个体级移动性预测在城市模拟、交通规划及政策分析中处于核心地位。监督序列模型虽能实现高准确性，但需要针对特定任务进行训练，且决策层面的透明度有限。近期基于大语言模型（LLM）的方法提升了可解释性，但大多依赖静态提示和单次推理，当移动信号微弱或冲突时，限制了其寻求额外证据的能力。本文提出 \method{}（AgentMob），一种无需训练的 LLM 驱动代理框架，将下一位置预测形式化为自适应证据控制的决策制定。\method{} 基于历史规律，通过快速路径解决常规案例；而对于模糊案例，则触发针对近期轨迹、历史行为、停留 - 移动可能性及地理证据的迭代工具调用。在三个移动性数据集上，AgentMob 在无训练的基于 LLM 的方法中取得了最强的整体性能，其中 GPT-5.4 在 BW 数据集上 Acc@1 达到 71.42%，在 YJMob100K 上为 33.14%，在 Shanghai ISP 上为 33.50%。在 BW 数据集的非快速路径案例上，LLM 控制器相比同工具统计基线，将 Acc@1 从 30.65% 提升至 48.62%，这表明其主要优势在于通过自适应证据收集来解决模糊预测。我们的代码可在 https://github.com/Unknown-zoo/AgentMob 获取。

Individual-level mobility prediction is central to urban simulation, transportation planning, and policy analysis. Supervised sequence models achieve strong accuracy but require task-specific training and offer limited decision-level transparency. Recent LLM-based methods improve interpretability, yet mostly rely on static prompts and single-pass inference, limiting their ability to seek additional evidence when mobility signals are weak or conflicting. We propose \method{}, a training-free LLM-driven agent framework that formulates next-location prediction as adaptive evidence-controlled decision making. \method{} resolves routine cases through a fast path based on historical regularity, while ambiguous cases trigger iterative tool use over recent trajectories, historical behavior, stay-move likelihood, and geographical evidence. Across three mobility datasets, AgentMob achieves the strongest overall performance among training-free LLM-based methods, with GPT-5.4 reaching 71.42\% Acc@1 on BW, 33.14\% on YJMob100K, and 33.50\% on Shanghai ISP. On BW non-fast-path cases, the LLM controller improves Acc@1 from 30.65\% to 48.62\% over a same-tool statistical baseline, showing that its main benefit lies in resolving ambiguous predictions through adaptive evidence gathering. Our code is available at https://github.com/Unknown-zoo/AgentMob.

评分理由: 论文提出了一种基于 LLM 驱动的证据引导移动性预测框架（AgentMob），核心在于自适应证据收集与决策，而非模型架构的统一或表征学习。未涉及 Tokenizer 设计、视觉编码器架构或基于模型的强化学习训练过程。虽然使用 LLM，但未明确强调多模态融合或世界模型构建，因此与给定技术关键词的相关性较低至中等。作者列表中未包含指定的专家，无额外加分。

Mobility Prediction, LLM-Driven Agent, Evidence-Grounded, Training-Free, Adaptive Decision Making, Tool Use, Next-Location Prediction

主动推断将决策视为推断，其中预期自由能 (EFE) 统一了目标导向行为与信息寻求行为。近期研究表明，预期自由能 (EFE) 最小化可表述为在增加了认识先验的生成模型上的变分自由能 (VFE) 最小化。我们证明，增强模型的变分自由能 (VFE) 可重写为预测模型的变分自由能 (VFE) 加上显式熵修正项，从而使预期自由能 (EFE) 的贡献变得清晰。随后我们表明，恰当的基于预期自由能 (EFE) 的规划需将这些认识修正与一个规划修正相结合，该修正将边缘推断转化为策略优化，从而得到基于预期自由能 (EFE) 规划的完整变分表征。这明确了交叉熵规划与完整的基于预期自由能 (EFE) 规划分别需要的修正。同样的熵修正公式导出了基于预期自由能 (EFE) 规划的详细消息传递方案，以及更简化的变体。在三个网格世界环境中的实验表明，当观测具有决定性时，规划修正即可提供帮助；而当观测仅为提示性时，额外的观测侧认识修正则最为关键。

Active inference casts decision-making as inference, with the Expected Free Energy (EFE) unifying goal-directed and information-seeking behavior. Recent work showed that EFE minimization can be written as Variational Free Energy (VFE) minimization on a generative model augmented with epistemic priors. We prove that the VFE of the augmented model can be rewritten as the VFE of the predictive model plus explicit entropy-correction terms, making the EFE contribution transparent. We then show that proper EFE-based planning requires combining these epistemic corrections with a planning correction that turns marginal inference into policy optimization, yielding a full variational characterization of EFE-based planning. This clarifies which corrections are needed for cross-entropy planning and for full EFE-based planning. The same entropy-corrected formulation leads to a detailed message-passing scheme for EFE-based planning together with simpler ablations. Experiments on three grid-world environments show that the planning correction already helps when observations are decisive, whereas the additional observation-side epistemic corrections matter most when observations are merely suggestive.

评分理由: 论文聚焦主动推断（Active Inference）的理论推导，涉及生成模型与策略优化。model-based RL 相关性高（7.0），因主动推断基于生成模型进行规划；World Models 中度相关（4.0），因涉及世界生成模型；Unify Models 低度相关（3.0），仅统一行为目标而非模型架构；其余关键词（Tokenizer, Visual Encoder, MLLM, MultiModal）涉及多模态大模型具体组件，与本文理论内容无关（0.0）。作者列表未包含指定专家，无加分。加权总分 22.5，低于动态及格分 27.8。

Active Inference, Expected Free Energy, Variational Free Energy, Generative Model, Epistemic Priors, Policy Optimization, Planning Correction

大语言模型（LLMs）面临捷径学习问题：它们在分布外（OOD）输入上系统性地失效，尽管逻辑结构相同，但这些输入的语义表层与训练数据存在差异。这破坏了将思维链推理（CoT）知识蒸馏至较小学生模型（Student Models）的流程。我们提出不变梯度对齐（Invariant Gradient Alignment, IGA），这是一种训练框架，通过三项创新，在语义多样但逻辑同构的示例之间对齐梯度更新：(i) 逻辑同构集（Logical Isomer Sets），指在不同语义领域（数学、医学、法律、科学）中共享相同逻辑结构的问题组；(ii) 可微的“连续梯度冲突掩码”（Continuous Gradient Conflict Mask），用于抑制跨域梯度方差较高的参数维度，同时保留不变方向；(iii) 将掩码后的梯度通过截断奇异值分解（SVD）投影回 LoRA 低秩流形，从而在整个过程中保持参数效率。理论上，IGA 比经验风险最小化（ERM）具有更紧的分布外（OOD）泛化界，该界随异构体领域的数量缩放，且在温和正则性条件下以标准随机梯度下降（SGD）速率收敛。实验上，IGA 在四个基准测试上优于八个基线方法，相比经验风险最小化微调（ERM-SFT）准确率提升高达 14.3 个百分点，且逻辑一致性得分为 0.031（对比 0.142）——实现了表征不变性的四倍提升。

Large language models (LLMs) suffer from shortcut learning: they systematically fail on out-of-distribution (OOD) inputs whose semantic surface differs from training data, even when the logical structure is identical. This undermines knowledge distillation pipelines that transfer chain-of-thought reasoning to smaller students. We introduce Invariant Gradient Alignment (IGA), a training framework that aligns gradient updates across semantically diverse but logically isomorphic examples via three innovations: (i) Logical Isomer Sets, groups of problems sharing identical logical structure across distinct semantic domains (mathematics, medicine, law, science); (ii) a differentiable \emph{Continuous Gradient Conflict Mask}, that suppresses parameter dimensions with high cross-domain gradient variance while preserving invariant directions; and (iii) a truncated SVD projection of the masked gradient back onto the LoRA low-rank manifold, maintaining parameter efficiency throughout. Theoretically, IGA yields tighter OOD generalization bounds than ERM, scaling with the number of isomer domains, and converges at the standard SGD rate under mild regularity. Empirically, IGA outperforms eight baselines across four benchmarks with accuracy gains up to 14.3 pp over ERM-SFT and a Logical Consistency Score of 0.031 versus 0.142 -- a fourfold improvement in representational invariance.

评分理由: The paper focuses on LLM reasoning distillation and OOD generalization via gradient alignment, showing low relevance to Multimodal components (Visual Encoder, MultiModal), World Models, and Model-based RL. It shares moderate relevance with MLLM as an LLM study but lacks multimodal specifics. Tokenizer and Unify Models are not central themes. No expert authors from the specified list were found, so no bonus points were added.

Invariant Gradient Alignment, Reasoning Distillation, Out-of-Distribution Generalization, Logical Isomer Sets, Gradient Conflict Mask, LoRA, Large Language Models

我们介绍了 IRIS-GAN，这是一种针对跨生成器偏移下合成人脸图像的专业法医检测器。与解决通用合成图像检测不同，我们专注于由生成对抗网络（GANs）生成的人脸，这些网络在深度伪造内容中处于最先进水平，并通过分阶段暴露于要求日益提高的 GAN 家族（同时保留早期生成器）来训练该检测器。最终模型在考虑的 GAN 家族上实现了超过 99% 的伪造检测率，并在外部真实人脸数据集上实现了 98.9% 的分类准确率。Grad-CAM 分析进一步揭示了可测量的生成器依赖的空间响应模式，这些模式对于仅基于热力图的分类器仍具有信息量。在扩散模型生成的人脸上的家族外测试证实了 IRIS-GAN 是一种专业检测器，具备一定的检测非 GAN 深度伪造的能力。这些结果确立了分阶段训练作为一种稳健的 GAN 人脸法医分析的有效策略。

We introduce IRIS-GAN, a specialist forensic detector for synthetic face images under cross-generator shift. Rather than addressing universal synthetic-image detection, we focus on faces generated by generative adversarial networks (GANs), which are state-of-the-art in deepfake content, and train the detector through staged exposure to increasingly demanding GAN families while retaining earlier generators. The final model reaches fake-detection rates above 99% across the GAN families considered and classifies an external real-face dataset with 98.9% accuracy. Grad-CAM analysis further reveals measurable generator-dependent spatial response patterns, which remain informative for a secondary heatmap-only classifier. Out-of-family tests on diffusion-generated faces confirm that IRIS-GAN is a specialist detector, with some capability to reach non-GAN deepfakes. These results establish staged training as an effective strategy for robust GAN-face forensics.

评分理由: 论文主要关注深度伪造人脸检测（Deepfake Detection）及生成对抗网络（GAN）家族的分类，未涉及分词器（Tokenizer）、世界模型（World Models）、多模态大语言模型（MLLM）或基于模型的强化学习（model-based RL）。虽然图像分析隐含使用了视觉编码器（Visual Encoder），但整体内容与统一模型（Unify Models）或多模态（MultiModal）的核心定义关联较弱，因此大部分关键词评分较低。

Deepfake Detection, GAN Families, Staged Training, Specialist Detector, Synthetic Faces, Forensic Analysis, Grad-CAM, Cross-Generator Shift

近期基于生成模型的多视图图像编辑进展，使我们离通用 3D 内容生成和定制更近了一步。大多数现有工作利用未编辑场景的几何结构，专注于刚性或仅外观的编辑。这自然限制了这些方法仅适用于保留底层场景结构的编辑。其他方法则是针对特定图像编辑任务（如物体移除和添加）进行训练的。尽管取得了这些进展，通用非刚性编辑（即显著改变场景几何的编辑）对现有方法而言仍然具有挑战性。我们提出 GeM-NR，一种快速灵活的无需训练的通用多视图一致图像编辑方法，涵盖显著改变场景几何和外观的编辑。给定一个使用选定骨干编辑器 (backbone editor)（如 FLUX、Qwen、BrushNet）编辑过的锚图像 (anchor image) 和一个查询未编辑图像 (query image)，GeM-NR 能够根据锚编辑一致地编辑查询图像。该方法包含多个阶段：(i) 深度图 (depth map) 估计，其中我们提出一种策略以最大化编辑场景与未编辑场景之间 3D 点云 (3D point clouds) 的对齐度；(ii) 投影到查询视角；以及 (iii) 基于未编辑查询图像对所获图像进行细化。这种基于条件的表述在从两个视图扩展到多个物体视图时具有良好的可扩展性。我们展示了该方法处理几何和外观发生显著变化的编辑的能力，而这正是现有方法所难以应对的。我们进行了广泛的评估，结果表明该方法在多种编辑任务中提升了包括生成编辑场景 3D 表示在内的一致性。定性和定量结果表明，该方法在编辑质量以及多视图下的几何和光度 (photometric) 一致性方面均达到了最先进的性能。

Recent developments in multi-view image editing with generative models have brought us a step closer toward general 3D content generation and customization. Most existing works focus on rigid or appearance-only edits by utilizing the geometry of the unedited scene. This naturally limits these methods to edits that preserve the underlying scene structure. Other approaches are trained for specific image editing tasks, such as object removal and addition. Despite this progress, general nonrigid edits, i.e., edits that substantially change the scene geometry, remain challenging for existing methods. We propose GeM-NR, a fast and flexible training-free approach for general multi-view consistent image editing, including edits that drastically change the geometry and appearance of the scene. Given an anchor image edited with a chosen backbone editor (such as FLUX, Qwen, BrushNet) and a query unedited image, GeM-NR edits the query image consistently with the anchor edit. The method incorporates multiple stages: (i) depth map estimation, where we propose a strategy to maximize the alignment between the 3D point clouds of the edited and unedited scenes, (ii) projection onto a query viewpoint, and (iii) refinement of the obtained image conditioned on the unedited query. The conditioning-based formulation scales well from two to many views of an object. We demonstrate the ability of our method to handle edits with significant changes in geometry and appearance, something that existing methods struggle with. We perform an extensive evaluation showing that our method improves consistency for a wide variety of edit tasks, including generating 3D representations of the edited scene. Both quantitative and qualitative results indicate the state-of-the-art performance of our method in terms of edit quality as well as geometric and photometric consistency across multiple views.

评分理由: The paper focuses on geometry-aware multi-view image editing for nonrigid scene changes using generative models. It utilizes existing tools (e.g., Qwen, FLUX) but does not propose unified model architectures, tokenizers, world models, or reinforcement learning frameworks. Moderate relevance is assigned to Visual Encoder and MultiModal due to vision and geometry processing, while other keywords are largely irrelevant to the core contribution. No target expert authors (Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang) were found in the author list.

Multi-View Editing, Nonrigid Scene Changes, Geometry-Aware, Generative Models, 3D Content Generation, Depth Map Estimation, Image Consistency

基于评分标准的强化学习（RL）利用大语言模型作为裁判（LaaJ），依据评分标准对模型输出进行评分，并将其作为奖励。然而，策略模型可能会利用裁判中的潜在偏见，导致奖励黑客行为（reward hacking），从而产生无效或不安全的训练结果。在实际的基于评分标准的强化学习中，此类黑客行为往往较为微妙，且与多种裁判偏见相互纠缠，导致难以分析、检测和缓解。本文介绍了一种名为 CHERRL 的可控黑客环境，专门用于基于评分标准的强化学习。通过向 LaaJ 注入已知偏见，CHERRL 能够实现奖励黑客行为的重现、奖励分歧的明确观察以及黑客起始点的精确识别。这为研究基于评分标准的强化学习中奖励黑客的机制及缓解方法提供了一个清晰的实验测试平台。为了展示其效用，我们从可发现性和可利用性的角度分析了不同的裁判偏见，并探索了一种基于智能体的系统，用于从训练日志中自动检测奖励黑客行为的起始点。代码及环境已在 https://github.com/THUAIS-Lab/CHERRL 上公开提供。

Rubric-based reinforcement learning (RL) uses an LLM-as-a-Judge (LaaJ) to score model outputs according to rubrics as rewards. However, policy models may exploit latent biases in the judge, leading to reward hacking and ineffective or unsafe training outcomes. In real-world rubric-based RL, such hacking behaviors are often subtle and entangled with multiple judge biases, making them difficult to analyze, detect, and mitigate. In this paper, we introduce CHERRL, a controllable hacking environment for rubric-based RL. By injecting known biases into LaaJ, CHERRL enables stable reproduction of reward hacking, explicit observation of reward divergence, and precise identification of hacking onset. This provides a clean experimental testbed for studying the mechanisms and mitigations of reward hacking in rubric-based RL. To demonstrate its utility, we analyze different judge biases from the perspectives of discoverability and exploitability, and explore an agent-based system for automatically detecting reward hacking onset from training logs. The code and environment are publicly available at https://github.com/THUAIS-Lab/CHERRL.

评分理由: 论文主要研究 Rubric-based RL 中的奖励黑客问题，使用 LLM 作为评判者。关键词中的 Unify Models、Tokenizer、Visual Encoder、World Models、MultiModal 与论文内容无直接关联，相关性极低（1 分）。MLLM 虽涉及 LLM 但未涉及多模态，相关性较低（3 分）。model-based RL 虽属于 RL 范畴，但论文侧重于奖励建模而非环境模型学习，相关性中等（3 分）。作者列表中未包含指定的专家，无加分。加权总分低于动态及格分，表明论文与给定关键词主题匹配度较低。

Rubric-based Reinforcement Learning, Reward Hacking, LLM-as-a-Judge, CHERRL, Judge Biases, Controllable Hacking Environment, Reward Divergence

车辆车身类型是超车事故中自行车骑行者受伤严重程度的重要决定因素，然而，目前公开文献中尚无能够从自然主义道路视频中自动将车辆分类为与伤害风险相关类别的工具。标准目标检测基准仅提供粗略的车辆标签（如汽车、卡车、公交车、摩托车），而现有的细粒度识别系统通常在受控图像数据上训练，且缺乏跨不同录制站点部署鲁棒性的评估。本文提出了一种开源的两阶段计算机视觉流水线，该流水线结合了预训练的 RT-DETR 检测器用于粗略车辆定位，以及微调的视觉 Transformer（ViT-Base/16）用于六类车身类型分类：乘用车、SUV、皮卡车、Minivan、大型货车和商用卡车。该流水线包含一个基于置信度的拒绝机制：当 softmax 输出低于 0.60 时，第二阶段将拒绝输出预测，从而生成未知标签，而非隐蔽的错误分类。在密歇根州安娜堡自行车道走廊的 3,805 个标注超车事件（分布内）上评估，该流水线实现了 0.94 的准确率，各类别 F1 分数范围从 0.91（Minivan）至 0.97（SUV）。在未重新训练的情况下，对来自开放骑行数据集的 311 个事件进行独立的分布外评估，准确率为 0.89。在域偏移下，四个样本充足的类别中有三个保持了 F1 分数在 0.90 或以上。性能下降最显著的是 Minivan（F1 = 0.72），这主要由拒绝率从 2.4% 上升至 25.0% 驱动，而非主动错误分类，表明该机制确实传播了真实的模型不确定性。完整的流水线，包括推理脚本、训练代码、评估工具及模型权重，均以开源软件形式发布，以支持路边视频档案和骑行安全研究中的可复现性与重用。

Vehicle body type is a significant determinant of cyclist injury severity in overtaking crashes, yet automated tools for classifying vehicles into injury-risk-relevant categories from naturalistic roadway video do not exist in the open literature. Standard object detection benchmarks provide only coarse vehicle labels (car, truck, bus, motorcycle), while existing fine-grained recognition systems are trained on controlled imagery and lack evaluation for deployment robustness across recording sites. This paper presents an open-source two-stage computer vision pipeline combining a pre-trained RT-DETR detector for coarse vehicle localization with a fine-tuned Vision Transformer (ViT-Base/16) for six-category body-type classification: passenger car, SUV, pickup truck, minivan, large van, and commercial truck. A confidence-based abstention mechanism withholds Stage 2 predictions when softmax output falls below 0.60, producing unknown labels rather than silent misclassifications. Evaluated on 3,805 annotated overtaking events from a bicycle-lane corridor in Ann Arbor, Michigan (in-distribution), the pipeline achieved 0.94 accuracy with per-class F1 scores from 0.91 (minivan) to 0.97 (SUV). On an independent out-of-distribution evaluation of 311 events from an open cycling dataset without retraining, accuracy was 0.89. Three of four well-represented categories maintained F1 at or above 0.90 under domain shift. The largest degradation was observed for minivan (F1 = 0.72), driven by abstention rate rising from 2.4% to 25.0% rather than active misclassification, consistent with the mechanism propagating genuine model uncertainty. The full pipeline, including inference scripts, training code, evaluation utilities, and model weights, is released as open-source software to support reproducibility and reuse across roadside video archives and cycling safety research.

评分理由: The paper focuses on fine-grained vehicle classification using Vision Transformers and RT-DETR, showing moderate relevance to 'Visual Encoder' (ViT usage) and weak relevance to 'Unify Models' (two-stage pipeline structure). It is largely unrelated to 'World Models', 'MLLM', 'MultiModal', 'model-based RL', and 'Tokenizer' (in the context of language/generative models). No expert authors from the specified list are present.

