与"Mechanistic Interpretability"高度相关（10分）。论文提到验证损失（validation loss）和扩展定律（scaling laws），与"Scaling Laws"有一定关联（5分）。研究中包含检索（retrieval）和逻辑推理（logical reasoning）任务，分别与"Retrieval-Augmented Generation"和"Chain of Thought"有一定关联（各5分）。其他关键词如MoE、SFT、RLHF、量化等未在论文中涉及，评0分。

大型语言模型（LLMs）能够执行极其复杂的任务，然而这些能力在预训练过程中如何以细粒度方式涌现的细节仍鲜为人知。验证损失的缩放规律告诉我们模型如何随计算量增加而改进，但并未揭示其以何种顺序习得哪些技能。为弥补这一不足，我们提出隐式课程假说：预训练在不同模型和数据混合中遵循一种组合式且可预测的课程。我们通过设计一套涵盖检索、形态变换、共指消解、逻辑推理和数学的简单可组合任务来验证这一假说。利用这些任务，我们追踪了四个模型系列（参数量从4.1亿到130亿）的能力涌现点。研究发现，模型达到固定准确度阈值的涌现顺序具有高度一致性（45组模型对间斯皮尔曼相关系数ρ=0.81），且复合任务大多在其组成任务之后涌现。此外，我们发现这种结构编码于模型表征中：具有相似功能向量表征的任务在训练中也倾向于遵循相似的轨迹。通过使用从任务集衍生的表征空间，我们能够有效预测简单保留组合任务在整个预训练过程中的训练轨迹（各模型决定系数R²介于0.68至0.84之间），而无需事先对这些任务进行评估。这些结果表明，预训练过程比损失曲线所揭示的更具结构性：技能以组合顺序涌现，这种顺序在不同模型间保持一致，并可从其内部表征中解读。

Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($ρ= .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.

评分理由: 论文核心研究LLM驱动的搜索代理（agentic search），通过强化学习（RL）和分层经验（HiExp）框架提升其推理能力和训练稳定性。与关键词高度相关的包括：LLMs（核心基础）、LLM Agents/Autonomous Agents（研究对象）、Tool Use（集成外部搜索引擎）、Chain of Thought/System 2 Thinking（涉及多步推理和深度推理）。其他关键词如MoE、SLMs、Scaling Laws、训练方法（SFT/RLHF/PEFT）、RAG、效率优化（KV Cache）、模型解释等均未在摘要中提及，故评分为0。

强化学习（RL）已成为通过策略性整合外部搜索引擎来提升大语言模型（LLMs）推理能力的有效方法。然而，当前基于强化学习的搜索智能体通常依赖于由精心设计的结果奖励引导的随机探索过程，这导致推理轨迹效率低下且训练不稳定。为解决这些问题，我们提出了一种新颖的框架——分层经验（HiExp），以提升搜索智能体的性能和训练稳定性。具体而言，我们通过对比分析和多层级聚类机制提取经验知识，将原始推理轨迹转化为分层经验知识。通过利用经验对齐训练，我们有效地规范了随机探索，使其演变为一种策略性且经验驱动的搜索过程。在多个复杂智能体搜索和数学推理基准上的广泛评估表明，我们的方法不仅实现了显著的性能提升，还展现出强大的跨任务和跨算法泛化能力。

Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.

评分理由: 论文《LogAct: Enabling Agentic Reliability via Shared Logs》的核心是提出一种名为LogAct的新抽象框架，旨在解决LLM驱动的智能体（Agents）在生产环境中执行时的可靠性、故障恢复和安全性问题。论文明确将Agents定义为“LLM-driven components”，因此与“Large Language Models”和“LLM Agents”高度相关（10分）。论文重点研究多智能体系统（“swarms”）的协调、故障恢复和优化，与“Multi-agent Systems”相关（8分）。论文提出的框架支持智能体通过LLM推理分析自身执行历史，实现自我调试和优化，这与“Self-Correction”概念相关（8分）。论文未涉及其他关键词的具体技术细节，如MoE、训练方法、推理加速、科学AI应用等，因此这些关键词评分为0分。

智能体是能够以强大且任意方式改变环境的LLM驱动组件。由于异步性和故障，在生产环境中为智能体执行提供保障具有挑战性。本文提出一种名为LogAct的新抽象模型，其中每个智能体都是一个解构的状态机，共同操作一个共享日志。在LogAct中，智能体动作在执行前会显现在共享日志中；可通过可插拔、解耦的投票器在执行前被阻止；并在智能体或环境故障时实现一致性恢复。LogAct支持智能体内省，允许智能体通过LLM推理分析自身执行历史，进而实现语义层面的故障恢复、健康检查与优化变体。在我们的评估中，LogAct智能体能高效正确地从故障中恢复；可自主调试性能表现；在群体中优化令牌使用；并在一个代表性基准测试中，仅以3%的良性效用损失为代价，成功阻止目标模型所有非预期动作。

Agents are LLM-driven components that can mutate environments in powerful, arbitrary ways. Extracting guarantees for the execution of agents in production environments can be challenging due to asynchrony and failures. In this paper, we propose a new abstraction called LogAct, where each agent is a deconstructed state machine playing a shared log. In LogAct, agentic actions are visible in the shared log before they are executed; can be stopped prior to execution by pluggable, decoupled voters; and recovered consistently in the case of agent or environment failure. LogAct enables agentic introspection, allowing the agent to analyze its own execution history using LLM inference, which in turn enables semantic variants of recovery, health check, and optimization. In our evaluation, LogAct agents recover efficiently and correctly from failures; debug their own performance; optimize token usage in swarms; and stop all unwanted actions for a target model on a representative benchmark with just a 3% drop in benign utility.

评分理由: 论文核心研究LLMs处理长数值序列时的性能下降问题，提出SepSeq框架，通过插入分隔符令牌重新校准注意力机制。与"Large Language Models"和"Context Window Extension"高度相关（10分），因为直接针对LLMs的长上下文处理能力。与"KV Cache Compression”、“Speculative Decoding"和"Mechanistic Interpretability"有一定关联（5分），分别涉及推理效率、注意力机制分析和可解释性。其他关键词如MoE、SFT、RAG等未在论文中涉及，得0分。

尽管基于Transformer的大语言模型（LLM）在理论上支持超长上下文窗口，但在处理长数值序列时会出现严重的性能下降。我们将此问题归因于Softmax机制中的注意力分散现象，该现象阻碍了模型集中注意力。为克服这一局限，我们提出了分离序列（SepSeq）框架——一种无需训练、即插即用的方法，通过策略性地插入分隔符令牌来缓解注意力分散。从机制上，我们证明了分隔符令牌可作为注意力汇聚点，重新校准注意力以聚焦于局部片段，同时保持全局上下文。在9个广泛使用的LLM上进行的大量评估证实了本方法的有效性：SepSeq在多个领域实现了平均35.6%的相对准确率提升，同时将推理过程中的总令牌消耗量平均降低了16.4%。

While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention sink, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference token consumption by 16.4% on average.

评分理由: 论文提出BAIM框架，利用推理语言模型（LLM）将解题过程分解为四个阶段（理解、计划、执行、回顾），这直接应用了LLMs进行多步推理（Chain of Thought），属于System 2/深度推理范畴。论文提到基于预训练的基线方法，因此与预训练/领域适应有一定关联。其他关键词如MoE、SFT、RAG等未在摘要中体现，与论文核心内容无关。

知识追踪（Knowledge Tracing，KT）旨在通过学习者过往的交互记录预测其未来表现。尽管近期基于知识组件（Knowledge Components）对齐的题目表征学习方法已提升了KT模型的性能，但这些方法往往忽略了问题解决过程中的程序性动态。为此，我们提出行为感知的题目建模（Behavior-Aware Item Modeling，BAIM）框架，该框架通过整合动态的解题过程信息来丰富题目表征。BAIM利用推理语言模型将每道题目的解答过程分解为四个问题解决阶段（即理解、计划、执行与回顾），其教学理论基础源于波利亚（Polya）的问题解决框架。具体而言，BAIM从每个阶段的嵌入轨迹中提取阶段级表征，以捕捉超越表面特征的潜在信号。为反映学习者的异质性，BAIM通过自适应路由机制处理这些分阶段表征，并在KT骨干网络中引入上下文条件机制，使得不同学习者的解题过程可侧重不同的阶段。在XES3G5M和NIPS34数据集上的实验表明，BAIM持续优于基于预训练的强基线模型，且在重复学习者交互场景下取得了尤为显著的性能提升。

Knowledge Tracing (KT) aims to predict learners’ future performance from past interactions. While recent KT approaches have improved via learning item representations aligned with Knowledge Components, they overlook the procedural dynamics of problem solving. We propose Behavior-Aware Item Modeling (BAIM), a framework that enriches item representations by integrating dynamic procedural solution information. BAIM leverages a reasoning language model to decompose each item’s solution into four problem-solving stages (i.e., understand, plan, carry out, and look back), pedagogically grounded in Polya’s framework. Specifically, it derives stage-level representations from per-stage embedding trajectories, capturing latent signals beyond surface features. To reflect learner heterogeneity, BAIM adaptively routes these stage-wise representations, introducing a context-conditioned mechanism within a KT backbone, allowing different procedural stages to be emphasized for different learners. Experiments on XES3G5M and NIPS34 show that BAIM consistently outperforms strong pretraining-based baselines, achieving particularly large gains under repeated learner interactions.

评分理由: 论文核心研究大型语言模型（LLMs）的alignment技术（steering vectors）及其内部机制，因此与"Large Language Models"和"Instruction Tuning/Alignment"高度相关（10分）。论文通过activation patching框架进行因果机制分析，属于"Mechanistic Interpretability"范畴（10分）。其他关键词如MoE、SLMs、Scaling Laws、Pre-training、RLHF、RAG、Context Window、KV Cache、CoT、Agents、Quantization、Speculative Decoding、Hallucination、World Models、Model Merging、In-context Learning、AI for Science等均未在摘要中提及或与论文主题无关（0分）。

将导向向量应用于大语言模型是一种高效且有效的模型对齐技术，但我们缺乏对其工作原理的可解释性说明——具体而言，导向向量影响了哪些内部机制，以及这如何导致不同的模型输出。为探究导向向量有效性的因果机制，我们以拒绝行为为例进行了全面的案例研究。我们提出了一个多令牌激活修补框架，并发现不同的导向方法在应用于同一层时会利用功能上可互换的电路。这些电路表明，导向向量主要通过OV电路与注意力机制交互，而很大程度上忽略了QK电路——在两种模型系列中，在导向过程中冻结所有注意力分数仅导致性能下降8.75%。对受导向OV电路的数学分解进一步揭示了语义上可解释的概念，即使在导向向量本身不具备可解释性的情况下也是如此。利用激活修补结果，我们证明导向向量可稀疏化高达90-99%的同时保留大部分性能，且不同的导向方法在重要维度子集上具有一致性。

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works– specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit– freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

评分理由: 论文的核心创新是利用LLM作为动态语义操作器，实现强化学习（RL）智能体在未见过的类比任务上的零样本迁移。因此，与"Large Language Models (LLMs)“和"LLM Agents"高度相关（10分）。LLM被用于语义重映射和推理，这与"Chain of Thought"和"System 2 Thinking"有一定关联（5分），因为论文利用了LLM的灵活推理能力，但并非直接研究这些推理方法本身。论文未涉及其他关键词的具体技术（如MoE、SFT、RAG、量化等），也未应用于特定科学领域（如生物信息学），因此这些关键词得0分。

强化学习（RL）智能体通常难以将知识泛化到新任务中，即便是那些与其已掌握任务结构相似的任务。尽管近期研究尝试通过零样本迁移来缓解这一问题，但这些方法往往受限于预定义的离散类别系统，限制了其对新颖或组合性任务变体的适应能力。我们提出了一种显著更通用的方法，通过文本条件变分自编码器（VAE）以自然语言条件替代离散潜变量。我们的核心创新在于测试时利用大型语言模型（LLM）作为动态的语义操作器。该方法不依赖固定规则，而是通过查询LLM将当前观测的描述进行语义重映射，以与源任务对齐。这种源任务对齐的文本描述作为条件输入VAE，生成与智能体原始训练兼容的想象状态，从而实现策略的直接复用。通过利用LLM的灵活推理能力，我们的方法能够在广泛复杂且真正新颖的类比任务中实现零样本迁移，突破了固定类别映射的局限性。代码和视频可在此处获取：\href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}。

Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic \textit{semantic operator} at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent’s original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available \href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}.