Vehicle Classification, Vision Transformers, Fine-grained Recognition, Two-Stage Pipeline, Roadway Video, Open-Source, RT-DETR, Confidence-based Abstention

选择合适的 3D 表示（3D representation）是一个基本的设计决策，它决定了现代计算机视觉与图形学管线（computer vision and graphics pipelines）在 3D 重建、新视图合成与渲染、形状与运动分析、识别及生成等任务中的效率、质量和能力。虽然传统表示（例如网格（meshes）、点云（point clouds）和体素网格（volumetric grids））仍然是 3D 传感器（例如激光雷达（LiDAR）和 3D 扫描仪）的标准输出，并在下游应用（例如编辑和模拟）中广泛使用，但近期基于神经和基元（primitive）的表示（例如 3D 高斯泼溅（3D Gaussian Splatting））提供了紧凑且可微的替代方案，在游戏、AR/VR、自动驾驶、机器人导航和医学成像等应用中开启了广泛的机会，仅举几例。本文旨在综述 3D 表示的主要类别，涵盖从离散显式格式到基于神经渲染（neural rendering）或基元泼溅（primitive splatting）的连续隐式场（continuous implicit fields）。对于每种表示类型，我们阐述其一般形式化及其变体，讨论其优势与局限性，并重点阐述关键应用。本文最后概述了开放挑战及未来研究的潜在方向。区别于近期广泛覆盖 3D 物体与场景重建的综述，本文聚焦于 3D 表示本身演化的分析。我们特别强调向隐式表示（implicit representations）的范式转变，提供了一种新颖的视角，阐释这些新兴格式如何从根本上改变 3D/4D 工作流程。

The selection of an appropriate 3D representation is a fundamental design decision that dictates the efficiency, quality, and capabilities of modern computer vision and graphics pipelines for tasks such as 3D reconstruction, novel-view synthesis and rendering, shape and motion analysis, recognition, and generation. While traditional representations (\eg meshes, point clouds, and volumetric grids) remain standard outputs of 3D sensors (\eg LiDAR and 3D scanners) and are widely used in downstream applications (\eg editing and simulation), recent neural and primitive-based representations (\eg 3D Gaussian Splatting) offer compact and differentiable alternatives opening a wide range of opportunities in applications such as games, AR/VR, autonomous driving, robot navigation, and medical imaging, to name a few. The goal of this paper is to survey the main families of 3D representations from discrete explicit formats to continuous implicit fields based either on neural rendering or primitive splatting. For each type of representation, we present the general formulation and its variants, discuss its benefits and limitations, and highlight key applications. We conclude the paper by outlining the open challenges and potential directions for future research. Distinct from recent surveys that broadly cover 3D object and scene reconstruction, this paper provides a focused analysis on the evolution of 3D representations themselves. We specifically emphasize the paradigm shift toward implicit representations, offering a novel perspective on how these emerging formats fundamentally alter 3D/4D workflows.

评分理由: The paper is a survey on learning-based 3D geometric representations (meshes, point clouds, implicit fields, Gaussian Splatting) for computer vision and graphics. It does not focus on Multimodal Large Language Models (MLLM), tokenizers, or model-based reinforcement learning algorithms. While 3D representations are components in world models and visual encoders, the core contribution is a taxonomy of geometry rather than multimodal learning or control systems. Therefore, relevance to the provided keywords is low, resulting in a weighted score below the dynamic passing threshold.

3D Representations, Learning-based, Implicit Fields, Neural Rendering, Primitive Splatting, Computer Vision, Graphics Pipelines

大语言模型 (LLM) 的后训练过程经历多个阶段，例如监督微调 (SFT) 之后是基于人类反馈的强化学习 (RLHF) 或直接偏好优化 (DPO)，每个阶段的数据均来源于不同且可能不可信的来源。现有文献假设数据投毒攻击可能发生在每个训练阶段，却忽略了存在多个攻击者的可能性。为了研究整个后训练流程的可信度，我们提出了顺序数据投毒的威胁模型，其中多个攻击者分别对 SFT 数据集和偏好数据集进行投毒。在此威胁模型下，我们发现了一种“单攻击者幻觉”：每个攻击者若单独评估，似乎构成的威胁微不足道。然而，当攻击者在不同阶段之间协作时，真正的漏洞才会显现。在 SFT → DPO 流程中，攻击者的贡献是可加的：将固定的投毒预算分散到各个阶段优于将其集中在任一阶段。在 SFT → 近端策略优化 (PPO) 流程中，攻击者的贡献是互补的：单独的 SFT 投毒或奖励模型投毒均无法成功，但两者的结合却能奏效。这些发现表明，对单个后训练阶段的安全分析系统性地低估了仅从其交互中才会出现的复合漏洞。代码可在 https://github.com/jcksanderson/sequential-poisoning 获取。

LLM post-training proceeds through multiple stages, e.g., supervised fine-tuning (SFT) followed by reinforcement learning from human feedback (RLHF) or direct preference optimization (DPO), where each stage draws data from different, potentially untrusted sources. Existing literature assumes data poisoning attacks may occur at each training stage, but neglects the possibility of multiple attackers. To study the trustworthiness of the entire post-training pipeline, we propose the threat model of sequential data poisoning, where multiple adversaries separately poison the SFT and preference datasets. Under this threat model, we identify the single-attacker illusion: each adversary, evaluated in isolation, appears to pose a negligible threat. Yet when adversaries collaborate across stages, the true vulnerability is revealed. In the SFT $\to$ DPO pipeline, their contributions are additive: splitting a fixed poison budget across stages outperforms concentrating it in either stage alone. In the SFT $\to$ PPO pipeline, their contributions are complementary: neither SFT nor reward model poisoning succeeds individually, yet their combination does. These findings show that security analyses of individual post-training stages systematically underestimate compound vulnerabilities that emerge only from their interaction. Code is available at https://github.com/jcksanderson/sequential-poisoning.

评分理由: The paper focuses on security vulnerabilities (sequential data poisoning) in LLM post-training pipelines (SFT, DPO, PPO). It does not address multimodal architectures (Tokenizer, Visual Encoder, MultiModal, MLLM), world models, or model-based reinforcement learning algorithms. While RLHF/PPO involves reinforcement learning, it is model-free, not model-based. Unify Models is not the core theme. Thus, relevance to the provided keywords is low. Additionally, none of the listed expert authors (Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang) are present in the author list.

Sequential Data Poisoning, LLM Post-Training, SFT, DPO, PPO, Threat Model, Compound Vulnerabilities, Adversarial Collaboration

解剖标志点检测是医学图像分析中的基础任务，支持广泛的诊断与介入工作流程。尽管近期方法已实现亚毫米级定位，但仅凭准确性不足以支持临床部署，还需要预测具备可靠性和鲁棒性。尽管具有临床相关性，但在此背景下表征学习的影响仍未被充分探索。在本文中，我们引入了 CDPM-align，一种用于解剖标志点检测的多尺度引导对齐条件扩散预训练方法。我们的实验设置专注于少量图像和少量标注方案。具体来说，我们通过条件生成预训练，利用三个流行的异构小尺度基准数据集进行表征学习。此外，我们考虑了下游标志点检测任务中的低标注场景，分别使用 10 张和 25 张标注图像，反映了临床工作量与标注资源约束之间的现实权衡。我们的结果表明，生成式预训练使模型能够学习鲁棒的表征。这提高了下游任务的准确性和不确定性估计，朝着安全高效的临床部署迈进。

Anatomical landmark detection is a fundamental task in medical image analysis supporting a wide range of diagnostic and interventional workflows. Although recent methods have achieved sub-millimetric localisation, accuracy alone is not sufficient for clinical deployment, requiring reliability and robustness in prediction. Despite its clinical relevance, the impact of representation learning in this context is still underexplored. In this work, we introduce CDPM-align, a multi-scale guidance-aligned conditional diffusion pre-training for anatomical landmark detection. Our experimental setup focuses on a few images and a few annotation regimes. Specifically, we employ three popular heterogeneous small-scale benchmark datasets for representation learning via conditional generative pre-training. Furthermore, we consider low-annotation scenarios for the downstream task of landmark detection, with 10 and 25 annotated images, reflecting realistic trade-offs between clinical effort and resource constraints for annotations. Our results confirm that generative pre-training enables the model to learn a robust representation. This improves both accuracy and uncertainty on the downstream tasks, advancing towards safe and efficient clinical deployment.

评分理由: The paper focuses on medical image analysis using diffusion models for few-shot landmark detection. It does not address Unify Models, Tokenizers, World Models, MLLM, or Model-Based RL. Visual Encoder is implicitly part of the diffusion architecture but not a core contribution. Multi-modality is limited to image and annotations. Thus, relevance to the provided keyword list is low.

Anatomical Landmark Detection, Diffusion Pretraining, Few-Shot Learning, Robustness, Medical Image Analysis, Conditional Generative, Uncertainty Estimation

扩散语言模型（DLMs）通过迭代去噪生成文本，分块解码通过在局部块中确定词元提高了其实用性。然而，现有的分块方法通常依赖于固定块大小或基于分隔符的运行时信号，这些信号不一定与语义边界对齐。在本文中，我们提出 SemBlock，一种面向扩散大语言模型（LLMs）的语义边界驱动动态块解码框架。SemBlock 将动态块构建形式化为语义边界预测，并在冻结的 LLaDA 隐藏状态上训练轻量级预测器。为了提供监督，我们构建 SemBound，一个语义边界数据集，该数据集从自然语言、数学和代码任务的话语单元、推理步骤和实现片段中推导边界标签。在推理过程中，SemBlock 使用预测的边界概率来确定每个动态块的结束位置。在 GSM8K、IFEval、MATH 和 HumanEval 上的实验表明，SemBlock 一贯优于固定块解码和 AdaBlock。我们的代码已开源：https://github.com/TH-AI-Lab-PKU/SemBlock.

Diffusion language models (DLMs) generate text through iterative denoising, and blockwise decoding improves their practicality by committing tokens in local blocks. However, existing blockwise methods typically rely on fixed block sizes or delimiter-based runtime signals, which do not necessarily align with semantic boundaries. In this paper, we propose SemBlock, a semantic-boundary-driven dynamic block decoding framework for diffusion LLMs. SemBlock formulates dynamic block construction as semantic boundary prediction and trains lightweight predictors on frozen LLaDA hidden states. To provide supervision, we construct SemBound, a semantic-boundary dataset that derives boundary labels from discourse units, reasoning steps, and implementation spans across natural language, math, and code tasks. During inference, SemBlock uses predicted boundary probabilities to select the ending position of each dynamic block. Experiments on GSM8K, IFEval, MATH, and HumanEval show that SemBlock consistently improves over fixed-block decoding and AdaBlock. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SemBlock.

评分理由: 论文核心在于扩散语言模型（Diffusion LLMs）的解码策略优化，具体为语义边界动态块解码。与给定关键词高度相关的领域（如多模态、视觉编码器、世界模型、强化学习）在论文中均未涉及，因此相关性评分较低。Tokenizer 仅作为生成单元存在，非核心创新点；Unify Models 仅体现在解码逻辑的统一性上，关联较弱。

Diffusion Language Models, Blockwise Decoding, Semantic Boundaries, Dynamic Block, Text Generation, Denoising Process, LLaDA

长时程在线视觉映射是机器人感知的一项核心能力，需要在有界内存与计算的约束下，从视觉流中持续进行相机运动与场景几何估计。近期前馈 3D 重建模型提供了强大的几何先验，但其流式变体通常在与第一帧绑定的固定坐标系或持久场景记忆中预测位姿。这种固定规范设计会导致训练 - 测试不匹配、对早期锚点的注意力偏差，以及在远长于训练期间所见序列上的累积漂移。我们提出 Anchor3R，一种流式 3D 重建框架，该框架将前馈重建视为以当前为中心的局部测量预测，而非持久全局规范回归。在每个时间步，Anchor3R 在当前帧坐标系中预测窗口相对位姿和局部点图，从而将流式重建转化为相对位姿测量生成。这些测量支持在线位姿更新，而闭环重插入和运动平均则用于对齐轨迹，并将局部点图转换为连贯的全局重建。在室内、室外、驾驶和 RGB-D 基准测试上的实验表明，Anchor3R 相比现有流式基线提高了长时程位姿精度和密集重建质量，同时支持有界内存在线推理。

Long-horizon online visual mapping is a core capability for robot perception, requiring continuous camera-motion and scene-geometry estimation from visual streams under bounded memory and computation. Recent feed-forward 3D reconstruction models provide strong geometric priors, but their streaming variants often predict poses in a fixed coordinate system tied to the first frame or a persistent scene memory. This fixed-gauge design leads to train--test mismatch, attention bias toward early anchors, and accumulated drift on sequences much longer than those seen during training. We propose \emph{Anchor3R}, a streaming 3D reconstruction framework that treats feed-forward reconstruction as current-centric local measurement prediction rather than persistent global-gauge regression. At each time step, Anchor3R predicts window-relative poses and a local pointmap in the current-frame coordinate system, turning streaming reconstruction into relative-pose measurement generation. These measurements support online pose updates, while loop-closure reinsertion and motion averaging align the trajectory and transform local pointmaps into a coherent global reconstruction. Experiments on indoor, outdoor, driving, and RGB-D benchmarks show that Anchor3R improves long-horizon pose accuracy and dense reconstruction quality over existing streaming baselines, while supporting bounded-memory online inference.

评分理由: 该论文专注于流式 3D 重建与长时程视觉映射，属于机器人感知与计算机视觉领域。关键词中 Unify Models、Tokenizer、MLLM、model-based RL 与论文内容无直接关联，评分为 0。Visual Encoder 虽隐含用于特征提取，但非核心创新点，评分为 5。World Models 涉及场景理解但非生成式世界模型，MultiModal 仅涉及 RGB-D 视觉输入，关联度较低，评分为 2。作者列表中未包含指定的 Yang Shi 等专家。加权总分 13.5，低于动态及格分 27.8。

Streaming 3D Reconstruction, Long-Horizon Visual Mapping, Transient Anchors, Relative-Pose Measurement, Loop-Closure Reinsertion, Motion Averaging, Pointmap Generation, Robot Perception

中医 (TCM) 眼部望诊为评估巩膜表面异常提供了经验性依据，但其临床应用仍具有主观性且难以量化。为支持智能化和可量化的眼部望诊，本研究提出了中医启发的人工智能眼部辅助诊断系统 (TAO)，并专注于像素级巩膜表面异常分割。针对受多源分布差异、多样异常形态及巩膜镜面反射 (SSR) 影响的临床及用户获取图像，我们提出了 HD-DinoMoE，一种类别感知层次化双混合专家网络。HD-DinoMoE 结合了类别感知双流 DINOv3 特征融合与类别特定多专家解码，以分割血管 (Vessels)、黄黑斑 (Yellow and Black Spots) 及血斑 (Blood Spots)。三阶段骨干冻结路由策略稳定了双骨干适应；渐进式置信度惩罚 (PCP) 损失减少了 SSR 区域中高置信度的假阳性及分割泄漏；类别感知自适应样本加权 (CA-ASW) 平衡了样本级与类别级的训练贡献。我们进一步构建了多标签巩膜异常分割数据集 (ML-SASD)，这是一个包含临床、野外 (Wild) 和混合设置的新基准，并对三种异常类别进行了像素级标注。在 ML-SASD-Mix 上，HD-DinoMoE 达到了 72.11% 的平均 Dice 系数和 58.44% 的平均交并比 (mIoU)，同时保持了良好的边界定位能力及镜面区域假阳性控制。它在公共 SBVPI 数据集的血管子集上也表现出具有竞争力的泛化性能。这些结果表明，HD-DinoMoE 为 TAO 在复杂采集场景下提供了一种可行的分割方案。代码和数据访问信息可在 https://github.com/FX-CMX/HD-DinoMoE 获取。

Traditional Chinese Medicine (TCM) ocular inspection provides empirical cues for assessing scleral surface anomalies, but its clinical use remains subjective and difficult to quantify. To support intelligent and quantifiable ocular inspection, this study presents the TCM-inspired Artificial Intelligence Ocular Auxiliary Diagnosis System (TAO) and focuses on pixel-level scleral surface anomaly segmentation. For clinical and user-acquired images affected by multi-source distributional discrepancies, diverse anomaly morphologies, and scleral specular reflection (SSR), we propose HD-DinoMoE, a class-aware hierarchical dual mixture-of-experts network. HD-DinoMoE combines class-aware dual-stream DINOv3 feature fusion with class-specific multi-expert decoding to segment Vessels, Yellow and Black Spots, and Blood Spots. A three-stage backbone-frozen routing strategy stabilizes dual-backbone adaptation; Progressive Confidence Penalty (PCP) Loss reduces high-confidence false positives and segmentation leakage in SSR regions; and Class-Aware Adaptive Sample Weighting (CA-ASW) balances sample- and class-level training contributions. We further construct the Multi-label Scleral Anomaly Segmentation Dataset (ML-SASD), a new benchmark with Clinical, Wild, and Mix settings and pixel-wise annotations for three anomaly categories. On ML-SASD-Mix, HD-DinoMoE achieves a mean Dice of 72.11% and a mean Intersection-over-Union of 58.44%, while maintaining favorable boundary localization and specular-region false-positive control. It also shows competitive generalization on the Vessels subset of the public SBVPI dataset. These results indicate that HD-DinoMoE provides a feasible segmentation solution for TAO under complex acquisition scenarios. The code and data access information are available at https://github.com/FX-CMX/HD-DinoMoE.

评分理由: The paper focuses on medical image segmentation using a Mixture-of-Experts network (HD-DinoMoE) and DINOv3 backbone. It shows high relevance to 'Visual Encoder' (DINOv3 is used as the feature extractor) and slight relevance to 'Unify Models' (dual-stream fusion mechanism). It is unrelated to 'Tokenizer', 'World Models', 'MLLM', 'MultiModal' (in the context of multimodal LLMs), and 'model-based RL', as the work is supervised computer vision rather than generative world modeling or reinforcement learning. The research background keywords target LLM/RL domains, while this paper is in medical AI.

HD-DinoMoE, Scleral Anomaly Segmentation, Mixture-of-Experts, DINOv3, Multi-label Segmentation, TCM Ocular Inspection, Dual Stream, Complex Acquisition Scenarios

当后训练语言模型（Post-trained Language Models）在推理问题上失败时，常见的测试时扩展（test-time-scaling）响应是投入更多计算资源进行额外尝试，而失败轨迹不再发挥进一步作用。我们认为这丢弃了一个关键信号：部分失败源于不幸采样（unlucky sampling），此时更多展开（rollouts）有所帮助；而另一些失败则是结构性的，无论预算如何，重采样都无法克服。我们提出，失败轨迹编码了可恢复性结构（recoverability structure）：即哪些测试时干预（test-time interventions）能够挽救特定失败的推理时特征（inference-time signature）。三个问题级轨迹特征（problem-level trajectory features）源自可用干预的结构，能够从失败展开的分布特征（distributional signature）中恢复此结构，而非基于其文本内容。它们将失败聚类为稳定模式（stable regimes），表征不同后训练方法的失败地形（failure topography）（准确率为 84.3±4.3%，比多数类基线（majority-class baseline）高出 20%），并支持一个无需训练的路由规则（training-free routing rule），在部署相关的 Steerable-Hard 子集（Steerable-Hard subset，即重试不足且有界干预可达的失败）上提升救援率 12.2%。这些特征和路由规则在两个跨模型族探针（cross-family probes）上均具有迁移性。因此，相同的三个特征将失败轨迹从被丢弃的数据转化为诊断对象，支持测试时路由（test-time routing）和后训练分析，且无需训练数据或权重空间访问。

When post-trained language models fail on reasoning problems, the common test-time-scaling response is to spend more compute on additional attempts, and the failed traces play no further role. We argue this discards a crucial signal; some failures come from unlucky sampling, where more rollouts help, while others are structural and resist resampling regardless of budget. We propose that failed traces encode recoverability structure: the inference-time signature of which test-time interventions can rescue a given failure. Three problem-level trajectory features, derived from the structure of available interventions, recover this structure from the distributional signature of failed rollouts, not their text. They cluster failures into stable regimes, characterize the failure topography of different post-training methods ($84.3{\pm}4.3\%$ accuracy, $+20\%$ over a majority-class baseline), and support a training-free routing rule that lifts rescue by $+12.2\%$ on the deployment-relevant Steerable-Hard subset (failures where retry is insufficient and a bounded intervention is reachable). The features and the routing rule transfer across two cross-family probes. The same three features thus convert failed traces from discarded data into a diagnostic object, supporting test-time routing and post-training analysis without training-time or weight-space access.

评分理由: 该论文专注于后训练语言模型的推理失败分析与测试时间干预策略，核心贡献在于利用轨迹特征区分采样错误与结构故障。提供的关键词集主要面向多模态大模型架构、世界模型及模型强化学习领域，与本文纯语言模型推理诊断的主题存在显著偏差。因此，除 MLLM 和 Unify Models 有微弱语义关联外，其余关键词（如 Visual Encoder, Tokenizer, MultiModal, model-based RL）相关性极低，导致加权总分（12.0）远低于动态及格分（27.8）。

Failed Reasoning Traces, Post-trained Language Models, Test-time Scaling, Intervention Routing, Failure Topography, Structural Failures, Unlucky Sampling

基于大语言模型（LLM）的多智能体系统正日益应用于战略决策任务。在此类场景中，性能不仅取决于单个模型的能力，还取决于智能体之间交互与适应的策略。多智能体强化学习（MARL）可以优化这些交互策略，但其奖励设计往往仍局限于特定任务，且与交互结构的结合较弱。为了解决这一差距，我们提出 GARL（博弈论强化学习框架），用于多智能体战略优先级排序。GARL 将战略优先级排序形式化为一个两阶段博弈：竞争智能体首先在一个共享候选集上分配战略资源，随后由高层仲裁者生成最终排名。由此产生的博弈论效用被转换为角色特定的强化信号，从而使策略优化能够受到结构化交互的引导。我们将 GARL 实例化于争议问题排名任务中，其目标是在法律程序中优先处理核心问题。实验表明，GARL 提升了排名性能，使小型开源大语言模型在相同的候选排名设置下能够与强大的闭源大语言模型相竞争，并在法律领域能力及更广泛的战略决策中取得了收益。总体而言，GARL 展示了如何将博弈论交互结构转化为强化学习目标，为多智能体战略优先级排序中的策略优化提供了一种原则性的方法。

LLM-based multi-agent systems are increasingly used for strategic decision-making tasks. In such settings, performance depends not only on individual model capabilities, but also on the policies by which agents interact and adapt. Multi-agent reinforcement learning can optimise these interaction policies, but its reward design often remains task-specific and weakly grounded in interaction structure. To address this gap, we propose GARL, a GAme-theoretic Reinforcement Learning framework for multi-agent strategic prioritisation. GARL formalises strategic prioritisation as a two-stage game: competing agents first allocate strategic resources over a shared candidate set, and a higher-level arbiter then produces the final ranking. The resulting game-theoretic utilities are converted into role-specific reinforcement signals, allowing policy optimisation to be guided by structured interaction. We instantiate GARL on issues-in-dispute ranking, where the goal is to prioritise core issues in legal proceedings. Experiments show that GARL improves ranking performance, enables small open-source LLMs to become competitive with a strong closed-source LLM under the same candidate-ranking setting, and yields gains in legal-domain competence and broader strategic decision-making. Overall, GARL demonstrates how game-theoretic interaction structure can be turned into reinforcement-learning objectives, providing a principled approach to policy optimisation in multi-agent strategic prioritisation.