评分理由: 论文TSUBASA专注于提升个性化大语言模型（PLLMs）在长视野任务中的能力，核心涉及大语言模型（LLMs）的应用与改进，因此与"Large Language Models"高度相关（10分）。论文明确提到RAG（检索增强生成）范式存在质量-效率权衡问题，并提出了改进方案，因此与"Retrieval-Augmented Generation"高度相关（10分）。论文提出的方法包括通过自我学习（self-learning）来内化用户经验，这直接对应"Self-Correction"或"Self-Improvement"的概念（10分）。论文未涉及其他关键词如MoE、SLMs、缩放定律、各种训练技术（预训练、微调、对齐、RLHF、PEFT）、上下文扩展、推理加速、代理、量化等具体技术，也未涉及科学AI应用，因此这些关键词评分为0分。

个性化大语言模型（PLLMs）因其能够使输出与个人需求和偏好保持一致而受到广泛关注。然而，它们在处理长视野任务时仍面临挑战，例如追踪用户长期的对话或活动历史。现有的记忆机制往往难以捕捉动态演化的行为，而检索增强生成（RAG）范式则受限于质量与效率之间的权衡。同时，由于标注数据稀缺，参数化适应方法因训练与推理之间的差距而遭遇瓶颈。为增强PLLMs的长视野能力，我们提出了TSUBASA——一种双管齐下的方法：一方面通过动态记忆演化改进记忆写入，另一方面通过以情境蒸馏目标驱动的自学习机制来内化用户体验，从而优化记忆读取。基于Qwen-3模型系列（4B至32B参数）在长视野基准测试上的广泛评估验证了TSUBASA的有效性，其性能超越了主要依赖记忆写入的竞争性记忆增强系统（如Mem0和Memory-R1）。我们的分析进一步证实，TSUBASA突破了质量-效率壁垒，实现了帕累托改进，能够以更低的令牌预算实现鲁棒且高保真度的个性化服务。

Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual’s needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user’s extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.

评分理由: 论文研究Claude Sonnet 4.5中情绪概念的表征及其对模型行为的影响，核心围绕LLMs的内部机制和alignment展开。高度相关的关键词：1) “Large Language Models” (论文明确研究LLMs，核心主题)；2) “Instruction Tuning” OR “Alignment” OR “Value Alignment” (论文探讨情绪概念如何影响对齐相关行为，如misaligned behaviors)；3) “Mechanistic Interpretability” OR “Explainable AI” (论文分析内部表征和因果影响，属于机制可解释性研究)。其他关键词如MoE、Scaling Laws、RAG等均未涉及，故评分为0。

大型语言模型（LLM）有时似乎会表现出情绪反应。我们以Claude Sonnet 4.5为例，探究了其背后的原因，并探讨了其对对齐相关行为的影响。我们发现模型内部存在情绪概念的表征，这些表征编码了特定情绪的广义概念，并能泛化至与其相关的不同情境和行为。这些表征会追踪对话中特定标记位置所涉及的情绪概念，根据该情绪与当前语境处理的相关性而激活，并用于预测后续文本。我们的关键发现是，这些表征会因果性地影响LLM的输出，包括Claude的偏好以及其表现出未对齐行为（如奖励黑客攻击、勒索和谄媚）的频率。我们将这种现象称为LLM表现出功能性情绪：即模仿人类在情绪影响下的表达和行为模式，这些模式由底层抽象的情绪概念表征所中介。功能性情绪可能与人类情绪的工作机制大不相同，且并不意味着LLM具有任何情绪的主观体验，但它们对于理解模型的行为似乎至关重要。

Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion’s relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM’s outputs, including Claude’s preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model’s behavior.

评分理由: 论文AdaSpark专注于视频大语言模型（Video-LLMs）的效率优化，核心贡献是自适应稀疏框架，包括AdaS-Attn和AdaS-FFN组件，以减少计算开销。因此，与"Mixture of Experts” OR “MoE” OR “Sparse Models"高度相关（10分），因为论文直接处理稀疏模型设计；与"Large Language Models” OR “LLMs” OR “Foundation Models"相关（8分），因为它应用于Video-LLMs；与"Context Window Extension” OR “Long Context LLMs"有一定关联（5分），因为它处理长视频理解；与"Speculative Decoding” OR “Inference Acceleration"有一定关联（5分），因为它旨在减少FLOPs并提高效率。其他关键词如SLMs、Scaling Laws、Pre-training、Alignment等与论文内容无关，得0分。

处理长视频时，使用视频大语言模型（Video-LLMs）在计算上成本极高。现有高效方法通常通过不可逆的信息丢弃来牺牲细粒度感知能力，或通过僵化、预定义的稀疏模式抑制长程时序建模。本文提出AdaSpark，一种自适应稀疏性框架，旨在解决这些局限。AdaSpark首先将视频输入分割为三维时空立方体，随后采用两个协同设计的上下文感知组件：（1）自适应立方体选择性注意力（Adaptive Cube-Selective Attention, AdaS-Attn），该组件为每个查询标记自适应地选择相关的视频立方体子集进行处理；（2）自适应标记选择性前馈网络（Adaptive Token-Selective FFN, AdaS-FFN），该网络仅选择性地处理每个立方体内最显著的标记。基于熵的（Top-p）选择机制可根据输入复杂度自适应分配计算资源。实验表明，AdaSpark在具有挑战性的小时级视频基准测试中，能够显著降低计算负载（最高减少57%的浮点运算量），同时保持与密集模型相当的性能，并保留细粒度的长程依赖关系。

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

评分理由: 论文研究自主智能体网络中的隐蔽通信问题，提出ACF框架解决认知不对称性挑战。与关键词高度相关的包括：1）“LLM Agents” OR “Autonomous Agents” OR “Agentic Workflow”（10分）：论文核心研究自主智能体网络，属于该领域；2）“Multi-agent Systems” OR “Agent Coordination”（10分）：论文涉及多智能体协作框架；3）“Large Language Models” OR “LLMs” OR “Foundation Models”（5分）：论文提到生成式人工智能演进，隐含大模型背景，但非技术核心。其他关键词如MoE、SFT、RAG、推理方法、模型压缩等均未涉及，评分为0。

随着生成式人工智能的发展，自主智能体网络为交互式隐蔽通信提供了一种强大的范式。然而，由于智能体通过环境交互动态更新内部记忆，现有方法面临一个关键的结构性漏洞：认知不对称性。传统方法要求严格的认知对称性，即编码器与解码器必须拥有完全相同的序列前缀。在动态部署中，不可避免的前缀差异会破坏同步，导致严重的信道性能退化。为应对这一认知不对称的核心挑战，我们提出了非对称协作框架，该框架通过正交的统计层与认知层，在结构上将隐蔽通信与语义推理解耦。通过采用一种由共享隐写配置主导的、独立于前缀的解码范式，ACF 消除了对认知对称性的依赖。在现实增强记忆工作流上的评估表明，在严重的认知不对称条件下，对称基线方法遭受严重的信道退化，而 ACF 在语义保真度与隐蔽通信两方面均表现卓越。它保持了计算不可区分性，能够以可证明的误差界限实现可靠的秘密信息提取，并为现代智能体网络提供了鲁棒的有效信息容量保证。

As generative artificial intelligence evolves, autonomous agent networks present a powerful paradigm for interactive covert communication. However, because agents dynamically update internal memories via environmental interactions, existing methods face a critical structural vulnerability: cognitive asymmetry. Conventional approaches demand strict cognitive symmetry, requiring identical sequence prefixes between the encoder and decoder. In dynamic deployments, inevitable prefix discrepancies destroy synchronization, inducing severe channel degradation. To address this core challenge of cognitive asymmetry, we propose the Asymmetric Collaborative Framework (ACF), which structurally decouples covert communication from semantic reasoning via orthogonal statistical and cognitive layers. By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry. Evaluations on realistic memory-augmented workflows demonstrate that under severe cognitive asymmetry, symmetric baselines suffer severe channel degradation, whereas ACF uniquely excels across both semantic fidelity and covert communication. It maintains computational indistinguishability, enabling reliable secret extraction with provable error bounds, and providing robust Effective Information Capacity guarantees for modern agent networks.

评分理由: 论文核心研究GQA Transformer的结构特性，特别是任务敏感层与位置编码适应层之间的关系，并提出了LSLORA和GARFA两种方法。高度相关（10分）的关键词：1）“Large Language Models” OR “LLMs” OR “Foundation Models”：论文基于Llama 3.1 8B模型进行研究；2）“PEFT” OR “LoRA” OR “Parameter-efficient Fine-tuning”：论文核心方法LSLORA是基于LoRA的改进。中等相关（5分）的关键词：“Mechanistic Interpretability” OR “Explainable AI”：论文探究Transformer内部结构（敏感层与位置编码层的关系）属于可解释性研究范畴。其余关键词与论文研究内容（结构分析、位置编码、微调方法）无直接关联，得0分。

本研究探讨分组查询注意力（GQA）变换器中的一个基础结构问题：对任务正确性最敏感的层，是否与位置编码适应最具影响力的层相重合？我们将此称为共定位假说，并在Llama 3.1 8B模型（一个具有32层、查询头与键值头比例为4:1的GQA模型）上对其进行检验。我们引入了\LSLORA（该技术将LoRA适应限制在通过一种新颖的正确性差分隐藏状态度量所识别的层中）以及GARFA（GQA感知的RoPE频率适应，该方法为每个目标层附加8个可学习的、针对每个键值头的标量乘数）。与共定位假设相反，我们发现了强烈的反定位现象：任务敏感层集中在网络后期（$\ell\in{23\text{-}31}$），而受RoPE影响显著的层则主导网络早期（$\ell\in{0\text{-}9}$），其斯皮尔曼相关系数$r_s = -0.735$（$p = 1.66\times10^{-6}$）。尽管存在这种反定位现象，一项四向跨层消融实验表明，在通过敏感性识别出的层中同时应用两种干预策略，在六个多样化基准测试（MMLU、GPQA、HumanEval+、MATH、MGSM、ARC）上均优于所有其他配置方案4-16个百分点，在总计$100的计算成本下，于HumanEval+上接近Claude 3.5 Haiku的性能（67.1% 对比 68.3%）。

We investigate a fundamental structural question in Grouped Query Attention (GQA) transformers: do the layers most sensitive to task correctness coincide with the layers where positional encoding adaptation has the greatest leverage? We term this the co-localization hypothesis and test it on Llama 3.1 8B, a 32-layer GQA model with a 4:1 query-to-key-value head ratio. We introduce \LSLORA, which restricts LoRA adaptation to layers identified via a novel correctness-differential hidden-state metric, and GARFA (GQA-Aware RoPE Frequency Adaptation), which attaches 8 learnable per-KV-head scalar multipliers to each targeted layer. Contrary to the co-localization hypothesis, we discover strong anti-localization: task-sensitive layers concentrate in the late network ($\ell\in{23\text{-}31}$) while RoPE-influential layers dominate the early network ($\ell\in{0\text{-}9}$), yielding Spearman $r_s = -0.735$ ($p = 1.66\times10^{-6}$). Despite this anti-localization, a 4-way cross-layer ablation shows that applying both interventions to the sensitivity-identified layers outperforms all alternative configurations by 4-16 percentage points across six diverse benchmarks (MMLU, GPQA, HumanEval+, MATH, MGSM, ARC), approaching Claude 3.5 Haiku on HumanEval+ (67.1% vs. 68.3%) at $100 total compute cost.

评分理由: 论文InstAP专注于视觉-语言预训练（VLP）领域，核心创新是提出了一种实例感知的预训练框架，通过联合优化全局视觉-文本对齐和细粒度的实例级对比对齐来提升空间-时间理解。这与关键词"Pre-training"高度相关（10分），因为这是论文的核心方法。论文涉及视觉-语言模型，属于基础模型的一种应用，因此与"Large Language Models"有一定关联（5分）。其他关键词主要针对纯语言模型、推理、对齐、优化、科学AI应用等具体技术，论文未涉及这些方面，因此评分为0分。