评分理由: The paper focuses on Game-Theoretic Reinforcement Learning for multi-agent strategic prioritization in the legal domain. It does not address multimodal architectures, tokenizers, visual encoders, or world models. While it utilizes LLMs and reinforcement learning, it lacks the core components of the specified keywords (e.g., no visual components, no world model learning, no explicit model-based dynamics learning), resulting in low relevance scores.

Game-Theoretic Reinforcement Learning, Multi-Agent Systems, Strategic Prioritisation, LLM-based Multi-Agent Systems, Policy Optimisation, Legal Domain, Resource Allocation, Interaction Structure

视频全景分割（VPS）旨在联合检测、分割与跟踪所有对象，同时将视频划分为语义一致的区域。我们引入了无监督 VPS 的任务设定，摒弃了任何人工监督。现有的无监督场景理解工作主要集中于图像分割任务；视频领域仍探索不足。我们提出了 VideoCUPS，这是首个无监督 VPS 方法。VideoCUPS 通过利用无监督的深度、运动和视觉线索，从以场景为中心的视频中生成时间一致的视频全景伪标签。利用一种新颖的 Video DropLoss 在这些伪标签上进行训练，可获得一个准确的无监督 VPS 模型。为了评估进展，我们引入了一种全面的评估协议和四个竞争基线，将最先进的无监督全景图像分割和实例视频分割模型扩展至 VPS。VideoCUPS 优于所有基线，并展示了强大的标签高效学习能力。借助 VideoCUPS、我们的评估协议和基线，我们为未来的无监督 VPS 研究提供了坚实基础。

Video panoptic segmentation (VPS) aims to jointly detect, segment, and track all objects while partitioning the video into semantically consistent regions. We introduce the task setting of unsupervised VPS, omitting any human supervision. Existing unsupervised scene understanding works mainly focused on image segmentation tasks; the video domain remains underexplored. We propose VideoCUPS, the first unsupervised VPS approach. VideoCUPS generates temporally consistent panoptic video pseudo-labels from scene-centric videos by exploiting unsupervised depth, motion, and visual cues. Training on these pseudo-labels using a novel Video DropLoss yields an accurate, unsupervised VPS model. To benchmark progress, we introduce a comprehensive evaluation protocol and four competitive baselines, extending state-of-the-art unsupervised panoptic image and instance video segmentation models to VPS. VideoCUPS outperforms all baselines and demonstrates strong label-efficient learning. With VideoCUPS, our evaluation protocol, and baselines, we provide a strong foundation for future research on unsupervised VPS.

评分理由: The paper focuses on unsupervised video panoptic segmentation using depth and motion cues, which is a computer vision task unrelated to MLLM, Tokenizers, World Models, or RL. It shows weak alignment with 'Unify Models' (task unification) and 'MultiModal' (RGB+Depth+Motion). No expert authors from the specified list are present.

Video Panoptic Segmentation, Unsupervised Learning, Pseudo-labels, Depth Cues, Motion Cues, Scene-Centric, VideoCUPS

我们探究人类数学教学法的方法能否引导语言模型的训练，使其具备算术推理能力。基于 GASING 方法（一种印尼教学法，通过从左到右的过程解决基本算术，该过程与 token 生成的因果顺序一致），我们将每个运算操作化为一个计算过程，其执行轨迹被序列化为自然语言思维链（Chain-of-Thought，简称 CoT）监督。一个小型的 GPT-2 解码器（8600 万（86M）参数），配备用于印尼语的音节 - 黏着式 TOBA 分词器，仅使用下一个 token 预测目标，基于此数据从头开始训练，未使用强化学习（Reinforcement Learning）或基于奖励的优化。监控训练揭示了三个不同的学习阶段，而机制分析——包括对思维链信息图的注意力遮蔽干预、残差流探测（Residual-Stream Probing）以及 logit 透镜检查（Logit-Lens Inspection）——表明模型首先内化了一种程序路径，随后发展出一种关联式“心算”能力，能够在无需显式逐步计算的情况下检索中间结果。该训练模型在保留问题（Held-out Problems）上达到超过 80% 的准确率，并在性能上与规模大得多的语言模型相当，这表明基于教学法的目标训练能够在小规模下产生强大且经济高效的算术能力。

We investigate whether methods of human mathematics pedagogy can guide the training of language models toward arithmetic reasoning. Building on the GASING method -- an Indonesian pedagogy that solves basic arithmetic through a left-to-right procedure aligned with the causal order of token generation -- we operationalize each operation as a computational procedure whose execution trace is serialized into natural-language Chain-of-Thought (CoT) supervision. A small GPT-2 decoder (86M parameters) with a syllabic-agglutinative TOBA tokenizer for Indonesian is trained from scratch on this data using only a next-token prediction objective, without reinforcement learning or reward-based optimization. Monitoring training reveals three distinct learning phases, and mechanistic analyses -- attention-masking interventions on the CoT information graph, residual-stream probing, and logit-lens inspection -- show that the model first internalizes a procedural pathway and subsequently develops an associative, ``mental-arithmetic'' capacity that retrieves intermediate results without explicit step-by-step computation. The trained model reaches over 80% accuracy on held-out problems and attains competitive performance against substantially larger language models, indicating that targeted, pedagogically grounded training can yield strong and economical arithmetic capability at small scale.

评分理由: 该论文主要研究语言模型的算术推理训练方法，核心贡献在于教学法指导下的思维链监督生成。关于关键词相关性：1. Tokenizer 文中明确提到使用'TOBA tokenizer'，相关性中等（5 分）；2. Unify Models 虽涉及模型训练，但未涉及多模型架构统一，相关性较低（2 分）；3. Visual Encoder, World Models, MLLM, MultiModal, model-based RL 均与论文内容无关，论文明确排除了强化学习且无视觉或世界模型组件，相关性为 0 分。

Arithmetic Pedagogy, Language Models, Chain-of-Thought, Next-token Prediction, Syllabic Tokenizer, Mechanistic Analysis, Small-scale Training

基于大语言模型（LLM）的智能体越来越多地通过与外部工具、检索系统、内存模块、环境及其他智能体交互来解决复杂任务。这些能力扩展了智能体的自主性，但也使得智能体行为更难验证、调试和审计。仅凭最终答案准确性无法解释输出是如何产生的、哪些证据支持了每个主张、工具调用是否合理、内存如何影响后续决策，或者执行失败源自何处。证据追踪和执行溯源通过建模检索到的证据、工具输出、内存条目、环境观察、中间主张、动作和最终答案如何在整个智能体执行过程中相互连接，从而填补这一空白。本文综述提供了关于 LLM 智能体中证据追踪和执行溯源的系统性回顾和概念框架。我们围绕统一的溯源视角组织相关工作，该视角连接了检索锚定、主张支持、工具使用安全、内存谱系、可观察性、调试、审计和恢复。我们引入一个涵盖追踪来源、证据和执行单元、溯源关系、追踪粒度与时机、表示形式和信任功能的分类法。我们回顾了关键的方法论方向，包括溯源表示、证据归因、工具使用溯源、运行时护栏、承载溯源的内存、基于追踪的可观察性和故障诊断。我们还映射现有的基准、数据集和评估指标到与溯源相关的功能，并讨论评估如何从最终答案正确性转向过程级问责。最后，我们概述了开放挑战，包括统一追踪模式、主张级和语义溯源、感知溯源的安全机制、真实的执行追踪基准、面向恢复的评估以及隐私感知审计基础设施。

Large language model (LLM)-based agents increasingly solve complex tasks by interacting with external tools, retrieval systems, memory modules, environments, and other agents. These capabilities expand agent autonomy, but also make agent behavior harder to verify, debug, and audit. Final-answer accuracy alone cannot explain how an output was produced, which evidence supported each claim, whether tool calls were justified, how memory influenced later decisions, or where execution failures originated. Evidence tracing and execution provenance address this gap by modeling how retrieved evidence, tool outputs, memory items, environment observations, intermediate claims, actions, and final answers are connected throughout agent execution. This survey provides a systematic review and conceptual framework for evidence tracing and execution provenance in LLM agents. We organize related work around a unified provenance perspective that connects retrieval grounding, claim support, tool-use safety, memory lineage, observability, debugging, audit, and recovery. We introduce a taxonomy covering trace sources, evidence and execution units, provenance relations, tracing granularity and timing, representation forms, and trust functions. We review key methodological directions, including provenance representation, evidence attribution, tool-use provenance, runtime guardrails, provenance-bearing memory, trace-based observability, and failure diagnosis. We also map existing benchmarks, datasets, and evaluation metrics to provenance-related capabilities, and discuss how evaluation can move from final-answer correctness toward process-level accountability. Finally, we outline open challenges, including unified trace schemas, claim-level and semantic provenance, provenance-aware safety mechanisms, realistic execution-trace benchmarks, recovery-oriented evaluation, and privacy-aware audit infrastructure.

评分理由: 该论文聚焦于 LLM 代理的可解释性与审计性，核心贡献在于证据追踪和溯源框架。关键词中的 Tokenizer、Visual Encoder 涉及模型架构组件，文中未提及；World Models 和 model-based RL 通常指环境动力学建模与规划算法，本文仅涉及代理与环境交互的追踪，非模型构建；MLLM 和 MultiModal 强调多模态融合，本文主要基于 LLM；Unify Models 虽提到统一视角，但指概念统一而非模型架构统一。因此相关性普遍较低。未检测到指定专家作者。

LLM Agents, Evidence Tracing, Execution Provenance, Trust, Tool Use, Memory Lineage, Process-level Accountability

实时数据分析需要具备在非平稳数据流中准确且自适应地应对非线性动力学的能力，同时保持计算效率。然而，非线性动力学如此复杂，以至于在严格的时间约束下捕捉动态变化的非线性模式并将其用于下游任务并非易事。为了弥合非线性复杂性与计算可处理性之间的差距，本研究采用了 Koopman 算子理论 (Koopman operator theory)，该理论指出非线性动力学可表示为无限维空间中的线性变换。基于该算子的有限维近似，我们提出了 AdaKoop，一种用于在非平稳数据流上建模非线性动力学的高效流式算法。该方法基于 Koopman 算子理论构建了一个概率框架，将原始观测值和再生核希尔伯特空间 (RKHS) 特征均视为源自潜在向量的发射。这种双视角表述使得非线性动力学能够被表达为一个可处理的线性系统。因此，AdaKoop 能够在流式方式下高效且稳定地建模非线性动力学，避免了迭代非线性优化所带来的高昂计算成本。此外，为应对数据流中的非平稳性，AdaKoop 通过统计假设检验自适应地检测模式突变引起的切换，并增量更新模型参数以应对连续变化。在涵盖多个领域的总共 71 个实际基准数据集上的广泛实验表明，AdaKoop 在实时预测精度和计算效率方面均优于现有最先进方法。

Real-time data analysis requires the ability to accurately and adaptively address nonlinear dynamics in a nonstationary data stream while preserving computational efficiency. However, nonlinear dynamics are so complex that capturing dynamically changing nonlinear patterns and utilizing them for downstream tasks under strict time constraints is nontrivial. To bridge the gap between nonlinear complexity and computational tractability, this study applies Koopman operator theory, which states that nonlinear dynamics can be represented as linear transitions in an infinite-dimensional space. Building upon finite-dimensional approximations of this operator, we present AdaKoop, an efficient streaming algorithm for modeling nonlinear dynamics over nonstationary data streams. Our approach utilizes a probabilistic framework grounded in Koopman operator theory, treating both raw observations and reproducing kernel Hilbert space (RKHS) features as emissions from latent vectors. This dual-view formulation allows nonlinear dynamics to be expressed as a tractable linear system. Therefore, AdaKoop enables the efficient and stable modeling of nonlinear dynamics in a streaming fashion, avoiding the prohibitive computational costs of iterative nonlinear optimization. Furthermore, to address nonstationarity in data streams, AdaKoop adaptively detects the switching of patterns via statistical hypothesis testing for abrupt pattern shifts and incrementally updates model parameters to handle continuous changes. Extensive experiments on a total of 71 practical benchmark datasets across various domains demonstrate that AdaKoop outperforms state-of-the-art methods in terms of real-time forecasting accuracy and computational efficiency.

评分理由: 论文基于 Koopman 算子理论处理非线性动态流数据，属于控制理论与流式算法领域。关键词集（如 Tokenizer、Visual Encoder、MLLM）主要指向多模态大模型，与本文主题高度不匹配。虽然动力学建模与 World Models 和 Model-Based RL 有潜在关联，但本文未涉及生成式建模或强化学习，故相关性评分较低。

Koopman Operator, Nonlinear Dynamics, Nonstationary Data Streams, Streaming Algorithm, Probabilistic Framework, RKHS Features, Pattern Switching, Efficient Modeling

深度主动学习此前已被探索用于大语言模型（LLM）的上下文样本选择，但尚未采用利用 Transformer（变换器）激活理解最新进展的方法。本文测试了这样一个假设：模型激活可提供细粒度信号，以优化上下文示例的选择。本文对应用于上下文学习的基于 MLP（多层感知机）激活的深度主动学习方法进行了迄今为止最全面的分析，包括不同注意力掩码策略如何影响多样分类和生成数据集上的主动学习，实验使用了 Llama-3.2-3B 和 Qwen2.5-3B 基础模型。然而，我们发现了一个负面结果：通过大规模激活或前四阶矩视角观察的 MLP 输出，与示例质量或任务性能无相关性。具体而言，在我们测试的所有任务和模型中，斯皮尔曼相关系数的绝对值最高仅为 0.33，表明此类基于激活的采样不应被用于上下文学习。我们假设这可能是由于叠加现象所致，即模型表示的特征数量超过其维度，这表明稀疏自编码器（SAEs）等方法可能是一个有前途的未来方向。

Deep active learning has previously been explored for LLM in-context sample selection, but not with methods that utilise recent advances in understanding of transformer activations. In this paper, we test the hypothesis that model activations could provide a fine-grained signal to optimise the selection of in-context examples. We present the most comprehensive analysis to date of MLP activation-based deep active learning methods applied to in-context learning, including how different attention masking strategies impact active learning across diverse classification and generative datasets, using both Llama-3.2-3B and Qwen2.5-3B base models. However, we find a negative result: MLP outputs, viewed through the lenses of massive activations or the first four moments, do not correlate with example quality or task performance. Specifically, the absolute Spearman correlation coefficient is at most 0.33 for all tasks and models we tested, showing that such activation-based sampling should not be used for in-context learning. We hypothesise that this may be due to superposition, whereby models represent more features than they have dimensionality, suggesting that methods like Sparse Autoencoders (SAEs) may be a promising future direction.

评分理由: The paper focuses on activation-based active learning for in-context learning in LLMs. It shows low relevance to World Models, Visual Encoders, and Model-Based RL as these concepts are not discussed. While it uses LLMs (MLLM/MultiModal context), it does not address multimodal architecture or world modeling specifically. Tokenizer and Unify Models have minimal relevance. No expert authors from the specified list were found.

Activation-Based Active Learning, In-Context Learning, Transformer Activations, MLP Activation, Sample Selection, Sparse Autoencoders, Llama-3.2, Qwen2.5

音频深伪检测（ADD）模型在对抗文本到语音（TTS）模型的恶意使用方面至关重要。评估和加强 ADD 模型需要开发能够覆盖生成音频空间并突出高错误区域的数据集。现有的数据集开发策略面临两个挑战：（i）人工收集，以及（ii）低效地发现 ADD 模型中的盲点。为了解决这些挑战，我们提出了 FoeGlass，这是首个用于音频深伪检测（ADD）的黑盒自动化红队方法，它能够有效地在生成音频空间中发现 ADD 的失效模式，而这些空间未被最先进的深伪基准测试充分探索。FoeGlass 利用大语言模型（LLM）的上下文学习能力来探索文本到语音（TTS）模型的输入空间，生成能够欺骗目标音频深伪检测（ADD）模型的音频样本，且仅需对所有组件进行黑盒访问。通过基于多样性度量精心设计的上下文，FoeGlass 缓解了自动化红队系统中常见的模式坍塌问题。在多个开源音频深伪检测（ADD）和文本到语音（TTS）模型上的实证评估表明，基于 FoeGlass 生成的数据相较于无条件采样基线和近期的欺骗数据集，显著降低了假阴性率（最高达 94%），且无需人工监督。此外，我们还表明，FoeGlass 生成的攻击在不同目标音频深伪检测（ADD）模型之间具有可迁移性，这展示了其在 ADD 系统自动化红队中的广泛适用性和易用性。最后，在 FoeGlass 生成的样本上对音频深伪检测（ADD）模型进行微调，显著增强了检测器的鲁棒性（提升高达 41%）。

Audio deepfake detection (ADD) models are critical for countering the malicious use of text-to-speech (TTS) models. Evaluating and strengthening ADD models requires developing datasets that span the space of generated audio and highlight high-error regions. Existing dataset development strategies face two challenges: (i) manual collection, and (ii) inefficient discovery of blind spots in the ADD models. To address these challenges, we propose FoeGlass, the first black-box automated red-teaming method for ADDs, which effectively discovers ADD failure modes in the space of generated audio underexplored by state-of-the-art deepfake benchmarks. FoeGlass uses the in-context learning capabilities of an LLM to explore the input space of a TTS model, generating audio samples that fool the target ADD using only black-box access to all components. By using a carefully designed context based on diversity measurements, FoeGlass mitigates the common problem of mode collapse in automated red-teaming systems. Empirical evaluations on several open-source ADD and TTS models demonstrate that data generated from FoeGlass substantially improves the false negative rates over unconditional sampling baselines and recent spoofing datasets by up to 94%, while requiring no manual supervision. Furthermore, we show that the attacks generated by FoeGlass are transferable across different target ADDs, demonstrating its broad applicability and ease of use for the automated red teaming of ADD systems. Finally, fine-tuning ADD models on FoeGlass-generated samples notably enhances the robustness of the detectors (up 41%).

评分理由: The paper focuses on audio deepfake red-teaming using LLM in-context learning, lacking content on visual encoders, world models, model-based RL, or unified architectures. Relevance to specific keywords is low. No expert authors from the specified list are found.

Audio Deepfake Detection, Red Teaming, In-Context Learning, Large Language Model, Text-to-Speech, Adversarial Robustness, Black-box Attack

随着 ChatGPT 等公共大型语言模型（LLMs）的广泛部署，保护用户提示（prompt）隐私已成为一个日益关键的问题。现有的隐私保护推理方法往往在效用（utility）或效率（efficiency）之间做出妥协，并且通常需要针对特定模型的修改，这限制了它们的兼容性。本文提出 SharedRequest，这是一个与模型无关（model-agnostic）的隐私保护 LLM 推理框架，它将隐私保护重新定义为在批次（batch）级别而非单个提示（individual-prompt）级别上进行。其核心思想是通过将原始提示与噪声变体混合来模糊敏感信息，同时分组语义等效的指令，以在大量查询的批次上分摊推理成本，且对 LLM 响应质量的影响最小。该设计独立于 LLM 架构，无需访问模型参数或进行架构修改。实验结果表明，SharedRequest 相较于先前的差分隐私（differential privacy）基线方法，实现了超过 20% 的效用提升，且其共享提示机制将查询成本降低了最多 5 倍，相较于非批次推理。

With the widespread deployment of public large language models (LLMs) such as ChatGPT, protecting user prompt privacy has become an increasingly critical issue. Existing privacy-preserving inference methods sacrifice either utility or efficiency, and often require model-specific modifications that limit their compatibility. In this paper, we propose SharedRequest, a model-agnostic framework for privacy-preserving LLM inference that reformulates privacy protection at the batch level rather than the individual-prompt level. The key idea is to obscure sensitive information by mixing original prompts with noisy variants, while grouping semantically equivalent instructions to amortize the inference cost over a large batch of queries with minimal impact on LLM response quality. This design is independent of the LLM architecture, requiring no access to model parameters or architectural modification. Empirical results demonstrate that SharedRequest achieves over $20\%$ higher utility compared to prior differential privacy baselines, and its shared-prompt mechanism reduces query cost by up to $5\times$ compared to non-batched inference.