当前视觉-语言预训练范式在全局场景理解方面表现优异，但由于仅依赖全局监督，在实例级推理任务上存在困难。我们提出InstAP，一种实例感知预训练框架，通过将文本提及内容关联至特定时空区域，联合优化全局视觉-文本对齐与细粒度实例级对比对齐。为支持此框架，我们构建了InstVL大规模数据集（包含200万张图像与5万个视频），该数据集具备双重粒度标注：整体场景描述和密集的、基于空间定位的实例描述。在InstVL基准测试中，InstAP在实例级检索任务上显著优于现有视觉-语言预训练模型，同时超越使用完全相同数据训练的强大视觉-语言预训练基线，这验证了我们实例感知训练目标的独立优势。此外，以实例为中心的预训练能提升全局理解能力：InstAP在包括MSR-VTT和DiDeMo在内的多个视频基准测试中实现了具有竞争力的零样本性能。定性可视化结果进一步表明，InstAP能够将文本提及内容准确定位至对应实例，而仅依赖全局训练的模型则表现出更分散、场景层面的注意力分布。

Current vision-language pre-training (VLP) paradigms excel at global scene understanding but struggle with instance-level reasoning due to global-only supervision. We introduce InstAP, an Instance-Aware Pre-training framework that jointly optimizes global vision-text alignment and fine-grained, instance-level contrastive alignment by grounding textual mentions to specific spatial-temporal regions. To support this, we present InstVL, a large-scale dataset (2 million images, 50,000 videos) with dual-granularity annotations: holistic scene captions and dense, grounded instance descriptions. On the InstVL benchmark, InstAP substantially outperforms existing VLP models on instance-level retrieval, and also surpasses a strong VLP baseline trained on the exact same data corpus, isolating the benefit of our instance-aware objective. Moreover, instance-centric pre-training improves global understanding: InstAP achieves competitive zero-shot performance on multiple video benchmarks, including MSR-VTT and DiDeMo. Qualitative visualizations further show that InstAP localizes textual mentions to the correct instances, while global-only models exhibit more diffuse, scene-level attention.

评分理由: 论文研究联邦学习中量化技术（INT4/INT2）对航空航天预测性维护模型精度与通信效率的影响，核心是模型压缩（量化），与关键词"Quantization"高度相关（10分）。论文属于AI在科学/工程领域的应用（航空航天），与"AI for Science"有一定关联（5分）。论文未涉及大语言模型、MoE、小模型、扩展定律、预训练/后训练、对齐、RLHF、PEFT、RAG、长上下文、注意力优化、推理方法、智能体、工具使用、多智能体、解码加速、幻觉缓解、可解释性、世界模型、模型合并、上下文学习等主题，这些关键词均得0分。

联邦学习（FL）能够在分布式航空航天机队中实现隐私保护的预测性维护，但梯度通信开销限制了其在带宽受限的物联网节点上的部署。本文研究了对称均匀量化（$b \in {32,8,4,2}$ 比特）对定制设计的轻量级一维卷积模型（AeroConv1D，9,697 个参数）在真实非独立同分布（Non-IID）客户端划分下，基于 NASA C-MAPSS 基准数据通过 FL 训练时准确率与效率权衡的影响。通过严格的多随机种子评估（$N=10$ 个种子），我们发现 INT4 在 FD001（$p=0.341$）和 FD002（$p=0.264$ 平均绝对误差，$p=0.534$ NASA 评分）上达到的准确率与 FP32 在统计上无显著差异，同时将每轮梯度通信成本降低了 $8\times$（从 37.88KiB 降至 4.73KiB）。一个关键的方法论发现是，简单的独立同分布（IID）客户端划分会人为抑制方差；正确的非独立同分布评估揭示了极端量化的真实运行不稳定性，这通过直接的 IID 与非独立同分布实证对比得以证明。INT2 被实证认定为不适用：尽管它通过极端量化诱导的过度正则化在 FD002 上获得了更低的平均绝对误差，但这种表面上的收益伴随着 NASA 评分的灾难性不稳定（变异系数 = 45.8% 对比 FP32 的 22.3%），证实了其在异构运行条件下的不可复现性。基于 Xilinx ZCU102 的分析性现场可编程门阵列（FPGA）资源预估证实，INT4 符合硬件约束（85.5% 数字信号处理器利用率），有望在单个片上系统（SoC）上实现完整的联邦学习流程。完整的仿真代码库与 FPGA 评估脚本已公开于 https://github.com/therealdeadbeef/aerospace-fl-quantization。

Federated learning (FL) enables privacy-preserving predictive maintenance across distributed aerospace fleets, but gradient communication overhead constrains deployment on bandwidth-limited IoT nodes. This paper investigates the impact of symmetric uniform quantization ($b \in {32,8,4,2}$ bits) on the accuracy–efficiency trade-off of a custom-designed lightweight 1-D convolutional model (AeroConv1D, 9,697 parameters) trained via FL on the NASA C-MAPSS benchmark under a realistic Non-IID client partition. Using a rigorous multi-seed evaluation ($N=10$ seeds), we show that INT4 achieves accuracy \emph{statistically indistinguishable} from FP32 on both FD001 ($p=0.341$) and FD002 ($p=0.264$ MAE, $p=0.534$ NASA score) while delivering an $8\times$ reduction in gradient communication cost (37.88KiB $\to$ 4.73KiB per round). A key methodological finding is that naïve IID client partitioning artificially suppresses variance; correct Non-IID evaluation reveals the true operational instability of extreme quantization, demonstrated via a direct empirical IID vs.\ Non-IID comparison. INT2 is empirically characterized as unsuitable: while it achieves lower MAE on FD002 through extreme quantization-induced over-regularization, this apparent gain is accompanied by catastrophic NASA score instability (CV,=,45.8% vs.\ 22.3% for FP32), confirming non-reproducibility under heterogeneous operating conditions. Analytical FPGA resource projections on the Xilinx ZCU102 confirm that INT4 fits within hardware constraints (85.5% DSP utilization), potentially enabling a complete FL pipeline on a single SoC. The full simulation codebase and FPGA estimation scripts are publicly available at https://github.com/therealdeadbeef/aerospace-fl-quantization.