评分理由: 论文核心聚焦于大语言模型（LLM）的隐私保护推理框架（SharedRequest），主要贡献在于批处理级别的提示词混淆与语义分组以保护隐私并降低成本。提供的关键词集主要涵盖多模态、世界模型及强化学习领域，与本文文本主导的隐私优化主题存在显著领域偏差。'Unify Models'因模型无关性设计给予 2 分，'Tokenizer'因属于 LLM 基础组件给予 2 分，'MLLM'因涉及大模型范畴给予 2 分；其余关键词（Visual Encoder, World Models, MultiModal, model-based RL）与本文内容完全无关，给予 0 分。作者列表中未包含指定的 Yang Shi 等专家，无额外加分。加权总分为 9.0，低于动态及格分 27.8，表明关键词匹配度较低。

Privacy-Preserving Inference, Model-Agnostic Framework, Large Language Models, Batch-level Prompt Mixing, Semantic Grouping, Inference Cost Reduction, Prompt Obfuscation

控制实验是机器学习研究的基石，但在现代基础模型（Foundation Models）的规模下，其代价已变得难以承受。取而代之的是，研究社区越来越依赖那些以极低成本近似理想实验的研究策略：代理实验（Proxy Experiments）和缩放定律（Scaling Laws）、基于公开模型的观察性研究（Observational Studies），以及利用单次训练运行内变异性的单次运行设计（Single-Run Designs）。本文认为，在计算预算（Compute Budget）约束下近似大规模实验时，天下没有免费的午餐。具体而言，计算资源的节省是以有效性威胁（Validity Threats）为代价的——这些是隐藏的、有时甚至不可验证的假设，一旦被违背，便可能推翻研究主张。为应对此类威胁，我们提出一个评估框架，将基础模型研究视为一个因果推断（Causal Inference）问题。在此框架内，我们借鉴经验社会科学的四种有效性类型——统计有效性（Statistical Validity）、内部有效性（Internal Validity）、外部有效性（External Validity）和构念有效性（Construct Validity）——来评估不同的研究策略。我们发现每种策略都带有其独特的有效性特征：代理实验以牺牲外部有效性和构念有效性为代价，换取统计有效性和内部有效性；观察性研究面临混杂（Confounding）效应和效应异质性（Effect Heterogeneity）；而单次运行设计则受到处理单元（Treated Units）间干扰（Interference）的困扰。该分析揭示了文献中未受到足够关注的若干有效性威胁。总体而言，我们的评估框架为研究人员提供了一个实用工具包，用于审视基础模型研究设计中的有效性威胁。

Controlled experiments are the backbone of machine learning research, but at the scale of modern foundation models, they have become prohibitively expensive. Instead, the community increasingly relies on research strategies that approximate the ideal experiment at a fraction of the cost: proxy experiments and scaling laws, observational studies with publicly available models, and single-run designs that leverage variation within individual training runs. In this work, we argue that there is no free lunch when approximating large-scale experiments on a compute budget. Specifically, savings in compute come at the cost of validity threats -- hidden and sometimes untestable assumptions that, when violated, can invalidate research claims. To help navigate such threats, we propose an evaluation framework that casts foundation model research as a causal inference problem. Within this framework, we evaluate different research strategies through four types of validity adapted from the empirical social sciences -- statistical, internal, external, and construct validity. We find that each strategy comes with a characteristic validity profile: proxy experiments trade external and construct validity for statistical and internal validity; observational studies face confounding and effect heterogeneity; and single-run designs are strained by interference between treated units. This analysis reveals several validity threats that have received insufficient attention in the literature. Overall, our evaluation framework provides researchers with a practical toolkit for scrutinizing validity threats in foundation model research~designs.

评分理由: The paper focuses on research methodology and validity evaluation in foundation model research (causal inference, proxy experiments), rather than technical architectures or specific reinforcement learning strategies. While it mentions foundation models broadly, it does not discuss tokenizers, visual encoders, world models, or model-based RL specifically, resulting in low relevance scores for technical keywords. Total Weighted Score: 9.0, which is below the dynamic passing threshold of 27.8.

Foundation Model Research, Validity Threats, Causal Inference, Proxy Experiments, Scaling Laws, Statistical Validity, Internal Validity, External Validity

大语言模型正越来越多地由其他模型进行评估，这引发了一个自然的问题：一个模型能否预测评判者将如何对其自身输出进行评分？我们发现，这种能力在很大程度上存在于任何针对性训练之前：通过提示式少样本（few-shot），基础模型已在三个基准测试上，显著高于随机水平地预测了外部评判者对开放式回答的多属性质量分数。我们引入了自我评估诱导（Self-Evaluation Elicitation, SEE），该方法通过一个短周期揭示这种潜在能力：首先是一个校准耦合强化学习阶段，用于改进答案并预测评判者；随后是一个掩码蒸馏阶段，该阶段细化预测同时保持答案不变。仅使用 160 个唯一示例（约为强化学习基线的 1/31），SEE 在三个基准测试上提高了保留集校准效果，同时保持了答案质量。诱导出的自我评估紧密局限于模型自身的词元分布中，且对从未训练过的评判者保持稳定，这表明这是一种可迁移的质量概念，而非单个评判者的偏好。这些结果将与评判者对齐的自我评估问题重新定义为诱导问题，而非习得问题。

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.

评分理由: 论文聚焦于文本大模型的自我评估与法官校准，未涉及多模态融合（无 MultiModal, MLLM, Visual Encoder），也未构建环境世界模型（World Models）。虽使用了强化学习阶段，但并非典型的环境模型构建（model-based RL），Tokenizer 仅提及 token 分布而非架构设计，Unify Models 不适用。因此与给定关键词集相关性较低。

Self-Evaluation, Judge Calibration, Base LLMs, Reinforcement Learning, Masked Distillation, Minimal Data, Elicitation

训练数据归因（TDA）旨在追溯模型预测背后的训练数据。TDA 的黄金标准依赖于因果干预，即观察数据添加或移除时模型的变化，但对于大型语言模型（LLMs）而言，反复重训练在计算上极具挑战性。因此，大多数方法在参数空间中利用梯度来近似这种效应。然而，追踪数十亿参数上的梯度不仅成本高昂，而且依赖于局部近似。在本文中，我们提出了一种范式转变：不估计参数变化，而是在激活空间中建模训练数据的功能效应。我们提出了 STRIDE（基于引导的训练数据影响分解），这是一个将 TDA 表述为稀疏恢复问题的框架，其思路源于压缩感知。STRIDE 学习轻量级的“引导算子”，以模仿在子集数据上训练所引发的行为偏移。通过测量这些算子如何扰动测试预测，我们通过稀疏线性分解恢复单个训练样本的影响。STRIDE 在 LLM 预训练归因任务上达到了最先进水平，同时比先前方法快一个数量级（13 倍）。我们进一步通过下游应用验证其实用性，包括数据选择、数据污染及定性分析。

Training Data Attribution (TDA) seeks to trace a model's predictions back to its training data. The gold standard for TDA relies on causal interventions, observing how a model changes when data is added or removed, but repeated retraining is computationally challenging for Large Language Models (LLMs). Consequently, most approaches approximate this effect in the parameter space using gradients. However, tracking gradients across billions of parameters is not only prohibitively expensive but relies on local approximations. In this work, we propose a shift: rather than estimating parameter changes, we model the functional effect of training data in the activation space. We introduce STRIDE (Steering-based Training Data Influence Decomposition), a framework that formulates TDA as a sparse recovery problem in the spirit of compressive sensing. STRIDE learns lightweight "steering operators" that mimic the behavioral shift caused by training on data subsets. By measuring how these operators perturb test predictions, we recover individual training example influences via sparse linear decomposition. STRIDE achieves state-of-the-art for LLM pre-training attribution while being an order of magnitude ($13\times$) faster than previous art. We further validate its practical utility through downstream applications including data selection, data contamination, and qualitative analysis.

评分理由: 论文核心为训练数据归因（TDA）与稀疏恢复，主要在大语言模型（LLM）的激活空间进行操作。提供的关键词涉及统一模型架构、分词器设计、视觉编码器、世界模型构建、多模态大模型（MLLM）、多模态融合或基于模型的强化学习，与本文主题（文本模型的数据归因）高度不相关，因此相关度评分极低。

Training Data Attribution, Sparse Recovery, Activation Space, Steering Operators, Large Language Models, Subset Perturbations, Compressive Sensing

机器学习工程（MLE）代理有望实现从原始数据和自然语言指令到端到端 ML 流水线开发的自动化，从而使非技术背景的领域专家能够使用机器学习。然而，在敏感且受监管的领域中，这种抽象会产生一个责任鸿沟：最终用户可能无法洞察影响正确性、鲁棒性、公平性及监管合规性的设计选择。我们认为现有的基准不足以评估 MLE 代理是否能在此类环境中安全应用。我们提出了以责任为中心的评估框架的期望，并在黑色素瘤分类任务上开展了一项探索性研究，将肤色公平性作为责任约束重点考察。在对两个近期 MLE 代理进行评估时，我们发现，代理生成的流水线表现出高方差，且在预测质量和公平性方面始终低于手动设计的基线，尽管采用了面向公平的提示。这些初步结果表明，需要进一步研究重新设计 MLE 代理，以便人类能够引导搜索过程，并可靠地评估所生成的 ML 流水线的合规性与质量。

Machine learning engineering (MLE) agents promise to automate end-to-end ML pipeline development from raw data and natural language instructions, potentially making ML accessible to non-technical domain experts. However, in sensitive and regulated domains, this abstraction creates a responsibility gap: end-users may lack visibility into design choices that affect correctness, robustness, fairness, and regulatory compliance. We argue that existing benchmarks are insufficient to assess whether MLE agents can be safely applied in such settings. We propose desiderata for a responsibility-centered evaluation framework and conduct an exploratory study on melanoma classification, focusing on fairness across skin tones as a responsibility constraint. When evaluating two recent MLE agents, we find that agent-generated pipelines show high variance and consistently underperform manually designed baselines in both predictive quality and fairness, despite fairness-oriented prompts. These preliminary results suggest that further research is needed towards redesigning MLE agents to allow humans to guide the search process and reliably assess the compliance and quality of the generated ML pipelines.

评分理由: 论文核心在于机器学习工程（MLE）代理的公平性约束与责任评估，未涉及世界模型、视觉编码器、Tokenizer 或模型强化学习等核心技术。关键词主要指向多模态大模型架构及强化学习领域，与本文主题（自动化 ML 流程评估）存在显著差异。虽然任务涉及图像数据（黑色素瘤），但未讨论视觉编码器架构或多模态模型统一，故相关度评分较低。

Machine Learning Engineering Agents, Fairness Constraints, Responsibility Gap, Melanoma Classification, Automated ML Pipelines, Evaluation Framework, Predictive Quality

扩散大语言模型（DLLMs）近期已成为一种有前景的替代自回归大语言模型（autoregressive LLMs）的方案，其通过结合双向上下文的迭代掩码去噪机制生成文本。然而，其庞大的模型规模和迭代去噪过程引入了显著的内存和计算开销，这促使采用后训练量化（PTQ）以实现高效部署。本文针对低比特 DLLM 量化识别出两个关键挑战：状态依赖的激活差异（state-dependent activation disparity）和时间误差累积（temporal error accumulation）。在每个去噪步骤中，掩码 token 与非掩码 token 表现出不同的激活分布，而在迭代解码过程中，量化误差可能会跨步骤累积。为应对这些挑战，本文提出了一种面向 DLLMs 的状态 - 时间一致后训练量化（PTQ）框架，即 STaR-Quant。STaR-Quant 引入了状态引导激活变换（State-Guided Activation Transformation, SGAT），通过统一的静态权重侧变换将掩码与非掩码 token 映射至不同的激活变换空间。此外，它还引入了时间注意力补偿（Temporal Attention Compensation, TAC），通过轻量级块对角仿射映射来校正量化注意力表示。在代表性 DLLMs 上的实验表明，STaR-Quant 在低比特权重 - 激活量化方面始终优于强 PTQ 基线，同时相比 FP16 部署可实现高达 1.69 倍的加速比和 3.14 倍的内存节省。

Diffusion large language models (DLLMs) have recently emerged as a promising alternative to autoregressive LLMs by generating text through iterative masked denoising with bidirectional context. However, their large model sizes and iterative denoising process introduce substantial memory and computational overhead, motivating post-training quantization for efficient deployment. In this paper, we identify two key challenges for low-bit DLLM quantization: state-dependent activation disparity and temporal error accumulation. Masked and unmasked tokens exhibit different activation distributions within each denoising step, while quantization errors can accumulate across steps during iterative decoding. To address these challenges, we propose STaR-Quant, a state-time consistent PTQ framework for DLLMs. STaR-Quant introduces State-Guided Activation Transformation (SGAT) to assign masked and unmasked tokens to different activation transformation spaces with a unified static weight-side transformation. It further introduces Temporal Attention Compensation (TAC) to correct the quantized attention representation via a lightweight block-diagonal affine mapping. Experiments on representative DLLMs demonstrate that STaR-Quant consistently improves low-bit weight-activation quantization over strong PTQ baselines, while delivering up to 1.69x speedup and 3.14x memory saving over FP16 deployment.

评分理由: 论文核心内容为扩散语言模型（DLLMs）的后训练量化（PTQ），旨在解决内存和计算开销问题。与提供的关键词相比，相关性普遍较低：'Tokenizer'因涉及 token 激活态略有相关性（2 分）；'Unify Models'和'World Models'因涉及生成模型架构有微弱关联（1 分）；'Visual Encoder'、'MultiModal'、'model-based RL'完全无关（0 分）；'MLLM'因论文聚焦文本而非多模态（1 分）。作者列表中未包含指定的 Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang 专家，故无加分。加权总分为 7.5，远低于动态及格分 27.8。

Diffusion Large Language Models, Post-Training Quantization, State-Time Consistent, Activation Transformation, Temporal Attention Compensation, Low-bit Quantization, Efficient Deployment

大型语言模型（LLMs）中的自回归思维链（CoT）推理本质上是前向导向的：每一步仅依赖于先前的标记。这种单向归纳偏置使得即使是有能力的模型也容易受到错误雪球效应的影响，即早期步骤中的单个逻辑或算术错误会不可逆地破坏整个推理链。我们引入目的论推理填充（TRI），这是一种训练框架，赋予仅解码器变换器原生的目标条件化桥接能力。关键洞察是将错误的推理片段重新表述为中间填充（FIM）任务：给定一个已验证的前缀前提 $P$、一个已验证的下游里程碑 $S$ 以及原始查询 $Q$，模型必须合成出逻辑桥接 $M$，该桥接需严格且完整地连接 $P$ 与 $S$。为了在标准因果架构下实现这一点，我们引入前缀 - 后缀 - 中间（PSM）序列重排，该重排采用三个非重叠的哨兵标记，使 $M$ 能够同时关注 $P$ 和 $S$，而无需对自注意力机制进行任何结构修改。训练分为两个阶段：(i) 在从形式数学语料库中提取的符号验证过的 $(P, S, M)$ 三元组上进行监督微调（SFT），(ii) 使用确定性符号验证器（Lean 4 / Python）作为唯一奖励预言机进行直接偏好优化（DPO），从而消除大语言模型评判者的谄媚。在推理阶段，TRI 在双系统循环中作为一个手术式修复模块运行：因果草稿模型生成初始推理轨迹，验证器定位故障，TRI 仅填充受损部分，保持已验证部分完整。在三个基准上的全面实验表明，TRI 在所有任务上实现了最先进的性能，同时将每个问题的令牌消耗减少了 31.2%。

Autoregressive chain-of-thought (CoT) reasoning in large language models (LLMs) is fundamentally forward-directed: each step conditions only on prior tokens. This unidirectional inductive bias renders even capable models susceptible to error snowballing, wherein a single logical or arithmetic mistake in an early step irreversibly corrupts the entire reasoning chain. We introduce Teleological Reasoning Infilling (\TRI{}), a training framework that endows decoder-only transformers with a native \emph{goal-conditioned bridging} capability. The key insight is to reframe erroneous reasoning segments as fill-in-the-middle (FIM) tasks: given a verified prefix premise $P$, a verified downstream milestone $S$, and the original query $Q$, the model must synthesise the logical bridge $M$ that connects $P$ to $S$ rigorously and completely. To achieve this with standard causal architectures, we introduce a Prefix-Suffix-Middle (PSM) sequence rearrangement with three non-overlapping sentinel tokens, enabling $M$ to attend to both $P$ and $S$ without any structural modification to the self-attention mechanism. Training proceeds in two stages: (i) Supervised Fine-Tuning (SFT) on symbolically verified $(P, S, M)$ triples extracted from formal mathematics corpora, and (ii) Direct Preference Optimisation (DPO) with a deterministic symbolic verifier (Lean 4 / Python) as the sole reward oracle, eliminating LLM-judge sycophancy. At inference, TRI operates as a surgical repair module within a dual-system loop: a causal draft model generates an initial trace, the verifier pinpoints failures, and TRI infills only the damaged segment, leaving verified sections intact. Comprehensive experiments on three benchmarks demonstrate that TRI achieves state-of-the-art performance across all tasks, while reducing per-problem token expenditure by 31.2%.

评分理由: 该论文专注于通过双向填充（TRI）修复 LLM 推理链，与多模态学习、视觉编码器或世界模型无关。虽然使用了 DPO（与 RL 相关），但并非模型强化学习。标记器使用标准，未涉及模型统一。评分反映了与所提供关键词集的低相关性。

Large Language Models, Chain-of-Thought, Fill-in-the-Middle, Bidirectional Logic, Reasoning Repair, Symbolic Verification, Prefix-Suffix-Middle

Self-attention (自注意力机制) 在序列上自由选择信息，但在深度维度上，Transformer (变换器) 仅将每一层的输出添加到残差流中，因此后续层无法有选择地重用早期层的表示。近期的跨层方法改进了这一流程，但它们在注意力机制之外的隐藏状态上操作，在推理时引入了超出 key-value cache (键值缓存) 之外的状态——这一成本随着现代 LLM (大语言模型) 利用 grouped-query (组查询) 和 multi-head latent attention (多头潜在注意力) 压缩缓存而日益显著。我们提出了 Depth-Attention (深度注意力)，它在 attention 模块内部执行这种选择：在某一层对序列进行注意力计算之前，其 query 会在同一 token 位置关注早期层的 keys，并将它们的 values 混合到 self-attention 随后读取的值中。由于 Depth-Attention 重用了标准的 attention query、key 和 value-cache 槽位，将深度混合值存储在原值位置，因此它不增加任何参数，也不引入超出标准 key-value cache 之外的持久推理状态——其缓存大小与 vanilla decoder (原始解码器) 相同，且少于基于隐藏状态的跨层方法。在参数量为 15 亿和 30 亿的 Qwen3 风格解码器上，Depth-Attention 达到了最低的 perplexity (困惑度) 和最高的平均下游准确率，相比原始 Transformer 提升了高达 2.3 个百分点的准确率，并在 perplexity 和平均准确率上超越了强大的跨层基线方法，同时仅增加了不到 0.01% 的额外 arithmetic FLOPs (算术浮点运算次数)，且未引入额外的持久推理状态。这些增益在 3.6 亿到 30 亿参数范围内均成立，并适用于 looped Transformers (循环变换器)。

Self-attention selects information freely across the sequence, but across depth, Transformers merely add each layer's output to the residual stream, so later layers cannot selectively reuse earlier-layer representations. Recent cross-layer methods improve this flow but operate on hidden states outside attention, adding state beyond the key-value cache at inference--a cost that becomes increasingly salient as modern LLMs compress the cache with grouped-query and multi-head latent attention. We introduce Depth-Attention, which performs this selection inside the attention module itself: before a layer attends over the sequence, its query attends over the keys of earlier layers at the same token position and mixes their values into the value that self-attention then reads. Because Depth-Attention reuses the standard attention queries, keys, and value-cache slots, storing depth-mixed values in place of the original values, it adds no parameters and introduces no persistent inference state beyond the standard key-value cache--the same cache size as a vanilla decoder and less than hidden-state-based cross-layer methods. On Qwen3-style decoders at 1.5B and 3B parameters, Depth-Attention attains the lowest perplexity and the highest average downstream accuracy, improving over the vanilla Transformer by up to 2.3 accuracy points and surpassing strong cross-layer baselines in perplexity and average accuracy, while adding under 0.01% extra arithmetic FLOPs and no additional persistent inference state. The gains hold from 360M to 3B parameters and extend to looped Transformers.

评分理由: 论文核心为语言模型架构优化（Depth-Attention），涉及跨层注意力与值混合，旨在提升困惑度与准确率。提供的关键词集侧重于多模态（MLLM, MultiModal, Visual Encoder）、世界模型及强化学习，与本文纯文本 LLM 主题关联度低。Tokenizer 与视觉编码器未在摘要中体现。作者列表中未发现指定专家。加权总分远低于动态及格分。

Depth-Attention, Cross-Layer Value Mixing, Language Models, Self-attention, Key-Value Cache, Perplexity, Downstream Accuracy, Transformer Architecture

扩散大语言模型（DLLMs）通过迭代去噪含噪声 token 序列并利用双向上下文，实现了非自回归生成。尽管它们能够并行更新多个位置，但由于高质量生成需要大量的去噪步骤，推理开销仍然较大。我们提出 SAID，一种感知骨架的迭代解码框架，通过重新分配 token 间的计算来加速 DLLMs。SAID 首先将去噪计算集中在骨架 token 上以建立粗略语义结构，随后用更少的步骤完成可预测的细节 token。我们进一步将 SAID 适配到块级扩散解码，并引入置信度分层生成（CHLG），该机制仅向低置信度 token 分配额外步骤。在 LLaDA-8B 和 LLaDA 1.5 上进行的数学、编码及知识基准测试实验表明，SAID 显著加速了 DLLM 推理，最大加速比达 9.1 倍，同时保持了具有竞争力的性能。我们的代码公开可用：https://github.com/TH-AI-Lab-PKU/SAID.

Diffusion large language models (DLLMs) enable non-autoregressive generation by iteratively denoising corrupted token sequences with bidirectional context. Despite their ability to update multiple positions in parallel, inference remains costly due to the many denoising steps required for high-quality generation. We propose SAID, a Scaffold-Aware Iterative Decoding framework that accelerates DLLMs by reallocating computation across tokens. SAID first spends denoising computation on scaffold tokens to establish the coarse semantic structure, and then completes predictable detail tokens with fewer steps. We further adapt SAID to block-wise diffusion decoding and introduce Confidence-Hierarchical Layered Generation (CHLG), which assigns additional steps only to low-confidence tokens. Experiments on LLaDA-8B and LLaDA 1.5 across math, coding, and knowledge benchmarks show that SAID significantly accelerates DLLM inference with a maximum speedup of 9.1x while maintaining competitive performance. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SAID.

评分理由: 该论文专注于扩散语言模型（DLLMs）的推理加速，核心贡献在于解码策略优化（SAID 框架）。提供的关键词集主要涵盖多模态、世界模型及强化学习领域，与本文纯文本生成加速的主题匹配度极低。Visual Encoder、World Models、MLLM、MultiModal、model-based RL 完全无关（0 分）。Tokenizer 涉及 token 处理但未深入设计（3 分），Unify Models 仅指代生成步骤统一而非模型统一（2 分）。

Diffusion Language Models, Scaffold-Aware Iterative Decoding, Inference Acceleration, Denoising Process, Block-wise Decoding, Confidence-Hierarchical Generation, Text Generation

可验证奖励的强化学习（例如 GRPO）现已成为提升大语言模型（LLMs）数学推理能力的常用方法。然而，当前方法通常将一个序列级优势广播至所有词元，或使用昂贵的过程奖励模型（PRMs）进行步骤级监督。均匀的优势分布假设所有词元对最终奖励的贡献相等。这会稀释梯度信号，因为有缺陷的推理步骤和填充词会被像有效的逻辑推理一样强烈地更新。为了解决这一问题，我们引入了梯度重加权优势（GRAIL），这是一种内在的词元级优势重加权方法。GRAIL 利用梯度 - 激活显著性，赋予那些对最终答案更局部敏感的词元更高的权重。在来自 Qwen3、R1-distilled 和 OctoThinker 系列的五个模型上的评估表明，GRAIL 一贯优于 GRPO。GRAIL 在准确率和 Pass@3 上分别平均提升了 3.60% 和 3.05%，这表明无需过程级监督即可实现细粒度推理对齐。

Reinforcement learning with verifiable rewards (e.g. GRPO) is now a common way to improve mathematical reasoning in Large Language Models (LLMs). However, current methods usually broadcast one sequence-level advantage to all tokens, or use costly process reward models (PRMs) for step-level supervision. Uniform advantage distribution assumes that all tokens contribute equally to the final reward. This dilutes the gradient signal, since flawed reasoning steps and filler words are updated as strongly as valid logical inferences. To address this, we introduce Gradient-Reweighted Advantage (GRAIL), an intrinsic token-wise advantage reweighting method. GRAIL uses gradient-activation saliency to place more weight on tokens that are more locally sensitive to the final answer. Evaluations across five models from the Qwen3, R1-distilled and OctoThinker families show that GRAIL consistently outperforms GRPO. GRAIL achieved an average improvement of 3.60% in accuracy and 3.05% in Pass@3, demonstrating that fine-grained reasoning alignment can be achieved without process-level supervision.

评分理由: The paper focuses on token-wise advantage reweighting (GRAIL) for reinforcement learning with verifiable rewards in Large Language Models (LLMs). It does not address model unification, tokenizers, visual encoders, world models, or multimodal processing, resulting in low scores for these keywords. 'MLLM' and 'model-based RL' receive minimal credit (2.0) as the paper involves LLMs and RL respectively, though it specifically uses model-free GRPO variants rather than model-based dynamics or explicit multimodality. None of the specified expert authors (Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang) are present in the author list.

Gradient-Reweighted Advantages, Reinforcement Learning, Verifiable Rewards, Token-wise Advantage, Large Language Models, Reasoning Alignment, Gradient-activation Saliency

多智能体推理系统采用“生成后传输（generate-then-transfer）”范式，迫使端到端延迟随流水线深度线性增长。我们引入 StreamMA，这是一种多智能体推理系统，它在生成后立即将每个推理步骤流式传输至下游代理，通过对相邻代理进行流水线化处理从而降低延迟。令人惊讶的是，这种流水线化同时也提升了有效性：由于多步推理质量存在非均匀性，早期步骤比后期步骤更为可靠，使用这些可靠的早期步骤而非完整链条，可防止易出错的后期步骤误导下游代理。我们通过首次对流式（stream）、串行（serial）和单（single）协议进行闭联合分析，形式化地阐述了这两种优势，推导出了有效性排序、加速比上界及成本比率。在涵盖数学、科学和代码的八个推理基准上，针对两种前沿大语言模型（Claude Opus 4.6 和 GPT-5.4）及三种拓扑结构（链、树、图），StreamMA 均优于基线模型（平均提升 +7.3 个百分点，在 HMMT 2026 上最高提升 +22.4 个百分点；基于 Claude Opus 4.6-high）。除上述贡献外，我们发现了一种“步骤级扩展定律”：增加每个代理的步骤数一致地提升了有效性和效率，这是一种新的扩展维度，与代理数量扩展正交且可组合。

Multi-agent reasoning systems adopt a "generate-then-transfer" paradigm that forces end-to-end latency to scale linearly with pipeline depth. We introduce StreamMA, a multi-agent reasoning system that streams each reasoning step to downstream agents as soon as it is generated, pipelining adjacent agents and thus reducing latency. Surprisingly, this pipelining also improves effectiveness: because multi-step reasoning quality is non-uniform and early steps are more reliable than later ones, working with these reliable early steps instead of the full chain prevents error-prone late steps from misleading downstream agents. We formalize both advantages with the first closed-form joint analysis of stream, serial, and single protocols, deriving the effectiveness ordering, speedup upper bound, and cost ratio. Across eight reasoning benchmarks spanning mathematics, science, and code, two frontier LLMs (Claude Opus 4.6 and GPT-5.4), and three topologies (Chain, Tree, Graph), StreamMA outperforms both baselines (avg. +7.3 pp, max +22.4 pp on HMMT 2026; Claude Opus 4.6-high). Beyond these contributions, we discover a "step-level scaling law": increasing per-agent steps consistently improves both effectiveness and efficiency, a new scaling dimension orthogonal to and composable with agent-count scaling.

评分理由: The paper focuses on multi-agent reasoning communication protocols (StreamMA) rather than multimodal architecture, tokenization strategies, visual encoders, world models, or reinforcement learning. While it uses frontier LLMs (potentially MLLM), the core contribution is pipeline streaming, not model unification or representation learning, resulting in low relevance to the provided keyword theme.

Multi-Agent Reasoning, Streaming Communication, Latency Reduction, Pipelining Agents, Scaling Laws, StreamMA, LLM-based Reasoning

在 3D Gaussian Splatting (3DGS) 成功应用于新视角合成之后，许多工作也探索了如何利用它进行几何表面表示。然而，直接从 3DGS 中提取准确的几何信息仍然具有挑战性，并且往往会降低外观渲染质量。本文通过训练使用完整的 ground-truth（真实值）纹理和几何信息，表明默认形式的 3DGS 固有地不适合同时表示纹理和几何。我们还提出了一种简单解决方案，即对每个高斯点（splat）应用一个额外的几何不透明度参数，并配合一个可选的透明度引导优化流程。我们的实验，无论是使用 ground-truth 还是 vision foundation model（视觉基础模型）的几何输入，都表明这种改进在各种数据集上带来了渲染和几何性能的提升，尤其是包含透明物体的复杂场景从我们的方法中获益显著。

After the success of 3D Gaussian Splatting (3DGS) for novel view synthesis, many works have explored how to also use it for geometric surface representation. However, extracting accurate geometric information directly from 3DGS remains challenging and can often reduce the appearance rendering quality. In this work, we show that 3DGS in its default form is inheritedly unsuited to represent texture and geometry at the same time, by training with complete ground-truth texture and geometry information. We also propose a simple solution by applying a single additional geometry opacity parameter to each splat, together with an optional transparency-curated optimization pipeline. Our experiments, both with ground-truth and vision foundation model geometric input, show that this change leads to improved rendering and geometry performance on a wide variety of dataset, and especially complex scenes with transparent objects benefit significantly from our method.

评分理由: The paper focuses on 3D Gaussian Splatting for geometry and appearance decoupling in novel view synthesis. It does not involve Tokenizers, MLLMs, World Models, or Reinforcement Learning. While it handles geometry and appearance (loosely MultiModal), it lacks alignment with the Multimodal AI/LLM context of the other keywords. No expert authors from the specified list were found.

3D Gaussian Splatting, Geometry Representation, Appearance Decoupling, Novel View Synthesis, Geometry Opacity, Transparent Objects, Rendering Performance

大型语言模型（LLMs）正被越来越多地部署于各类应用中，这引发了关于治理、问责和数据溯源的关键问题。理解哪些训练数据对模型输出影响最大，仍然是一个根本性的开放性问题。我们通过针对自回归大语言模型（auto-regressive LLMs）的训练数据归因（TDA）来解决这一挑战，该方法扩展了逆向表述：如果模型在训练过程中看到了生成的输出，训练数据会受到怎样的影响？我们的方法通过对生成的文本样本使用双向梯度优化（bidirectional gradient optimization，包括梯度上升和下降）来扰动基础模型，并测量训练样本上损失函数的变化。我们的框架支持任意数据粒度的归因，从而实现事实归因和风格归因。我们在具有已知数据集的预训练模型上，将我们的方法与基线进行了对比评估，结果显示其在影响力指标上优于先前工作，从而增强了模型的可解释性，这是可问责 AI 系统的基本要求。

Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance. Understanding which training data most influenced a model's output remains a fundamental open problem. We address this challenge through training data attribution (TDA) for auto-regressive LLMs by expanding upon the inverse formulation: How would training data be affected if the model had seen the generated output during training? Our method perturbs the base model using bidirectional gradient optimization (gradient ascent and descent) on a generated text sample and measures the resulting change in loss across training samples. Our framework supports attribution at arbitrary data granularity, enabling both factual and stylistic attribution. We evaluate our method against baselines on pretrained models with known datasets, and show that it outperforms previous work on influence metrics, thereby enhancing model interpretability, an essential requirement for accountable AI systems.

评分理由: The paper focuses on training data attribution for text-based LLMs using bidirectional gradient optimization. It is unrelated to World Models, MLLM, MultiModal, Visual Encoder, and model-based RL as it focuses solely on text generation and attribution. Tokenizers are used in LLMs but not analyzed, hence low score (2.0). 'Unify Models' is not discussed, hence low score (2.0). Consequently, the paper has very low relevance to the provided keyword set, which is centered on multimodal world models and RL, resulting in a total weighted score significantly below the passing threshold.

Data Attribution, Large Language Models, Bidirectional Gradient Optimization, Training Data Attribution, Model Interpretability, Accountability, Auto-regressive LLMs, Gradient Ascent and Descent

量化深度神经网络对于在资源受限设备上进行高效推理至关重要。然而，大多数现有方法是为单域和类别平衡数据设计的，导致存在领域偏移或严重类别不平衡的实际场景未被充分探索。我们通过高效多域对齐量化（EmaQ）来解决这些挑战，该方法通过基于累积分布函数（CDF）的投影对齐领域分布，并使用敏感度感知的权重聚合来稳定多域量化。我们进一步将 EmaQ 扩展为 EmaQ-LT 以进行长尾量化，通过引入类别条件方差缩放和基于置信度的 logit 调整来缓解多数类过度自信。理论分析确立了收敛性保证，并为所提出的敏感度和缩放机制提供了理论依据。在标准、多域（Office-31, Digits）和长尾（SynDigits-LT, CIFAR-10-LT, CIFAR-100-LT）基准上的实验表明，EmaQ 和 EmaQ-LT 在领域偏移和类别不平衡下实现了优异的低比特性能。

Quantizing deep neural networks is essential for efficient inference on resource-constrained devices. However, most existing methods are designed for single-domain and class-balanced data, leaving practical settings with domain shifts or severe class imbalance underexplored. We address these challenges with Efficient Multi-Domain Alignment Quantization (EmaQ), which aligns domain distributions through a CDF-based projection and uses sensitivity-aware weight aggregation to stabilize multi-domain quantization. We further extend EmaQ to EmaQ-LT for long-tailed quantization by introducing class-conditioned variance scaling and confidence-based logit adjustment to mitigate majority-class overconfidence. Theoretical analyses establish convergence guarantees and motivate the proposed sensitivity and scaling mechanisms. Experiments on standard, multi-domain (Office-31, Digits), and long-tailed (SynDigits-LT, CIFAR-10-LT, CIFAR-100-LT) benchmarks show that EmaQ and EmaQ-LT achieve strong low-bit performance under domain shift and class imbalance.

评分理由: 该论文核心贡献在于深度神经网络的量化技术（Quantization），特别是针对多域（Multi-Domain）和长尾（Long-Tailed）分布的量化方法（EmaQ/EmaQ-LT）。提供的关键词集主要聚焦于多模态大模型（MLLM）、世界模型（World Models）及强化学习（RL）领域。论文内容与这些关键词高度不相关：未涉及 Tokenizer、World Models、MLLM 或 Model-Based RL。仅在视觉基准测试中使用了图像数据（与 Visual Encoder 微弱相关），且在分布对齐技术上涉及域对齐（与 Unify Models 微弱相关），MultiModal 指多模态而非多域，相关性极低。因此，加权总分远低于动态及格分。

Quantization, Multi-Domain, Long-Tailed, Feature Alignment, Scaling, Domain Shift, Class Imbalance

倡议与公投是瑞士民主的核心，然而手写签名列表的验证仍是一项劳动密集型的人工流程。本文探讨了自动化文档分析方法的潜力，包括 OCR（光学字符识别）和基于 AI 的手写分析，以支持这一任务。我们提出了一种结合基于模板的行分割、文本识别与笔迹检索技术的流程，并在包含 418 位作者 443 个手写条目的数据集上进行了评估。结果表明，OCR 在处理词汇表外手写体时表现不佳，名字部分的字符错误率（CER）为 29.6%。相比之下，笔迹检索表现更为稳健，达到了 50.6% 的平均精度（mAP）。此外，我们的实验表明，现成的 OCR 系统对于转录手写签名数据并不足够可靠，特别是对于名字或地址等短小且词汇表外的条目。然而，笔迹检索方法可以有效识别签名列表中视觉上相似的条目，使其成为基于笔迹相似性检测潜在重复提交的合适工具。

Popular initiatives and referendums are central to Swiss democracy, yet the validation of handwritten signature lists remains a labor-intensive manual process. This paper investigates the potential of automated document analysis methods, including OCR and AI-based handwriting analysis, to support this task. We propose a pipeline combining template-based line segmentation with text recognition and writer retrieval techniques, evaluated on a dataset of 443 handwritten entries from 418 writers. Results show that OCR struggles with out-of-vocabulary handwriting, with a CER of 29.6% for first names. In contrast, writer retrieval performs more robustly, reaching an mAP of 50.6%. Furthermore, our experiments indicate that off-the-shelf OCR systems are not sufficiently reliable for transcription of handwritten signature data, particularly for short, out-of-vocabulary entries such as names or addresses. However, writer retrieval methods can effectively identify visually similar entries across signature lists, making them a suitable tool for supporting the detection of potential duplicate submissions based on handwriting similarity.

评分理由: The paper focuses on traditional document analysis, OCR, and writer retrieval for validating Swiss signature lists. It does not involve World Models, MLLM, Model-Based RL, or Unify Models. There is minimal overlap with MultiModal (image+text processing) and Tokenizer/Visual Encoder in a general computer vision sense, but not in the context of modern large-scale learning architectures implied by the keywords.

Handwriting Extraction, Signature Lists, Swiss Popular Initiatives, OCR, Writer Retrieval, Document Analysis, Duplicate Submissions, Handwriting Similarity

道义推理是指通过将明确的规则与政策应用于案件特定事实来回答问题的任务，例如根据成文法计算纳税义务或确定移民上诉的结果。基于大语言模型（LLM）的道义推理面临的一个关键挑战是，相关的规则集可能很长且相互交叉引用，因此模型仍可能无法定位特定推理步骤所需的规则。我们引入了道义代理推理（DAR），这是一种代理推理设置，其中模型按需与成文法交互。我们在 DeonticBench 的困难子集上，利用多个评测框架对 DAR 进行了评估。在这些设置下，我们发现代理评测框架可以推动道义推理任务的前沿，但改进并不一致：性能较弱的模型在数值任务上往往性能下降，同时消耗更多的标记。

Deontic reasoning is the task of answering questions by applying explicit rules and policies to case-specific facts, for example computing tax liability under a statute or determining the outcome of an immigration appeal. A key technical challenge for LLM-based deontic reasoning is that the relevant ruleset can be long and cross-referenced, so models may still fail to locate the rules needed for a particular reasoning step. We introduce Deontic Agentic Reasoning (DAR), an agentic reasoning setup in which the model interacts with the statutes on demand. We evaluate DAR under multiple harnesses on hard subsets of DeonticBench. Across these settings, we find that agentic harnesses can push the frontier on deontic reasoning tasks, but improvements are not uniform: weaker models often degrade on numerical tasks while consuming far more tokens.

评分理由: 该论文主要研究大语言模型在规范推理任务中的代理式交互方法（DAR），核心在于规则检索与推理逻辑，未涉及视觉编码器、多模态融合、世界模型构建或模型强化学习等关键词所指向的技术领域。虽然文中讨论了 Token 消耗，但 Tokenizer 并非核心贡献。因此，论文与给定关键词集（偏向多模态与 RL 架构）的相关性非常低。

Deontic Reasoning, Agentic Harnesses, LLM, Statutes, Rules, DeonticBench, Token Consumption

我们提出了图集合 Transformer（Graph Set Transformer, GST），这是一种用于在图集合上进行学习的神经网络架构，专为那些依赖于集合级上下文及局部结构的逐元素预测任务而设计。现有的架构（包括 DeepSets 和 SetTransformer）需要从独立的图神经网络（GNN）中获取预编码的图嵌入，这在特征提取与集合级上下文建模之间形成了瓶颈。相比之下，GST 在每一层交错进行节点级特征传播与跨图上下文建模，并通过门控机制融合这两个层级的信息。我们在一个旨在隔离集合条件结构推理的受控合成测试套件上评估 GST，并在三个涵盖逐原子反应中心识别、反应产率预测及图像分类的真实数据基准上进行测试。在参数量预算相当的情况下，GST 在这些设置中的表现均优于基线方法。架构消融实验强烈表明，局部上下文与集合上下文的交错对此优势贡献显著。

We introduce the Graph Set Transformer (GST), a neural network architecture for learning on sets of graphs, designed for tasks in which per-element predictions depend on set-wide context as well as local structure. Existing architectures, including DeepSets and SetTransformer, require pre-encoded graph embeddings from a separate GNN, creating a bottleneck between feature extraction and set-level contextualisation. In contrast, GST interleaves node-level feature propagation and cross-graph contextual modelling at every layer, fusing the two levels of information through a gating mechanism. We evaluate GST on a controlled synthetic suite designed to isolate set-conditional structural reasoning and on three real-data benchmarks spanning per-atom reaction-centre identification, reaction yield prediction, and image classification. Under matched parameter budgets, GST performs better than the baselines across these settings. An architectural ablation strongly suggests that the interleaving of local and set context contributes substantially to this advantage.

评分理由: 该论文提出 Graph Set Transformer (GST) 架构，专注于集合图上的学习与推理，核心在于交织节点级特征传播与跨图上下文建模。提供的关键词主要围绕多模态大模型（MLLM）、世界模型、强化学习及视觉编码器等前沿领域。论文内容与这些关键词高度不相关：未涉及 Tokenizer、视觉编码器、世界模型或强化学习；仅在图像分类基准中使用了视觉数据，故 MultiModal 相关性极低；虽涉及上下文融合，但并非"Unify Models"范式。作者列表中未包含指定的 Yang Shi 等专家。因此，加权总分远低于动态及格分。

Graph Set Transformer, Sets of Graphs, Node-level Feature Propagation, Cross-graph Contextual Modelling, Reaction-centre Identification, Reaction Yield Prediction, Image Classification, Gating Mechanism

函数向量（FVs）是在上下文学习过程中提取的任务表示，可用于引导大型语言模型（LLMs）。然而，关于其定义方式的设计选择仍研究不足。本文研究了针对指令的 FV 定义在两个自由度上的变化所带来的影响：注意力头选择和引导方式。在注意力头选择方面，结合层相关传播（LRP）使用基于梯度的归因方法能显著提高效率和准确性。在 FV 引导方面，采用分布式方式应用比简单聚合能获得更高的准确性。代码已公开。

Function vectors (FVs) are task representations elicited during in-context learning that can be used to steer Large Language Models (LLMs). However, design choices in their formulation remain underexplored. In this work, we study the impact of varying FV definitions for instructions along two degrees of freedom: attention head selection and steering. For head selection, using gradient-based attributions with Layer-wise Relevance Propagation (LRP) substantially improves efficiency as well as accuracy. For FV steering, applying it in a distributed manner yields a higher accuracy compared to simple aggregation. Our code is publicly available.