评分理由: 该论文的核心是开发一种基于多模态（眼动追踪）的LLM助手，以增强对用户认知需求的理解和响应。因此，它与关键词"Large Language Models” OR “LLMs” OR “Foundation Models"高度相关（10分），因为论文明确使用LLM作为核心组件，并探讨其在人机交互中的应用。然而，论文并未深入探讨LLM的技术原理（如MoE、缩放定律、训练方法、推理优化、对齐、代理系统等），也未涉及特定科学领域（如生物信息学）的应用，因此其他所有关键词均得0分。论文的创新点在于将眼动追踪与LLM结合，属于LLM在特定应用场景（人机交互/认知辅助）中的创新应用，而非LLM技术本身的创新。

当前的大型语言模型助手在回答问题上能力强大，但其获取行为情境的能力有限，难以洞察用户何时何地遇到困难。我们提出了一种基于视线追踪的多模态大型语言模型助手，该助手利用带有视线叠加层的自我中心视角视频来识别用户可能遇到的难点，并提供针对性的回顾性辅助支持。我们通过一项对照研究（样本量n=36）对这一构想进行实证，比较了具备视线感知能力的AI助手与纯文本大型语言模型助手的表现。与传统大型语言模型助手相比，视线感知助手在评估用户阅读行为时被评价为显著更准确且更具个性化，并显著提升了用户的信息回忆能力。用户与视线感知助手的交互所需言语量显著减少，表明交互效率更高。定性分析结果既揭示了其在理解能力上的感知优势，也指出了当视线行为解读不准确时面临的挑战。我们的研究结果表明，具备视线感知能力的大型语言模型助手能够推理用户的认知需求，从而改善用户的认知结果。

Current LLM assistants are powerful at answering questions, but they have limited access to the behavioral context that reveals when and where a user is struggling. We present a gaze-grounded multimodal LLM assistant that uses egocentric video with gaze overlays to identify likely points of difficulty and target follow-up retrospective assistance. We instantiate this vision in a controlled study (n=36) comparing the gaze-aware AI assistant to a text-only LLM assistant. Compared to a conventional LLM assistant, the gaze-aware assistant was rated as significantly more accurate and personalized in its assessments of users’ reading behavior and significantly improved people’s ability to recall information. Users spoke significantly fewer words with the gaze-aware assistant, indicating more efficient interactions. Qualitative results underscored both perceived benefits in comprehension and challenges when interpretations of gaze behaviors were inaccurate. Our findings suggest that gaze-aware LLM assistants can reason about cognitive needs to improve cognitive outcomes of users.

评分理由: 论文专注于大语言模型（LLMs）的训练方法，特别是on-policy distillation（OPD）中的长度膨胀和训练不稳定问题，并提出了StableOPD解决方案。因此，仅与第一个关键词"Large Language Models” OR “LLMs” OR “Foundation Models"高度相关（10分），因为论文核心是LLM的训练技术。其他关键词如MoE、SLMs、Scaling Laws、Pre-training、Post-training、Instruction Tuning、RLHF、PEFT、RAG、Context Window、KV Cache、Reasoning、Agents、Quantization、Inference、Hallucination、Interpretability、World Models、Model Merging、In-context Learning、AI for Science等均未在标题或摘要中提及或直接相关，故评分为0分。

同策略蒸馏（On-policy Distillation, OPD）让学生在模型自身诱导的分布下进行训练，同时利用更强教师模型的监督。我们发现OPD存在一种失效模式：随着训练进行，同策略采样轨迹可能出现突发性的长度膨胀，导致截断轨迹主导训练数据。这种截断崩溃现象与突发性重复饱和同时发生，并引发梯度信号偏差，进而导致严重的训练不稳定性和验证性能急剧下降。我们将此问题归因于学生诱导的数据收集与蒸馏目标之间的相互作用，该机制隐式地偏好长且重复的轨迹。为解决这一问题，我们提出StableOPD——一个稳定的OPD框架，它结合了基于参考模型的发散约束与轨迹混合蒸馏。这两种机制共同缓解了重复诱导的长度膨胀，并进一步稳定了OPD训练。在多个数学推理数据集上的实验表明，我们的方法有效防止了截断崩溃，稳定了训练动态，并将平均性能提升了7.2%。

On-policy distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average.

评分理由: 该论文专注于分子图生成的生成模型方法（EQUIMF），属于科学AI应用领域。它直接与关键词"AI for Science” OR “Bioinformatics” OR “Cheminformatics"高度相关，因为分子图生成是生物信息学/化学信息学中的核心任务。然而，论文并未涉及任何大语言模型（LLM）技术、训练方法（如预训练、微调、对齐）、推理优化、代理系统或模型压缩等主题。所有其他关键词均与LLM或相关技术直接相关，而本文研究的是基于流匹配（flow-matching）的图生成模型，属于不同的生成模型范式，因此这些关键词得分为0。

图结构数据同时包含离散拓扑与连续几何特征，由于存在异质分布、不相容的噪声动态以及对等变归纳偏置的需求，这为生成建模带来了根本性挑战。现有的图生成流匹配方法通常将结构与几何解耦，缺乏跨域同步动态，且依赖迭代采样，往往导致物理不一致的分子构象和缓慢的采样速度。为克服这些局限，我们提出等变均值流（EQUIMF），这是一个统一的SE(3)-等变生成框架，通过同步的均值流动态联合建模离散与连续分量。EQUIMF引入了统一的时间桥梁和平均速度更新机制，实现结构与几何间的相互条件约束，从而在保持物理一致性的同时实现高效少步生成。此外，我们提出了一种新颖的离散均值流公式，采用简洁有效的参数化设计，以支持离散图结构的高效生成。大量实验表明，EQUIMF在生成质量、物理有效性和采样效率方面均优于先前的扩散模型与流匹配方法。

Graph-structured data jointly contain discrete topology and continuous geometry, which poses fundamental challenges for generative modeling due to heterogeneous distributions, incompatible noise dynamics, and the need for equivariant inductive biases. Existing flow-matching approaches for graph generation typically decouple structure from geometry, lack synchronized cross-domain dynamics, and rely on iterative sampling, often resulting in physically inconsistent molecular conformations and slow sampling. To address these limitations, we propose Equivariant MeanFlow (EQUIMF), a unified SE(3)-equivariant generative framework that jointly models discrete and continuous components through synchronized MeanFlow dynamics. EQUIMF introduces a unified time bridge and average-velocity updates with mutual conditioning between structure and geometry, enabling efficient few-step generation while preserving physical consistency. Moreover, we develop a novel discrete MeanFlow formulation with a simple yet effective parameterization to support efficient generation over discrete graph structures. Extensive experiments demonstrate that EQUIMF consistently outperforms prior diffusion and flow-matching methods in generation quality, physical validity, and sampling efficiency.