评分理由: 论文主要关注大语言模型（LLM）中的函数向量（Function Vectors）及其在上下文学习中的引导作用，涉及注意力头选择和梯度归因。这与关键词中的多模态（MultiModal, Visual Encoder）、世界模型（World Models）、强化学习（model-based RL）及分词器（Tokenizer）无直接关联。虽然涉及 LLM（接近 MLLM），但未体现模型统一（Unify Models）的核心架构特征，因此相关性较低。

Function Vectors, Large Language Models, In-context Learning, Steering, Gradient-based Attributions, Layer-wise Relevance Propagation, Task Representations

列车延误预测对乘客和铁路运营商而言均是一个重要问题，但由于缺乏标准化的数据集、预测目标及评估协议，该领域的进展仍难以评估。为填补这一空白，我们引入了 RIDE，这是一个在比利时铁路网络上以全国规模构建的列车延误预测开放数据集与基准。RIDE 涵盖了 2023 年至 2025 年间的 9450 万次列车事件、360 万次行程以及 3570 万次天气记录。该数据集被组织为一个分层数据管道，从原始铁路和天气数据源延伸至两个公开版本：一个可重用的中间关系型数据集以及模型就绪基准数据集。该基准标准化了预测任务以及训练和测试数据。此外，它还提供了一个统一评估协议，支持模型间的直接比较。利用该框架，我们首次提供了针对非学习模型、统计学习模型及深度学习模型的全面比较评估。结果表明，基于学习的方法明显优于非学习模型，其中图神经网络（Graph Neural Networks, GNNs）取得了最佳平均性能，而最强的基于学习模型之间性能相对接近。除了聚合平均绝对误差（MAE）和均方根误差（RMSE）外，该框架还提供了按预测时域（prediction horizon）和延误变化的分解，从而能够更详细地分析模型在不同预测范式（forecasting regimes）下的行为。

Train delay prediction is an important problem for both passengers and railway operators, yet progress in the field remains difficult to assess due to the lack of standardized datasets, prediction targets, and evaluation protocols. To address this gap, we introduce RIDE, an open dataset and benchmark for train delay prediction built at nationwide scale over the Belgian railway network. RIDE covers 94.5M train events, 3.6M journeys, and 35.7M weather records from 2023 to 2025. It is organized as a layered data pipeline from raw railway and weather sources to two public releases: a reusable intermediate relational dataset and model-ready benchmark datasets. The benchmark standardizes the prediction task and the training and testing data. It also provides a unified evaluation protocol that supports direct comparison across models. Using this framework, we provide the first comprehensive comparative evaluation of non-learning, statistical learning, and deep learning models. We show that learning-based methods clearly outperform non-learning models, with graph neural networks achieving the best mean performance, while the strongest learning-based models remain relatively close to one another. Beyond aggregate mean absolute error (MAE) and root mean squared error (RMSE), the framework also provides breakdowns by prediction horizon and delay change, enabling more detailed analysis of model behavior across forecasting regimes.

评分理由: 论文专注于铁路延误预测的数据集基准与深度学习模型比较，涉及图神经网络和时间序列预测。提供的关键词主要涉及多模态大模型、世界模型及强化学习领域，与论文内容高度不相关。除'统一评估协议'与'Unify Models'有微弱关联，'多源数据'与'MultiModal'有极微弱关联外，其余关键词在文中均未涉及。作者列表中不包含指定的专家。

Train Delay Prediction, Dataset Benchmark, Graph Neural Networks, Railway Network, Weather Records, Evaluation Protocol, Deep Learning, Time Series Forecasting

本文提出了一种用于识别仿射控制（Control-affine）降阶模型（ROMs）的框架。所提出的方法利用自编码器（AEs）将高维状态（以及潜在的高维输入）转换为适合仿射控制状态空间动力学（State-space Dynamics）的低维潜在表示。这是通过对自编码器和状态空间模型（State-space Model）进行联合训练实现的。此外，我们将离散降阶模型（ROM）的形式扩展为基于序列的模型（Sequence-based Model），该模型处理状态和输入历史以提高预测精度，同时保持仿射控制结构。我们通过将反馈线性化（Feedback Linearization）应用于推导出的模型来论证该框架，并提出了其高效使用的指南。所提出的框架在两个数值算例上进行了评估，其性能与一个基线模型（Baseline Model）进行了比较，在该基线模型中，自编码器识别出一个具有线性状态空间动力学的潜在空间。评估包括评估降阶模型（ROM）在测试数据上的预测精度及其将系统控制到期望状态或轨迹的有效性。

We present in this paper a framework for the identification of control-affine reduced-order models (ROMs). The proposed method utilizes autoencoders (AEs) to transform the high-dimensional states, and potentially the high-dimensional inputs, into reduced latent ones suitable for control-affine state-space dynamics. This is achieved by simultaneous training of the AE and the state-space model. In addition, we extend the discrete ROM formulation to a sequence-based model, which processes state and input histories to improve prediction accuracy while preserving the control-affine structure. We motivate our framework by applying feedback linearization to the derived models, and we present guidelines for its efficient use. The proposed framework is assessed on two numerical examples and its performance is compared to a baseline model, where the AE identifies a latent space with linear state-space dynamics. The assessment involves evaluating the prediction accuracy of the ROM on test data and its effectiveness in controlling the system to a desired state or trajectory.

评分理由: 论文核心为控制理论中的系统辨识与降阶建模，使用自编码器处理连续状态空间。提供的关键词主要聚焦于多模态大模型、语言模型及强化学习领域（如 Tokenizer, MLLM, MultiModal），与本文内容领域不符，故相关性为 0。仅'model-based RL'因涉及基于模型的控制策略有微弱关联，故给予低分。作者列表中不包含指定的专家，无额外加分。加权总分约为 4.5，低于动态及格分 27.8。

Autoencoders, Reduced-Order Models, Control-Affine, System Identification, State-Space Model, Feedback Linearization, Latent Space

推理模型发展迅速，但主流的基于可验证奖励的强化学习（RLVR）范式却出人意料地狭隘：即采样大量响应，并用单个比特奖励每个响应，以指示最终答案是否正确。然而，许多场景提供丰富的反馈，包括执行轨迹、工具输出、专家修正以及模型的自我评估。我们通过经典模仿学习算法 DAgger（Dataset Aggregation）的分布变体来研究如何利用此类反馈，其中学习者可局部访问当前策略所访问状态上的专家分布。这导出了一个简单的前向交叉熵目标函数，该函数支持黑盒专家，且其序列级梯度通过传播未来的专家 - 学生分歧回早期决策，从而执行丰富的信用分配。我们表明，先前基于反向 KL 或 Jensen-Shannon 散度的自蒸馏目标的强化学习无法保证单调策略改进：即使专家具有更高奖励，其更新也可能增加较差动作的概率。相比之下，我们证明前向交叉熵支持单调策略改进，并具备遗憾保证。此外，我们进一步表明，我们的目标优化了教师加权成功似然性的下界，从而提升了 Pass@N 指标。实验上，我们的方法 DistIL 在科学推理、代码生成及求解复杂数学问题等多种领域上，均优于 RLVR 及基于自蒸馏的强化学习基线方法。

Reasoning models have advanced rapidly, but the dominant reinforcement learning from verifiable rewards (RLVR) recipe remains surprisingly narrow: sample many responses and reward each with a single bit indicating whether the final answer is correct. Yet many settings provide rich feedback, including execution traces, tool outputs, expert corrections, and model self-evaluations. We study how to use such feedback through a distributional variant of the classic imitation learning algorithm DAgger, where the learner has local access to an expert distribution on states visited by the current policy. This yields a simple forward cross-entropy objective that admits a blackbox expert and whose sequence-level gradient {conduct rich credit assignment by propagating} future expert-student disagreement back to earlier decisions. We show that prior RL with self-distillation objectives based on reverse KL or Jensen-Shannon fail to guarantee monotonic policy improvement: even when the expert has higher reward, their updates may increase probability on worse actions. In contrast, we show that forward cross-entropy admits monotonic policy improvement and enjoys guarantees on regret. We further show that our objective optimizes a lower bound on teacher-weighted likelihood of success, leading to improved Pass@N. Empirically, our approach, DistIL, improves over RLVR and RL with self-distillation baselines across a variety of domains: scientific reasoning, coding, and solving hard mathematical problems.

评分理由: The paper focuses on Reinforcement Learning from Rich Feedback using a distributional DAgger variant, emphasizing policy improvement and forward cross-entropy objectives. It does not involve multimodal architectures, visual encoders, tokenizers, world models, or model unification, resulting in low relevance for most keywords. While it is an RL paper, it is not specifically model-based RL (which typically involves learning dynamics models), hence the moderate score for that keyword. The total weighted score (3.0) is below the dynamic passing threshold (27.8), indicating poor alignment with the provided keyword set. None of the listed expert authors are present in the author list.

Reinforcement Learning, Rich Feedback, Distributional DAgger, Forward Cross-Entropy, Policy Improvement, Scientific Reasoning, Pass@N

大型语言模型（LLM）的知识基准面临三个问题：规模驱动的设计未能实现学科代表性；允许懒惰共识的扁平支付注释；以及在有界测试预算下未经审计的排名不稳定性。我们提出了 KINA，这是一个涵盖 261 个细粒度学科的 899 项基准，包含两个正式结果。首先，我们将代表性视为基于专家提取锚点的覆盖式目标，并通过一个代理变量操作化学科代表性，从而得到一个 (1-1/e) 的贪婪近似（命题 1）；该保证仅适用于代理变量，而非总体代表性。其次，我们证明基于条形奖金的锦标赛在发布评论质量上弱一阶随机占优（FOSD）支配扁平支付，其激励相容阈值为 B > ΔC / Δp_min（定理 1）。评估来自 13 个实验室的 42 个模型后，表现最佳的模型是 Gemini-3.1-Pro-Preview，得分为 53.17%，其次是 Claude-Opus-4.6（49.92%）和 GPT-5.4（48.55%），表明在饱和点以下仍有较大的提升空间。完整的排行榜显示出层级结构而非平滑的全序：一个较小的前沿层级位于 48% 以上，一个密集的强模型层级跨度约为 38%-45%，而表现较差的模型仅略高于 10% 的机会基线。工具增强在五个工具使用评估中最多可增加 5.17 分，且增益在不同模型间差异显著。我们报告了基于自助法的排名稳定性统计，以明确有界预算下的方差，并防止对相邻排名的过度解读。

Knowledge benchmarks for LLMs face three issues: scaling-driven designs that do not operationalize disciplinary representativeness; flat-payment annotation that permits lazy consensus; and unaudited ranking instability under bounded test budgets. We introduce KINA, an 899-item benchmark across 261 fine-grained disciplines, with two formal results. First, we cast representativeness as a coverage-style objective over expert-elicited anchors and operationalize disciplinary representativeness through a proxy, yielding a (1-1/e) greedy approximation (Proposition 1); the guarantee applies to the proxy, not to population representativeness. Second, we prove a bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality, with incentive-compatibility threshold B > Delta C / Delta p_min (Theorem 1). Evaluating 42 models from 13 labs, the top model, Gemini-3.1-Pro-Preview, reaches 53.17%, followed by Claude-Opus-4.6 at 49.92% and GPT-5.4 at 48.55%, leaving substantial headroom below saturation. The full leaderboard shows a tiered structure rather than a smooth total order: a small frontier tier lies above 48%, a dense strong-model tier spans roughly 38-45%, and low-performing models remain only modestly above the 10% chance baseline. Tool augmentation adds up to 5.17 points across the five tool-use evaluations, with gains varying substantially across models. We report bootstrap ranking-stability statistics to make bounded-budget variance explicit and to discourage over-interpretation of adjacent ranks.

评分理由: The paper proposes KINA, a knowledge benchmark for LLMs focusing on disciplinary representativeness, annotation incentives, and ranking stability. It does not discuss model architecture (Tokenizer, Visual Encoder), unification strategies, world modeling, or reinforcement learning. While it evaluates models like Gemini (potentially MLLMs), the core contribution is evaluation methodology rather than multimodal representation learning or model training. Thus, relevance to the specific technical keywords is minimal. None of the listed expert authors appear in the author list.

Knowledge Benchmark, LLM Evaluation, Disciplinary Representativeness, Incentive Compatibility, Ranking Stability, Tool Augmentation, Expert-Elicited Anchors

研究论文的标题以清晰简洁的方式传达其主要思想，偶尔也会传达其结论。选择合适的标题往往具有挑战性，而自动生成标题可协助作者完成这一任务。在这项工作中，我们提出了一种利用开源预训练大语言模型从摘要生成论文标题的技术。我们使用了 CSPubSum 和 LREC-COLING-2024 数据集，并引入了一个新的数据集 SpringerSSAT，该数据集是从四本社会科学领域的 Springer 期刊中整理而成的。此外，我们在零样本设置中使用 GPT-3.5-turbo 生成标题。模型性能使用 ROUGE、METEOR、MoverScore、BERTScore 和 SciBERTScore 指标进行评估。我们的实验表明，微调后的 PEGASUS-large 在大多数指标上优于其他模型，包括微调后的 LLaMA-3-8B 和零样本 GPT-3.5-turbo。我们进一步展示了 ChatGPT 可以生成创造性的论文标题。总体而言，AI 生成的标题通常是合适且可靠的。

The title of a research paper conveys its primary idea and, occasionally, its conclusions in a clear and concise manner. Choosing an appropriate title is often challenging, and automated title generation can assist authors in this task. In this work, we propose a technique to generate paper titles from abstracts using open-weight pre-trained and large language models. We use the CSPubSum and LREC-COLING-2024 datasets and introduce a new dataset, SpringerSSAT, curated from four Springer journals in the social sciences. Additionally, we use GPT-3.5-turbo in a zero-shot setting to generate titles. Model performance is evaluated with ROUGE, METEOR, MoverScore, BERTScore, and SciBERTScore metrics. Our experiments show that fine-tuned PEGASUS-large outperforms other models, including fine-tuned LLaMA-3-8B and zero-shot GPT-3.5-turbo, across most metrics. We further demonstrate that ChatGPT can generate creative paper titles. Overall, AI-generated titles are generally appropriate and reliable.

评分理由: The paper focuses on text-based automatic title generation using LLMs (PEGASUS, LLaMA, GPT) and does not involve multimodal learning, world models, visual encoders, or reinforcement learning. Only minimal relevance exists for Tokenizer (implicit in LLMs) and Unify Models (comparison of different models), while all other keywords are completely unrelated to the paper's content.

Automatic Generation, Research Papers, Language Models, Title Generation, Text Generation, PEGASUS-large, Abstract-to-Title, Evaluation Metrics

编程领域的 AI 工具已不再仅仅是自动补全或聊天助手；它们正组织成开发框架，涵盖流程、角色、工件及验证环节。近期的调查已梳理了软件工程中的智能体（Agents）与大语言模型（LLMs），但尚缺乏一项专注于将此类能力转化为流程的操作框架的研究。我们对原始资料进行了定向搜索，基于功能性纳入标准及影响力度量，筛选出六个框架：GitHub Spec Kit、OpenSpec、BMAD Method、Get Shit Done (GSD)、Spec Kitty 和 Reversa。各框架通过不同路径切入 AI 开发：包括完整与轻量级变体的规范驱动开发、基于智能体的敏捷规划、针对智能体的上下文工程、工作树隔离与审查，以及从遗留系统中恢复操作规范。本文的核心贡献在于提出了一种六维过程分类法：规范（Specification）、上下文（Context）、角色（Roles）、执行（Execution）、验证（Validation）与可移植性（Portability），并附带一套评分量表，使其成为一种可复现的评估工具。我们将该分类法应用于上述六个框架以及一个样本外案例（Spec-Flow）。研究结果主要有两点。在已采用部分流程的框架中存在趋同现象：孤立提示（isolated prompt）的中心地位减弱，持久性工件、工作契约、可追溯性及人工审查成为减少歧义并协调智能体的机制。且没有任何一个框架能全面覆盖所有六个维度，这揭示了过程深度与跨智能体可移植性之间存在结构性权衡。此外，我们还发现了反复出现的风险：规范与代码之间的漂移、对生成工件的过度信任、社区扩展的脆弱性、平台依赖性以及缺乏针对完整流程的基准。最后，我们提出了一项实证评估的研究议程，重点关注中间质量指标、上下文治理、安装安全性及可复现性。

AI tools for programming are no longer just autocomplete or chat assistants: they organize themselves as development frameworks, with process, roles, artifacts and verification. Recent surveys map agents and LLMs for software engineering, but a study centered on the operational frameworks that turn these capabilities into process is missing. We ran a directed search of primary sources, with a functional inclusion criterion and traction measurement, and selected six frameworks: GitHub Spec Kit, OpenSpec, BMAD Method, Get Shit Done (GSD), Spec Kitty and Reversa. Each attacks AI development through a different path: spec-driven development in full and lightweight variants, agent-driven agile planning, context engineering over the agent, worktree isolation and review, and recovery of operational specifications from legacy systems. Our central contribution is a six-dimension process taxonomy: specification, context, roles, execution, validation and portability, with a scoring rubric that turns it into a replicable instrument. We apply it to the six frameworks and an out-of-sample case, Spec-Flow. Two results stand out. Among frameworks that already adopt some process there is convergence: the isolated prompt loses centrality, and persistent artifacts, work contracts, traceability and human review become mechanisms that reduce ambiguity and coordinate agents. And no framework strongly covers all six dimensions, exposing a structural trade-off between process depth and portability across agents. We also found recurring risks: drift between specification and code, excessive trust in generated artifacts, fragility of community extensions, platform dependence and a lack of benchmarks for the complete process. We close with a research agenda for empirical evaluation, focused on intermediate-quality metrics, context governance, installation security and reproducibility.

评分理由: The paper focuses on process taxonomy for AI software development agents, lacking content on multimodal architectures (Tokenizer, Visual Encoder), world models, or reinforcement learning. While it utilizes LLMs, it does not address Multimodal LLMs (MLLM) or model unification in representation learning, resulting in low relevance scores.

AI Software Development Agents, Process Taxonomy, Comparative Assessment, Software Engineering Frameworks, Specification, Roles, Execution, Portability

我们引入大语言模型代理架构 Agentic Redux，旨在用于需要线性可审计性的非平凡问题域。利用类型化 Lambda 演算，我们证明，在适用域上运行时，Agentic Redux 的执行在语义上保证正确，所有决策均记录于追加式账本中。我们展示了两个生产级适用领域，涵盖医疗账单合规性和安全漏洞披露。在两个领域上运行的 Agentic Redux 的可运行代码可在支持代码库中获取。我们还引入了本体优先代理设计（Ontology-First Agent Design），这是一种在问题域上创建代理框架的方法论：人类专家利用基本形式本体（Basic Formal Ontology）对该问题域进行本体化，然后指派大语言模型（LLM）推导出代理和人在回路（humans-in-the-loop）可承担的角色，以解决该领域的问题。

We introduce the LLM agent architecture Agentic Redux, intended for use with nontrivial problem domains that require linear auditability. Using the typed lambda calculus, we prove that, run on appropriate domains, Agentic Redux executions are semantically guaranteed to be correct, with all decisions recorded in an append-only ledger. We present two production-grade appropriate domains, in healthcare billing compliance, and security vulnerability disclosure. Working code for Agentic Redux run on both domains is available in a supporting code repository. We also introduce Ontology-First Agent Design, a methodology for creation of agent frameworks on a problem domain, in which a human expert ontologizes the problem domain with Basic Formal Ontology, and then assigns an LLM to derive roles that agents and humans-in-the-loop can fill, in order to work the problems in the domain.

评分理由: The paper focuses on LLM agent safety, auditability, and ontology-based design using formal verification (typed lambda calculus). It does not address multi-modality (Visual Encoder, MultiModal), tokenization strategies, world models, or model-based reinforcement learning. While it involves LLMs (MLLM), it lacks the multi-modal components implied by the keyword set, resulting in low relevance scores for most keywords.

LLM Agents, Auditable, Safe, Human-Authored Ontologies, Agentic Redux, Typed Lambda Calculus, Ontology-First Agent Design

寻找组合谜题（如 Rubik's Cube、滑动拼图和 Lights Out）的最优解路径仍然是人工智能领域的一个经典挑战。启发式搜索算法（如 A*）仅在采用可容许启发式函数（即从不高估真实剩余代价的函数）时，才能保证路径的最优性。深度强化学习（RL）方法（如 DeepCubeA）通过训练深度神经网络来近似剩余代价启发式函数。然而，标准的均方误差（MSE）训练通常会产生高估，违反可容许性并损害解的最优性。本文提出了一种通用框架，用于学习基于验证校准的可容许神经启发式函数。我们利用一个低估型的可容许贝尔曼算子结合非对称损失函数来惩罚高估，从而训练价值网络。为了考虑残留的神经网络函数近似误差，我们提出了一种事后校准安全偏移量，该偏移量基于验证打乱状态计算得出。实验表明，在评估协议下，我们的校准神经启发式函数未观察到任何可容许性违反，并在实践中保持路径最优性；与标准解析基线相比，其在 2×2 Rubik's Cube 上的搜索节点扩展减少了高达 83.0%，在 3×3 Lights Out 网格上减少了 19.9%，在 8-Puzzle 上减少了 1.9%。

Finding optimal solution paths for combinatorial puzzles like the Rubik's Cube, sliding tile puzzles, and Lights Out remains a classical challenge in artificial intelligence. Heuristic search algorithms, such as A* , guarantee path optimality only when using an admissible heuristic-one that never overestimates the true remaining cost-to-go. Deep reinforcement learning (RL) methods like DeepCubeA train deep neural networks to approximate cost-to-go heuristics. However, standard mean-squared error (MSE) training regularly yields overestimations, violating admissibility and compromising solution optimality. In this paper, we introduce a generalizable framework for learning validation-calibrated admissible neural heuristics. We train a value network using an underestimating Admissible Bellman Operator combined with an Asymmetric Loss function to penalize overestimation. To account for residual neural function approximation errors, we propose a post-hoc calibration safety offset computed over validation scrambles. We demonstrate that our calibrated neural heuristics achieve no observed admissibility violations under the evaluation protocol and preserve path optimality in practice while reducing search node expansions by up to 83.0% on a 2 by 2 Rubik's Cube, 19.9% on a 3 by 3 Lights Out grid, and 1.9% on an 8-Puzzle compared to standard analytical baselines.

评分理由: The paper focuses on learning admissible neural heuristics for combinatorial search (e.g., Rubik's Cube) using RL concepts like the Bellman operator. It does not address Unify Models, Tokenizers, Visual Encoders, World Models, MLLM, or MultiModal aspects, resulting in 0 relevance for these keywords. There is a slight connection to model-based RL due to RL terminology usage, but the core is heuristic search rather than model-based planning. No expert authors from the specified list were found, so no bonus applies.

Combinatorial Search, Admissible Heuristics, Neural Heuristics, Deep Reinforcement Learning, Bellman Operator, Path Optimality, Rubik's Cube

高通量测序技术的快速发展催生了大型、高维组学数据集。深度无监督学习架构，尤其是自编码器（Autoencoders, AEs），在该领域中日益被广泛用于降维和表示学习。然而，自编码器对架构选择和超参数高度敏感，且无监督优化通常依赖于重建损失，这作为下游任务性能的代理指标可能效果不佳。穷举超参数优化（HPO）计算成本高昂，致使研究人员常依赖次优的默认配置。为了降低大规模无监督 HPO 研究的门槛，我们引入了 BBOmix，这是首个面向真实世界生物数据的无监督表示学习的开源表格基准。我们的基准涵盖了来自 TCGA 和 SCHC 数据集的四种 AE 架构及七种多组学模态的 105,000 次评估。我们量化了重建损失与下游任务性能之间的相关性，并对最先进的单保真度、多保真度及迁移学习 HPO 方法进行了广泛评估，为未来的无监督生物表示学习研究建立了严格的基线。

The rapid advancement of high-throughput sequencing has led to large, high-dimensional omics datasets. Deep unsupervised learning architectures, particularly Autoencoders (AEs), are increasingly used for dimensionality reduction and representation learning in this domain. However, AEs are highly sensitive to architectural choices and hyperparameters, and unsupervised optimization typically relies on reconstruction loss, which may be a poor proxy for downstream utility. Exhaustive hyperparameter optimization (HPO) is computationally expensive, leading researchers to frequently rely on suboptimal default configurations. To democratize access to large-scale unsupervised HPO research, we introduce $\textbf{BBOmix}$, the first open-source tabular benchmark for unsupervised representation learning on real-world biological data. Our benchmark includes 105,000 evaluations across four AE architectures and seven multi-omics modalities from the TCGA and SCHC datasets. We quantify the correlation between reconstruction loss and downstream task performance and provide an extensive evaluation of state-of-the-art single-fidelity, multi-fidelity, and transfer learning HPO methods, establishing a rigorous baseline for future research in unsupervised biological representation learning.

评分理由: 论文核心为生物组学数据的无监督表示学习（自编码器）与超参数优化，属于生物信息学领域。提供的关键词主要指向多模态大模型（MLLM）、视觉编码器、世界模型及强化学习，技术栈不匹配。仅因涉及‘多组学模态’与 MultiModal 有微弱语义关联，其余关键词完全无关。

Unsupervised Representation Learning, Hyperparameter Optimization, Autoencoders, Multi-omics, Benchmark, Biological Data, Dimensionality Reduction, Reconstruction Loss

生成逼真的金融时间序列颇具挑战性，因为训练数据往往仅限于单一历史路径。面对如此稀缺的数据，过拟合难以避免，尤其是在对抗训练场景下，训练好的判别器可能会记住训练样本。为缓解这一问题，近期方法通过训练生成器来最小化真实与生成时间序列的未经训练特征表示之间的差异。在这些工作中，特征图基于路径签名（path signatures），这在可行的截断深度下可能无法捕捉相关的时间序列属性。本文则通过匹配真实与生成时间序列的随机卷积特征来训练生成器。现有的随机卷积特征映射（如 Rocket 和 Hydra）已被证明能提供关于真实世界时间序列的富有信息量的表示，但由于它们不可微，无法用于监督生成模型。我们提出 SOCK（SOft Competing Kernels），这是一种完全可微的随机卷积特征映射，适用于训练生成式时间序列模型。实验表明，通过匹配随机 SOCK 特征训练的生成器在广泛的小样本金融数据集上一致优于基于签名和扩散的基线模型。此外，我们还展示了 SOCK 在双样本假设检验和时间序列分类任务上的表达能力，在这些任务中，SOCK 匹配或优于现有的无监督特征映射。

Generating realistic financial time series is challenging as training data is often limited to a single historical path. With such scarce data, overfitting is hard to avoid, especially under adversarial training where a trained discriminator can memorize the training samples. To mitigate this, recent approaches train generators to minimize the discrepancy between untrained feature representations of real and generated time series. In these works, the feature maps are based on path signatures, which can fail to capture relevant time series properties at tractable truncation depths. In this work, we instead train generators by matching random convolutional features of real and generated time series. Existing random convolutional feature maps, such as Rocket and Hydra, have been shown to provide informative representations of real-world time series, but cannot supervise generative models because they are non-differentiable. We introduce SOCK (SOft Competing Kernels), a fully differentiable random convolutional feature map, suited to train generative time series models. We show that generators trained by matching random SOCK features consistently outperform signature and diffusion baselines across a wide range of small-sample financial datasets. We further demonstrate SOCK's expressiveness on two-sample hypothesis testing and time series classification tasks, where SOCK matches or outperforms existing unsupervised feature maps.

评分理由: 论文核心为金融时间序列生成，采用可微分随机卷积特征匹配方法（SOCK）。给定关键词主要涉及多模态大模型（MLLM, MultiModal）、视觉编码（Visual Encoder）、序列标记（Tokenizer）及强化学习（model-based RL）等领域，与本文的时间序列生成任务无直接技术关联。因此，除“世界模型”和“统一模型”因涉及生成建模有微弱关联外，其余关键词相关性均为 0。加权总分远低于动态及格分 27.8，表明该论文不属于该特定研究方向的范畴。

Financial Time Series, Generative Models, Random Convolutional Features, Differentiable Feature Matching, SOCK, Small-sample Learning, Two-sample Hypothesis Testing

缺失值插补（Missing Value Imputation）是机器学习中的基础任务，大多数现有方法假设所有缺失条目都对应于未观测到的常规值。然而，在许多真实世界的数据集中，缺失性（Missingness）可能源于两个不同的来源：一些条目是有意义缺失的（本质上缺失且语义有效），而其他条目是由于观测过程缺失且应被插补的。我们将这种区别形式化为一个选择性插补问题（Selective Imputation Problem），其目标是联合推断哪些缺失条目应被保留，哪些应被恢复。为了解决这一挑战，我们提出 Diff-Joint，一种基于扩散的框架（Diffusion-based Framework），联合建模表格数据与潜在缺失性掩码（Latent Missingness Mask）。该方法在条件采样（Conditional Sampling）和不确定性感知聚合（Uncertainty-aware Aggregation）之间交替进行，以迭代地优化插补值和缺失性标签。在合成数据集和真实世界数据集上的实证结果表明，Diff-Joint 能有效识别有意义缺失条目，同时实现具有竞争力的插补精度并提升下游任务性能。

Missing value imputation is a fundamental task in machine learning, with most existing methods assuming that all missing entries correspond to unobserved regular values. In many real-world datasets, however, missingness may arise from two distinct sources: some entries are meaningfully missing (intrinsically absent and semantically valid), while others are missing due to the observation process and should be imputed. We formalize this distinction as a selective imputation problem, where the goal is to jointly infer which missing entries should be preserved and which should be recovered. To address this challenge, we propose Diff-Joint, a diffusion-based framework that jointly models tabular data together with a latent missingness mask. The method alternates between conditional sampling and uncertainty-aware aggregation to iteratively refine both imputed values and missingness labels. Empirical results on synthetic and real-world datasets demonstrate that Diff-Joint effectively identifies meaningfully missing entries while achieving competitive imputation accuracy and improved downstream task performance.

评分理由: The paper focuses on tabular data missing value imputation using a diffusion framework, distinguishing meaningful missingness from observation missingness. It does not involve tokenizers, visual encoders, multimodal architectures, MLLMs, or reinforcement learning. There is a minor conceptual overlap with 'Unify Models' due to joint modeling of data and masks, but overall relevance to the provided keyword set is low.

Missing value imputation, Meaningful missingness, Diffusion framework, Tabular data, Uncertainty-aware, Latent missingness mask, Selective imputation

深度图生成的研究范畴涵盖了两个极端：一次性模型（one-shot models）与序列模型（sequential models）。前者联合生成节点和边，而后者则自回归地（autoregressively）采样它们。每种方法在基于大小和拓扑结构的不同图领域表现各异，但均无法适用于所有图类别。例如，一次性方法在生成大型图时面临挑战，而序列方法在小型图上的表现则欠佳。克服这些限制的一种可行方法是在一个统一系统中灵活地结合这两种方法。在本文中，我们提出了 FLAGG（Flexible Autoregressive Graph Generation）框架，该框架利用一次性模型顺序生成图的部分子图。FLAGG 可将任意一次性模型转化为自回归模式，从而在选择序列策略时提供灵活性。该策略通过随机节点移除过程来定义，一个插入模型（Insertion Model）学习逆转该过程。我们在多个具有不同图大小和领域的数据集上，使用 DiGress 一次性模型对 FLAGG 进行了评估。结果表明，该方法在采样质量上优于一次性模型和自回归基线（baselines）方法。

The Deep Graph Generation's panorama spans two extremes: one-shot and sequential models. The former generates nodes and edges jointly, while the latter samples them autoregressively. Each method performs better in different graph domains depending on size and topology, but neither is applicable to all graph categories. For instance, one-shot methods struggle with generating large graphs, while sequential methods underperform on smaller graphs. A possible way to overcome these limitations is to flexibly combine the two methods in a unique system. In this work, we propose the FLAGG (Flexible Autoregressive Graph Generation) framework, which sequentially generates portions of graphs with one-shot models. FLAGG can apply any one-shot model to make it autoregressive, allowing flexibility in choosing the sequential policy. This policy is specified through a stochastic node removal process, which an Insertion Model learns to reverse. We evaluate FLAGG with the DiGress one-shot model on several data sets of different graph sizes and domains. We show that the approach outperforms both one-shot and autoregressive baselines in terms of sampling quality.

评分理由: 论文主题聚焦于图生成（Graph Generation），提出 FLAGG 框架融合一阶与自回归生成策略。提供的关键词主要涉及多模态大模型（MLLM, MultiModal）、视觉编码器、Tokenizer 及模型强化学习，与论文领域（图结构数据）存在显著差异。仅 Unify Models 因涉及策略融合获得微弱相关性，其余关键词完全不相关。

Graph Generation, Autoregressive, One-shot, Flexible Framework, Node Removal, Sampling Quality, DiGress

离散图模型中的边缘推断迫使我们在精确性与可扩展性之间做出选择：精确算法在高树宽图上不可行，而迭代近似方法（Belief Propagation, 变分方法）在受阻拓扑上牺牲了收敛保证。我们认为这种对立源于不匹配的归纳偏置：迭代方法放弃了使精确推断正确的顺序消除结构。我们提出了一种 In-Context Graphical Inference（ICG-I），这是一种自回归 Graph Transformer，通过模仿 Variable Elimination，利用学习到的、经 Tensor-Train（TT）压缩的中间因子来恢复这种结构，并配合 Dirichlet 输出层和 Weighted Conformal Prediction（WCP），以在拓扑偏移下提供校准的、无分布的覆盖保证。我们证明了 TT 压缩误差在自回归链中最多线性传播，Dirichlet-Multinomial 损失是一个真评分规则，且 WCP 在估计密度比下保持覆盖性，尽管会有可量化的退化。我们进行了详尽的实验来评估 ICG-I，并在所有基准上实现了最先进性能。ICG-I 在标准实例上将 MAE 从 0.041（最佳基线）降低到 0.020，并在 N=500 的受阻自旋玻璃上取得了 0.048 的结果，此时 Belief Propagation（BP）完全发散。

Marginal inference in discrete graphical models forces a choice between exactness and scalability: exact algorithms are intractable for high-treewidth graphs, while iterative approximations (Belief Propagation, variational methods) sacrifice convergence guarantees on frustrated topologies. We argue that this dichotomy stems from a mismatched inductive bias: iterative methods abandon the sequential elimination structure that makes exact inference correct. We introduce In-Context Graphical Inference (ICG-I), an autoregressive Graph Transformer that restores this structure by mimicking Variable Elimination with learned, Tensor- Train-compressed intermediate factors, paired with a Dirichlet output layer and Weighted Conformal Prediction for calibrated, distribution-free coverage guarantees under topological shift. We prove that TT compression errors propagate at most lincarly through the autoregressive chain, that the Dirichlet-Multinomial loss is a proper scoring rule, and that WCP maintains coverage with a quantifiable degradation under estimated density ratios. We conducted intensive experiments to evaluate ICG-I and achieved state-of-the-art performance across all benchmarks. ICG-I reduces MAE from 0.041 (best baseline) to 0.020 on standard instances and achieves 0.048 on N=500 frustrated spin glasses where BP diverges entirely.

评分理由: 论文主要关注概率图模型中的推断算法（如变量消除、图变换器），而提供的关键词集主要围绕多模态大模型、世界模型和强化学习。两者领域差异较大。仅在'Unify Models'上略有关联（统一了精确与近似推断结构），其余关键词如 Tokenizer、Visual Encoder、MLLM 等均无涉及。未发现指定专家作者。

Graphical Inference, Variable Elimination, Graph Transformer, Tensor-Train Compression, Marginal Inference, Dirichlet Output, Weighted Conformal Prediction, Autoregressive

大型语言模型（LLM）越来越多地被提议作为临床智能体，然而静态、单轮基准无法捕捉模型如何在一次诊疗过程中动态提供护理：收集信息、规划治疗，并在连续的患者状态中调整纵向管理。医学教育长期以来通过标准化病人（SP）解决了一个类似挑战：训练有素的演员始终扮演临床病例，从而实现现实化的练习和客观的脚本化评估。本文介绍了 MedSP1000，一种基于 SP 的临床智能体评估交互式基准，包含 1,638 个 SP 病例和 24,602 个轨迹级同行评审评分标准。MedSP1000 将同行评审的 SP 教学案例转换为可执行场景，包含定义的 SP 病例脚本、临床环境背景以及经人工验证的结构化评分标准。在每次模拟评估运行中，临床智能体与患者智能体及环境控制器进行闭环交互，其整个诊疗过程中的行为表现均根据原始材料中指定的专家标准进行评分。将 MedSP1000 应用于一系列通用型和医学专用 LLM 时，我们发现静态基准上的表现无法可靠地转化为此类教育场景。表现最佳的模型 GPT-5.5 仅完成了 60.4% 的专家定义评分标准项，而最强的医学专用模型仅达到 40.0%；增加推理时计算量并未带来可测量的提升。这些结果表明，当前的 LLM 包括针对医学调优的智能体系统，其可靠性尚不足以安全地整合到实际临床实践中。更广泛而言，MedSP1000 展示了过程级、SP 风格的评估如何揭示单轮基准所遗漏的临床相关故障模式。

Large language models (LLMs) are increasingly proposed as clinical agents, yet static, single-turn benchmarks cannot capture how a model dynamically delivers care across an encounter: gathering information, planning treatment, and adapting longitudinal management across successive patient states. Medical education has long addressed an analogous challenge through standardized patients (SPs): trained actors who consistently portray clinical cases, enabling realistic practice and objective, scripted assessment. Here we introduce MedSP1000, an SP-derived interactive benchmark for clinical-agent evaluation, including 1,638 SP cases with 24,602 trajectory-level peer-reviewed rubrics. MedSP1000 converts peer-reviewed SP teaching cases into executable scenarios with defined SP case scripts, clinical environment contexts, and human-validated structured rubric. In each simulation evaluation run, a clinical agent interacts in closed loop with a patient agent and an environment controller, and its behaviour is scored throughout the encounter against expert criteria specified in the original materials. Applying MedSP1000 to a range of general-purpose and medically specialized LLMs, we find that performance on static benchmarks does not reliably translate to such educational scenarios. The best-performing model, GPT-5.5, completes only 60.4% of expert-defined rubric items, whereas the strongest medically specialized model reaches 40.0%; increasing test-time compute produces no measurable gain. These results suggest that current LLMs, including agentic systems tuned for medicine, are not yet reliable enough to be safely integrated into actual clinical practice. More broadly, MedSP1000 shows how process-level, SP-style evaluation can reveal clinically relevant failure modes that single-turn benchmarks miss.

评分理由: The paper focuses on evaluating LLMs in clinical decision-making using standardized patients (MedSP1000), lacking technical content on unification, tokenization, visual encoders, world models, or model-based RL algorithms. It involves LLMs (weak link to MLLM) and agent-environment loops (weak link to RL), but does not address the specific multimodal or architectural aspects implied by the keywords. No expert authors from the target list were found.

Large Language Models, Clinical Decision-Making, Standardized Patients, MedSP1000, Dynamic Evaluation, Closed Loop, Benchmark, Agent Evaluation

大型语言模型（LLMs）正日益广泛应用于生成 Lean 形式化证明的工作流中。这些工作流通常将问题分解为更小的引理，采样大量证明尝试，并利用编译器反馈来引导搜索过程。然而，这些方法可能成本过高，往往在最终失败的尝试上消耗大量的计算资源。本文提出了一种动作路由代理（action routing agent）来解决这一问题，该代理由数据平面（data plane）和控制平面（control plane）组成。数据平面负责生成自然语言的引理分解，在 Lean 中将其形式化，并为生成的定理和引理目标采样证明尝试。控制平面观察先前失败的 Lean 尝试，估计成功的可能性以及另一次尝试的成本，并决定是继续证明当前目标还是从新的分解重新开始。在 PutnamBench 的一个子集上，我们的代理相较于固定步骤基线（fixed-step baseline）平均降低了 $25.8\%$ 的成本，在保持性能的同时显著减少了计算资源的消耗。这些结果表明，失败的 Lean 轨迹为基于代理的形式证明（agentic theorem proving）中成本感知的资源分配提供了可操作的信号。

Large language models (LLMs) are increasingly used in workflows for generating formal proofs in Lean. These workflows often decompose problems into smaller lemmas, sample many proof attempts, and use compiler feedback to guide search. However, they can be prohibitively expensive, often spending substantial compute on attempts that ultimately fail. In this work, we address this problem with an action routing agent that consists of a data plane and a control plane. The data plane generates natural-language lemma decompositions, formalizes them in Lean, and samples proof attempts for the resulting theorem and lemma targets. The control plane observes previous failed Lean attempts, estimates both the likelihood of success and cost of another attempt, and decides whether to continue proving the current target or restart from a new breakdown. On a subset of PutnamBench, our agent decreases the cost by $25.8\%$ over a fixed-step baseline on average, preserving performance while using substantially less compute. These results suggest that failed Lean trajectories provide actionable signals for cost-aware resource allocation in agentic theorem proving.

评分理由: 该论文研究 Lean 语言中代理定理证明的成本 - 质量优化，核心在于 LLM 工作流调度与决策。提供的关键词集（如视觉编码器、多模态、世界模型）主要面向多模态感知与表征学习，与本文的文本/代码形式化验证任务无直接关联。虽然控制平面涉及基于模型的决策（与 model-based RL 有微弱概念交集），但论文未涉及多模态输入、世界模型构建或模型统一架构，因此整体相关性极低。

Agentic Theorem Provers, Lean, Cost-Quality Tradeoff, Action Routing Agent, Control Plane, Data Plane, Formal Proofs, LLMs

采用梯度下降算法训练的径向基函数神经网络 (RBFN) 在浅层和深层网络中均能提供有效的全连接结构。误差修正 (ErrCor) 作为一种最先进的基于梯度的训练方法，通过选择最优隐藏单元来提高准确性。另一方面，作为一种基于种群的算法，粒子群优化算法 (PSO) 利用群体经验来优化 RBFN 参数，具备全局搜索能力以及对局部极小值的鲁棒性。自适应粒子群优化算法 (APSO) 已成为 PSO 的一种改进变体。该 APSO 算法通过在优化过程中动态调整群体参数来提高收敛速度。ErrCor 和 PSO 均表现出改进的结果和具有竞争力的收敛性。然而，面对大型数据集，这些方法面临着可扩展性挑战，例如过多的核计算和庞大的隐藏层结构。最近提出的多列 RBFN 方法 (MCRN) 通过在并行系统中部署小型 RBFN 来改进 ErrCor 的性能。受 MCRN 成功的启发，我们提出了两种旨在提升 PSO 性能的新方法：基于 PSO 的多列 RBFN (MC-PSO) 和基于 APSO 的多列 RBFN (MC-APSO)。这些方法引入了采用进化群体方法训练的并行 RBFN 结构。每个 RBFN 均使用 PSO 或 APSO 算法，在数据集的特定空间子集上进行独立训练。由此产生的专门化训练 RBFN 针对其各自的子集进行了定制。测试时，仅测试实例邻居所在区域对应的选定 RBFN 参与多列输出的生成。这种专门化提高了准确性，而并行性则提升了速度。我们在多种基准数据集上对所提出的方法进行了评估。MC-PSO 和 MC-APSO 在准确性和召回率方面均优于 ErrCor、PSO、APSO 和 MCRN。此外，它们在大多数实验中均表现出更快的训练和测试速度。

The radial basis function neural network (RBFN) trained with a gradient descending algorithm provides an effective fully connected structure in both shallow and deep networks. The error correction (ErrCor), a state-of-the-art gradient-based training method, selects optimal hidden units to improve accuracy. Alternatively, as a population-based algorithm, the particle swarm optimization algorithm (PSO) uses the swarm experience to optimize RBFN parameters, offering global search and robustness to local minima. Adaptive PSO (APSO) has emerged as an improved variant of PSO. APSO algorithm improves convergence speed by dynamically adjusting swarm parameters during optimization. Both ErrCor and PSO demonstrate improved results and competitive convergence. However, with large datasets, these methods face scalability challenges such as excessive kernel computations and large hidden layer structures. A recent multi-column RBFN approach (MCRN) improves ErrCor performance by deploying small RBFNs in a parallel system. Inspired by MCRN's success, we propose two novel approaches to improve PSO performance: the multi-column RBFN with PSO (MC-PSO) and the multi-column RBFN with APSO (MC-APSO). These methods introduce parallel RBFN structures trained using evolutionary swarm methods. Each RBFN is independently trained on a specific spatial subset of the dataset using either PSO or APSO algorithms. These resulting specialist-trained RBFNs are tailored to their respective subsets. During testing, only selected RBFNs, where the test instance neighbors are located, contribute to the multi-column output. This specialization improves accuracy, while parallelism enhances speed. We evaluate the proposed methods on various benchmark datasets. The MC-PSO and MC-APSO outperform ErrCor, PSO, APSO, and MCRN in terms of accuracy and recall. They also demonstrate faster training and testing times in most experiments.