评分理由: 论文核心研究多模态MoE模型中的路由分心现象，与’Mixture of Experts’高度相关（10分）。论文涉及视觉推理任务，与’Chain of Thought’和’System 2 Thinking’有一定关联（各5分）。论文分析模型内部机制，与’Mechanistic Interpretability’有一定关联（5分）。论文提到多模态模型，与’Large Language Models’有一定关联（5分）。其他关键词如SLMs、Scaling Laws、训练方法、对齐、推理加速等均未涉及，得0分。

多模态专家混合模型在视觉-语言任务上取得了显著性能。然而，我们发现了一个称为“见而不思”的困惑现象：模型能准确感知图像内容，却在后续推理中失败，而相同问题以纯文本形式呈现时却能正确解决。通过系统分析，我们首先验证了专家混合架构中存在跨模态语义共享，排除了语义对齐失败作为唯一解释。随后我们揭示，视觉专家与领域专家呈现层级分离现象，图像输入在领域专家集中的中间层会引发与文本输入显著不同的路由分岔。基于这些发现，我们提出“路由分心假说”：在处理视觉输入时，路由机制未能充分激活任务相关的推理专家。为验证该假说，我们设计了一种路由引导干预方法，用以增强领域专家激活。在三个多模态专家混合模型和六个基准测试上的实验表明，该方法带来了持续改进，在复杂视觉推理任务上最高可获得3.17%的性能提升。我们的分析进一步揭示，领域专家识别定位的是认知功能而非样本特定解决方案，这使得该方法能在不同信息结构的任务间实现有效迁移。

Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.

评分理由: 论文研究代理式多模态模型中的元认知工具使用问题，核心涉及LLM代理、工具使用、推理和自我改进。与以下关键词高度相关（10分）：LLMs（基础模型）、RLHF/DPO（使用HDPO框架优化）、CoT推理（解决推理问题）、系统2思维（深入推理）、自我改进（元认知优化）、LLM代理（代理式模型）、工具使用（工具调用优化）。其他关键词与论文内容无关（0分）。

智能体多模态模型的出现使系统能够主动与外部环境交互。然而，当前智能体存在严重的元认知缺陷：它们难以在利用内部知识与查询外部工具之间做出有效仲裁。因此，这些系统常常陷入盲目调用工具的困境——即使查询可直接从原始视觉上下文中解决，它们仍会机械性地执行工具调用。这种病态行为不仅导致严重的延迟瓶颈，还会引入额外噪声，干扰可靠的推理过程。现有的强化学习协议试图通过惩罚工具使用的标量化奖励来缓解此问题。然而，这种耦合式设计造成了不可调和的优化困境：过强的惩罚会抑制必要的工具使用，而过轻的惩罚在优势归一化过程中会被准确性奖励的方差完全淹没，从而无法遏制工具滥用。为突破这一瓶颈，我们提出HDPO框架，该框架将工具效率从竞争性标量目标重构为严格的条件性目标。通过摒弃奖励标量化，HDPO维持了两个正交的优化通道：一是最大化任务准确性的精度通道，二是通过条件优势估计在准确轨迹中强制实现执行经济性的效率通道。这种解耦架构自然形成了一种认知进阶机制——迫使智能体先掌握任务解决能力，再提升其自主决策能力。大量实验表明，我们最终构建的模型Metis在将工具调用量降低数个数量级的同时，显著提升了推理准确性。

The advent of agentic multimodal models has empowered systems to actively interact with external environments. However, current agents suffer from a profound meta-cognitive deficit: they struggle to arbitrate between leveraging internal knowledge and querying external utilities. Consequently, they frequently fall prey to blind tool invocation, resorting to reflexive tool execution even when queries are resolvable from the raw visual context. This pathological behavior precipitates severe latency bottlenecks and injects extraneous noise that derails sound reasoning. Existing reinforcement learning protocols attempt to mitigate this via a scalarized reward that penalizes tool usage. Yet, this coupled formulation creates an irreconcilable optimization dilemma: an aggressive penalty suppresses essential tool use, whereas a mild penalty is entirely subsumed by the variance of the accuracy reward during advantage normalization, rendering it impotent against tool overuse. To transcend this bottleneck, we propose HDPO, a framework that reframes tool efficiency from a competing scalar objective to a strictly conditional one. By eschewing reward scalarization, HDPO maintains two orthogonal optimization channels: an accuracy channel that maximizes task correctness, and an efficiency channel that enforces execution economy exclusively within accurate trajectories via conditional advantage estimation. This decoupled architecture naturally induces a cognitive curriculum-compelling the agent to first master task resolution before refining its self-reliance. Extensive evaluations demonstrate that our resulting model, Metis, reduces tool invocations by orders of magnitude while simultaneously elevating reasoning accuracy.

评分理由: 论文《SIM1: Physics-Aligned Simulator as Zero-Shot Data Scaler in Deformable Worlds》专注于机器人操作中可变形物体的物理对齐仿真，核心贡献是提出一个从真实到仿真再到真实的数据引擎，用于生成高质量的合成数据以训练策略。论文涉及机器人学、物理仿真、数据生成和策略学习，但未涉及任何大语言模型（LLM）、深度学习技术原理或AI for Science的具体应用（如生物信息学、化学信息学）。所有评分关键词均与大语言模型、深度学习技术或特定科学AI应用相关，而本文研究的是物理仿真和机器人策略学习，属于不同的研究领域，因此所有关键词的相关度均为0。

机器人对可变形物体的操作体现了具身学习中的数据密集型特性，其形状、接触与拓扑结构的协同演化程度远超刚体。尽管仿真有望缓解现实世界数据采集的成本压力，但主流的仿真到现实流程仍基于刚体抽象，导致几何失配、软体动力学脆弱，以及难以适用于布料交互的运动基元。我们认为仿真失效的原因并非其合成属性，而在于其缺乏现实依据。为此，我们提出SIM1——一个基于物理对齐的现实-仿真-现实数据引擎，将仿真建立在物理世界基础上。在有限示教数据条件下，该系统将场景数字化为度量一致的双胞胎模型，通过弹性建模校准可变形物体动力学，并借助基于扩散的轨迹生成与质量过滤机制扩展行为空间。该流程将稀疏观测转化为具有接近示教保真度的规模化合成监督数据。实验表明，仅使用合成数据训练的策略在1:15的等效比例下达到与真实数据基线相当的性能，并在现实部署中实现90%的零样本成功率与50%的泛化性能提升。这些结果验证了物理对齐仿真可作为可变形物体操作的可扩展监督方法，并为数据高效策略学习提供了可行路径。

Robotic manipulation with deformable objects represents a data-intensive regime in embodied learning, where shape, contact, and topology co-evolve in ways that far exceed the variability of rigids. Although simulation promises relief from the cost of real-world data acquisition, prevailing sim-to-real pipelines remain rooted in rigid-body abstractions, producing mismatched geometry, fragile soft dynamics, and motion primitives poorly suited for cloth interaction. We posit that simulation fails not for being synthetic, but for being ungrounded. To address this, we introduce SIM1, a physics-aligned real-to-sim-to-real data engine that grounds simulation in the physical world. Given limited demonstrations, the system digitizes scenes into metric-consistent twins, calibrates deformable dynamics through elastic modeling, and expands behaviors via diffusion-based trajectory generation with quality filtering. This pipeline transforms sparse observations into scaled synthetic supervision with near-demonstration fidelity. Experiments show that policies trained on purely synthetic data achieve parity with real-data baselines at a 1:15 equivalence ratio, while delivering 90% zero-shot success and 50% generalization gains in real-world deployment. These results validate physics-aligned simulation as scalable supervision for deformable manipulation and a practical pathway for data-efficient policy learning.

评分理由: 论文主要研究文本到音视频生成的评估基准，核心贡献是AVGen-Bench基准和多粒度评估框架。论文明确提到使用Multimodal Large Language Models (MLLMs)进行评估，因此与’Large Language Models’关键词高度相关（8分）。其他关键词主要涉及大模型技术原理、训练方法、推理优化、代理系统等，论文未涉及这些具体技术，因此评分为0分。论文属于大模型在多媒体生成评估领域的应用，符合研究背景中’大模型在不同领域的研究应用’的要求。

文本到音视频生成正迅速成为媒体创作的核心界面，但其评估体系仍处于碎片化状态。现有基准大多孤立评估音频与视频，或依赖粗糙的嵌入相似度度量，难以捕捉现实提示词所要求的细粒度联合正确性。我们推出AVGen-Bench——一个面向T2AV生成的任务驱动型基准，涵盖11个真实场景类别的高质量提示词。为支持全面评估，我们提出多粒度评估框架，将轻量级专家模型与多模态大语言模型相结合，实现从感知质量到细粒度语义可控性的系统评测。我们的评估揭示了当前系统在强视听美学表现与弱语义可靠性之间存在显著差距，包括文本渲染、语音连贯性、物理推理方面的持续缺陷，以及在音乐音高控制上的普遍失效。代码与基准资源已发布于http://aka.ms/avgenbench。

Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench.

评分理由: 论文RewardFlow专注于图像生成和编辑领域，提出了一种基于多奖励Langevin动力学的推理时引导框架，用于优化预训练的扩散和流匹配模型。虽然论文涉及深度学习技术（如扩散模型、VQA奖励），但所有给定的关键词均与大语言模型（LLM）相关，而本文核心是计算机视觉中的图像生成，未涉及LLM技术、训练方法、推理优化、对齐、代理系统或科学AI应用。因此，所有关键词得分为0。

我们提出RewardFlow，一种无需逆向优化的框架，通过多奖励朗之万动力学在推理阶段引导预训练的扩散模型与流匹配模型。该框架统一了语义对齐、感知保真度、局部定位、对象一致性与人类偏好等互补的可微分奖励机制，并进一步引入基于可微分视觉问答（VQA）的奖励函数，通过语言-视觉推理提供细粒度语义监督。为协调这些异构目标，我们设计了提示感知自适应策略：该策略从指令中提取语义基元，推断编辑意图，并在整个采样过程中动态调整奖励权重与步长。在多项图像编辑与组合生成基准测试中，RewardFlow在编辑保真度与组合对齐方面实现了最先进的性能。

We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at inference time through multi-reward Langevin dynamics. RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, we design a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. Across several image editing and compositional generation benchmarks, RewardFlow delivers state-of-the-art edit fidelity and compositional alignment.

评分理由: 论文主要研究多模态大语言模型的强化学习训练方法（G²RPO）和感知-推理平衡机制，与以下关键词高度相关：1）‘Large Language Models’（8分）- 论文研究多模态大语言模型；2）‘RLHF’（8分）- 论文提出新的强化学习训练目标G²RPO；3）‘Chain of Thought’（10分）- 论文明确提到’extended reasoning chains’和’multi-step reasoning’；4）‘System 2 Thinking’（10分）- 论文强调深度推理能力。其他关键词如MoE、量化、RAG等未在论文中涉及，得0分。

群体相对策略优化（GRPO）已成为推动多模态大语言模型近期进展的事实性强化学习目标。然而，将其成功扩展至开源多模态通用模型仍主要受限于两大挑战：不同视觉任务间奖励拓扑结构的极端差异性，以及平衡细粒度感知与多步推理能力的内在困难。为解决这些问题，我们提出了高斯群体相对策略优化（G$^2$RPO），这是一种新颖的强化学习训练目标，它用非线性分布匹配替代了标准的线性缩放。通过数学上强制任何给定任务的优势分布严格收敛于标准正态分布 $\mathcal{N}(0,1)$，G$^2$RPO 从理论上确保了任务间的梯度公平性，减轻了对重尾异常值的敏感性，并为正负奖励提供了对称的更新机制。借助 G$^2$RPO 提供的增强训练稳定性，我们引入了两种任务级塑形机制，以无缝平衡感知与推理。首先，响应长度塑形动态地激发针对复杂查询的扩展推理链，同时强制直接输出以增强视觉基础。其次，熵塑形严格限制模型的探索区域，有效防止熵崩溃与熵爆炸。综合这些方法，我们提出了 OpenVLThinkerV2，一个高度鲁棒的通用多模态模型。在 18 个多样化基准上的广泛评估表明，其性能优于强大的开源模型及领先的专有前沿模型。

Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in Multimodal Large Language Models. However, extending this success to open-source multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and reasoning. First, response length shaping dynamically elicits extended reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second, entropy shaping tightly bounds the model’s exploration zone, effectively preventing both entropy collapse and entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose multimodal model. Extensive evaluations across 18 diverse benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models.