评分理由: The paper focuses on classical machine learning optimization (RBFN with PSO), unrelated to LLMs, RL, or multimodal systems. 'Unify Models' has minimal relevance due to multi-column architecture, but domain mismatch is significant.

Radial Basis Function Neural Network, Particle Swarm Optimization, Adaptive PSO, Multi-Column RBFN, Scalability, Accuracy, Parallelism, Benchmark Datasets

我们引入了 Graph Cascades，这是一种针对图神经网络（GNNs）和图变换器（GTs）的介观重连策略，旨在捕捉超越纯局部边或完全全局注意力的中间尺度图结构。基于传染性的扩散过程，Graph Cascades 可在 O(|V|+|E|) 时间内构建一个辅助图，其中由反复多跳强化支持的节点对将被升级为直接邻接关系。我们从理论上刻画了基于强化重连何时有效：包括基于强化边选择比直接邻接更具标签对齐性的充分条件、一个两跳强化完全同质性的 SBM（随机块模型）见证，以及通过图有效电阻形式化的介观连通性。实验上，在多个节点分类基准上，Graph Cascades 改进了多种 GNN 和稀疏 GT 骨干模型，其中在异质性图以及中等至高度数同质性图上观察到了最稳定的增益。这些理论条件还指出了介观重连不太可能有益的情形——低度数正则图以及存在结构瓶颈的图——且这些预测与观察到的失败结果相符。此外，我们还观察到重连图中的性能与结构属性之间存在紧密相关性。

We introduce Graph Cascades, a mesoscopic rewiring strategy for Graph Neural Networks (GNNs) and Graph Transformers (GTs) that captures intermediate-scale graph structure beyond purely local edges or fully global attention. Using contagion-based diffusion processes, Graph Cascades constructs, in O(|V|+|E|) time, an auxiliary graph where node pairs supported by repeated multi-hop reinforcement are promoted to direct neighbors. We theoretically characterize when reinforcement-based rewiring helps: sufficient conditions under which reinforcement-based edge selection is more label-aligned than direct adjacency, an SBM witness in which two-hop reinforcement is perfectly homophilic, and a formalization of mesoscopic connectivity via graph effective resistance. Empirically, across node-classification benchmarks, Graph Cascades improves multiple GNN and sparse-GT backbones, with the most reliable gains observed on heterophilic and moderate- to high-degree homophilic graphs. The theoretical conditions also identify regimes where mesoscopic rewiring is unlikely to be beneficial -- low-degree regular graphs and graphs with structural bottlenecks -- and these predictions match the observed failures. We additionally observe tight correlations between performance and structural properties in the rewired graphs.

评分理由: The paper focuses on Graph Neural Networks and mesoscopic graph rewiring, which is fundamentally unrelated to the provided keywords concerning Multimodal LLMs, World Models, Tokenizers, Visual Encoders, or Reinforcement Learning. 'Unify Models' receives a minimal score as the method applies to multiple backbones, but it does not align with the background context of unified multimodal architectures.

Graph Cascades, Mesoscopic Rewiring, Graph Neural Networks, Graph Transformers, Contagion-based Diffusion, Node Classification, Heterophilic Graphs, Structure-Aware

我们研究了多智能体深度强化学习，并针对多智能体深度确定性策略梯度（MADDPG）算法提出了两种改进。首先，我们引入了一种新颖的动作推断（Action Inference）机制，使每个智能体能够预测其他智能体的预期动作，从而提高其自身策略的准确性和稳定性。其次，我们在经验回放缓冲区中应用了一种基于几何分布的重要性采样策略，以优先处理更近期且更具信息量的经验，这有助于缓解多智能体环境中固有的非平稳性。我们在 PettingZoo 库提供的离散动作捕食者 - 猎物任务上评估了这两种修改，该库是一个用于通用多智能体强化学习基准的灵活 Python 接口。我们的结果表明，动作推断（Action Inference）在提高学习稳定性和智能体间合作方面是有效的，而使用几何分布的重要性采样相比标准 MADDPG 可以显著改善探索效率。代码可在 https://github.com/shaashwathsivakumar/MARL_Proj 获取。

We investigate multi-agent deep reinforcement learning and propose two enhancements to the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. First, we introduce a novel Action Inference mechanism that enables each agent to predict other agents' intended actions, thereby improving the accuracy and stability of its own policy. Second, we apply an importance sampling strategy, using geometric distribution, in the replay buffer to prioritize more recent and informative experiences, which helps mitigate the non-stationarity inherent in multi-agent environments. We evaluate both modifications on the discrete-action Predator-Prey task provided by the PettingZoo library, a flexible Python interface for general multi-agent reinforcement learning benchmarks. Our results indicate that Action Inference is effective in improving learning stability and inter-agent cooperation and that importance sampling using geometric distribution can lead to significant improvements in exploration efficiency over standard MADDPG. Code available at https://github.com/shaashwathsivakumar/MARL_Proj

评分理由: The paper focuses on enhancing the MADDPG algorithm for multi-agent reinforcement learning using action inference and importance sampling. It does not involve unifying models, tokenizers, visual encoders, world models, MLLMs, or multimodal architectures. Although it belongs to the RL domain, MADDPG is typically model-free, making the 'model-based RL' keyword only minimally relevant. No expert authors from the specified list are present.

Multi-Agent Reinforcement Learning, MADDPG, Action Inference, Importance Sampling, Policy Gradient, Non-stationarity, Predator-Prey

混合专家模型（MoE）架构通过稀疏专家激活来扩展模型容量，但其部署仍受限于内存，因为所有专家权重都必须驻留在内存中。混合精度量化可以通过为不同专家分配不同的位宽来显著减少这种内存占用。然而，现有方法通常依赖校准数据来估计专家重要性并确定位宽分配。对于前沿 MoE 大语言模型而言，原始训练数据（因而也是真实的训练分布）是专有的且不可访问的。因此，校准集不可避免地是不完美的代理，这可能导致对专家利用率的误估，并导致次优的位宽分配。鉴于现代 MoE 模型中观察到的显著专家间质量差异，以及重尾自正则化（HT-SR）理论在无需访问训练或测试数据的情况下成功预测神经网络模型质量，我们提出了 AlphaQ，一种用于 MoE 量化的无校准位分配方法。AlphaQ 基于 HT-SR 理论并遵循一个简单的原则：具有更重尾权重谱的专家通常训练得更好，因此应分配更高的位宽，而重尾结构较弱的专家可以采用更激进的量化策略。AlphaQ 通过测量专家层面的谱重尾性并求解一个预算约束优化问题来实现这一原则，该问题在全局位预算约束下最小化总量化误差。在多个 MoE 模型上，AlphaQ 在相同的位预算下始终优于基于校准的基线方法。值得注意的是，在 Qwen1.5-MoE 上，AlphaQ 仅以平均 3.5 位的专家位宽就实现了接近全精度的准确率，同时实现了超过 4 倍的内存压缩。我们的代码可在 https://github.com/Superone77/AlphaQ 获取。

Mixture-of-Experts (MoE) architectures scale model capacity through sparse expert activation, but their deployment remains memory-bound because all expert weights must reside in memory. Mixed-precision quantization can substantially reduce this footprint by assigning different bit-widths to different experts. Existing approaches, however, typically rely on calibration data to estimate expert importance and determine bit allocation. For frontier MoE LLMs, the original training data, and hence the true training distribution, is proprietary and inaccessible. As a result, calibration sets are inevitably imperfect surrogates, and this can misestimate expert utilization and lead to suboptimal bit allocation. Motivated by the substantial cross-expert quality variability observed in modern MoE models, and by the success of Heavy-Tailed Self-Regularization (HT-SR) theory at predicting neural network model quality without access to training or testing data, we propose AlphaQ, a calibration-free bit-allocation method for MoE quantization. AlphaQ draws on HT-SR theory and follows a simple principle: experts with more heavy-tailed weight spectra are typically better trained and hence should receive higher bit-widths, while experts with weaker heavy-tailed structure can be quantized more aggressively. AlphaQ operationalizes this principle by measuring expert-wise spectral heavy-tailedness and solving a budget-constrained optimization problem that minimizes total quantization error under a global bit-budget constraint. Across several MoE models, AlphaQ consistently outperforms calibration-based baselines under matched bit budgets. Notably, on Qwen1.5-MoE, AlphaQ achieves near full-precision accuracy with an average expert precision of only 3.5 bits, while delivering more than 4$\times$ memory compression. Our code is available at https://github.com/Superone77/AlphaQ.

评分理由: 该论文主要研究混合专家模型（MoE）的量化比特分配方法，基于重尾自正则化理论实现无校准量化。虽然涉及 LLM 模型（与 MLLM 关键词有微弱关联），但论文核心在于模型压缩与效率优化，未涉及模型统一、分词器设计、视觉编码器、世界模型、多模态融合或基于模型的强化学习等内容，因此与给定关键词的相关性极低。

Mixture-of-Experts, Quantization, Bit Allocation, Calibration-Free, Heavy-Tailed Self-Regularization, Model Compression, Memory Footprint

CARE-link 是一个开源的、基于网络的临床支持平台，旨在通过基于大语言模型（LLM）驱动的工作流连接临床医生和患者，以改善妊娠期糖尿病的管理。该系统聚合院外患者生成的数据，总结相关临床信息，并向临床医生提供情境感知的决策支持。对于患者，CARE-link 提供管理方案的清晰解释，并通过 WhatsApp 界面提供及时的生活方式指导。集成的双向设计旨在促进持续监测，支持个性化护理，并减少门诊随访的负担。采用模块化架构构建，该平台可适应其他需要纵向追踪和行为支持的慢性病。CARE-link 有望加强临床监管，促进患者依从性，并加强照护连续性，特别是在资源受限环境中。

CARE-link is an open-source, web-based clinical support platform designed to improve the management of gestational diabetes by linking clinicians and patients through an LLM-mediated workflow. The system aggregates patient-generated data outside the hospital, summarizes relevant clinical information, and delivers context-aware decision support to clinicians. For patients, CARE-link provides clear explanations of management plans and delivers timely lifestyle guidance through a WhatsApp interface. The integrated dual-facing design aims to promote continuous monitoring, support individualized care, and reduce the burden of in-clinic follow-ups. Built with a modular architecture, the platform can be adapted to other chronic conditions requiring longitudinal tracking and behavioral support. CARE-link has the potential to enhance clinical oversight, promote patient compliance, and strengthen continuity of care particularly in resource-constrained settings.

评分理由: The paper describes a healthcare application system (CARE-link) for diabetes management using LLMs, focusing on software architecture and clinical workflow. It does not address the technical model architectures implied by the keywords such as World Models, Visual Encoders, Tokenizers, or Model-Based Reinforcement Learning. While it utilizes an LLM (loosely related to MLLM), the core contribution is not model research, resulting in minimal relevance to the provided technical keywords.

Clinical Assistant, Remote Engagement, Electronic Health Records, Diabetes Management, LLM-mediated workflow, Web-based platform, Decision support, Patient-clinician linkage

大型语言模型（LLMs）在 CLadder 等因果推理基准上达到 50% 至 70% 的准确率，但目前尚不清楚这反映的是结构推理还是词汇模式匹配。我们引入了 Caliper，这是一种可控扰动方法，它在保留每个问题的因果图和概率规范的同时，将语义变量名替换为占位符。在九个参数量介于 38 亿至 671 亿之间的指令微调大型语言模型（LLMs）及三个因果推理基准上，词汇匿名化在本地 38 亿至 140 亿参数集上产生了稳健的准确率下降，分别为 +7.6、+27.0 和 +11.1 个百分点；而在跨越 2024 至 2026 代际的九个前沿模型上，CRASS 和 e-CARE 基准上的下降幅度分别增至 +29.6 和 +18.0 个百分点。在 40 个涉及的模型 - 基准单元中，有 39 个显示出正向差距，且在 CLadder 的伪词子集上，该差距缩小为原来的 1/17。结构化支架和少样本上下文学习各自缩小了该差距，但主要是通过降低较小模型上的 P0 准确率，而非恢复 P1 准确率。当前指令微调的大型语言模型（LLMs）在零样本评估下，一旦移除词汇锚点，便几乎没有显示出结构因果推理的证据。

Large language models reach 50 to 70% accuracy on causal reasoning benchmarks such as CLadder, but it is unclear whether this reflects structural reasoning or lexical pattern matching. We introduce Caliper, a controlled perturbation that replaces semantic variable names with placeholder tokens while preserving the causal graph and probabilistic specification of each question. Across nine instruction-tuned LLMs from 3.8B to 671B and three causal reasoning benchmarks, lexical anonymization yields robust accuracy drops of +7.6, +27.0, and +11.1 pp on a local 3.8B-14B set, rising to +29.6 and +18.0 pp on CRASS and e-CARE across nine frontier models spanning the 2024-2026 generations. Of 40 engaged model-by-benchmark cells, 39 show a positive gap, and the gap collapses by 17x on CLadder's pseudoword subset. Structured scaffolding and few-shot in-context learning each narrow the gap, but mainly by lowering P0 accuracy on smaller models rather than recovering P1. Current instruction-tuned LLMs, evaluated zero-shot, show little evidence of structural causal reasoning once lexical anchors are removed.

评分理由: 论文主题为大语言模型的可解释性与因果推理，通过词汇扰动分析模型行为，与关键词集中的多模态、世界模型、统一架构及强化学习方向严重不相关。仅'Tokenizer'因涉及占位符令牌操作而具有极低相关性（1.0 分），其余均为 0 分。作者列表中未包含指定专家，无额外加分。加权总分 1.5 分，远低于动态及格分 27.8 分。

Large language models, Causal reasoning, Lexical anchors, Causal structure, Placeholder tokens, Instruction-tuned, Perturbation, Accuracy drops

传统的生成对抗网络（GAN）在单图像超分辨率（SISR）任务中常面临幻觉伪影的困扰，主要是因为标准判别器评估的是整体图像的自然性，而非严格的条件真实性。为此，我们提出 MaCo-GAN，这是一种新颖的流形对比 GAN 框架，它用监督对比目标取代了传统的对抗损失。该方法的核心组件是一个动态假样本合成器，它将真实样本（GT）转化为一系列具有挑战性且感知上合理的假图像，这些图像严格保持低分辨率（LR）的一致性。利用这些合成样本，我们建立了一个稳健的对比极小极大博弈：生成器被训练将其预测吸引向流形内假样本（低失真），并排斥流形外假样本（高失真），而判别器则优化完全相反的目标。通过将基线超分辨率模型的对抗损失简单地替换为我们提出的目标，我们在各种基准上均实现了感知 - 失真权衡的一致改进。广泛的消融研究验证了该框架的有效性，并为这种条件对比博弈的动态机制提供了深入见解。

Conventional Generative Adversarial Networks (GANs) for Single Image Super-Resolution (SISR) often struggle with hallucinated artifacts, largely because standard discriminators evaluate overall image naturalness rather than strict conditional realism. To address this, we propose MaCo-GAN, a novel manifold-contrastive GAN framework that replaces the conventional adversarial loss with a supervised contrastive objective. A core component of our method is a dynamic fake sample synthesizer that transforms ground truth (GT) data into a spectrum of challenging, perceptually plausible fake images that strictly maintain low-resolution (LR) correspondence. Utilizing these synthesized samples, we establish a robust contrastive minimax game: the generator is trained to attract its predictions toward on-manifold fakes (low distortion) and repel them from off-manifold fakes (high distortion), while the discriminator optimizes the exact opposite. By simply replacing the adversarial loss of a baseline SR model with our proposed objective, we demonstrate consistent improvements in the perception-distortion trade-off across various benchmarks. Extensive ablation studies validate the effectiveness of our framework and provide deep insights into the dynamics of this conditional contrastive game.

评分理由: 该论文专注于单图像超分辨率（SISR）中的生成对抗网络（GAN）改进，采用流形对比学习解决幻觉伪影问题。提供的关键词主要围绕多模态大模型（MLLM）、世界模型、强化学习及模型统一等方向。论文内容与这些关键词高度不相关：未涉及文本分词器（Tokenizer）、多模态对齐（MultiModal/MLLM）、强化学习（model-based RL）或世界模型（World Models）。虽然涉及图像处理，但并非针对多模态视觉编码器（Visual Encoder）或模型统一（Unify Models）的研究，因此相关度极低，仅因涉及图像网络结构给予 Visual Encoder 极低分。

Single Image Super-Resolution, Generative Adversarial Networks, Manifold-Contrastive Learning, Perception-Distortion Trade-off, Fake Sample Synthesizer, Conditional Realism, GAN Framework

近年来，基于声明式交互协议的多智能体系统在建模与实现方面取得了重大进展。我们的贡献——Strabo，确立了这些进展与当前业界在 Agentic AI（智能体 AI）方面努力的相关性。具体来说，我们考察了 UCP（通用商务协议），这是谷歌最近牵头的一项旨在为 AI 智能体标准化电商交互的努力。本研究分为两部分。首先，我们将 UCP 中涉及结账的部分建模为声明式 Langshaw 协议，并使用 Peach（Langshaw 的一种编程模型）来实现智能体。这一部分凸显了形式化、声明式规范的优势。其次，我们展示了 Peach 智能体可以与谷歌实现的 UCP 智能体进行互操作，从而确立了我们的方法相对于 UCP 的保真度。这种互操作使得声明式协议和智能体能够以增量方式引入现有环境，表明了一种 EMAS 理念影响实践的路径，而无需进行全面更新。

The last few years have witnessed major advances in the modeling and implementation of multiagent systems based on declarative interaction protocols. Our contribution, Strabo, establishes the relevance of these advances to ongoing industry efforts in Agentic AI. Specifically, we consider UCP, the Universal Commerce Protocol, a recent Google-led effort to standardize e-commerce interactions for AI agents. Our exercise is in two parts. One, we model the part of UCP dealing with checkouts as a declarative Langshaw protocol and implement agents using Peach, a programming model for Langshaw. This part of the exercise brings out the advantages of formal, declarative specifications. Two, we show that Peach agents can interoperate with UCP agents implemented by Google, thereby establishing the fidelity of our approach with respect to UCP. Such interoperation enables the incremental introduction of declarative protocols and agents into a conventional setting, indicating a pathway by which EMAS ideas could influence practice without demanding a wholesale update.

评分理由: The paper addresses formal specification of agentic interaction protocols using Strabo and Langshaw, focusing on interoperability with industry standards (UCP). The provided keywords relate to deep learning components (Tokenizer, Visual Encoder), multimodal models (MLLM, MultiModal), and reinforcement learning paradigms (World Models, model-based RL). There is no intersection in technical domain or methodology, hence all scores are 0.

Strabo, Declarative Specification, Agentic Interaction Protocols, Multiagent Systems, Universal Commerce Protocol, Langshaw Protocol, Interoperability

当 AI 代理调用 API 并遭遇验证错误时，它所需的不仅仅是错误原因——它需要知道下一步该做什么。自反思 API (self-reflective API) 在验证失败时返回一个机器可读的 recovery_feedback.suggestions[] 载荷，足以使代理能够修复请求并重试，而无需外部推理。在经泄漏审计的试点实验中（每组 N=30，3 个 LLMs，10 个对抗性任务），相较于纯英文诊断，结构化建议将 Anthropic 模型的任务完成率提高了 +36.7--40.0 个百分点（Fisher 精确检验 p ≤ 0.0022），且每成功 token 效率提升了 1.8--2.2 倍。在 gpt-4o-mini 上提升不显著（p=0.435）；在计费 API 上的第二领域复现确认了这一模式。只有在审计了 LLMs 基准测试中两类未记录的答案泄露后，该比较才成立。我们发布了 audit_prompt_leakage.py 作为可重用的持续集成 (CI) 基础设施。代码和数据：https://github.com/arquicanedo/self-reflective-apis.

When an AI agent calls an API and hits a validation error, it needs more than what went wrong -- it needs what to do next. A self-reflective API returns, on validation failure, a machine-readable recovery\_feedback.suggestions[] payload sufficient for the agent to repair the request and retry without external reasoning. On a leak-audited pilot ($N{=}30$ per cell, 3 LLMs, 10 adversarial tasks), structured suggestions lift task-completion rate by $+36.7$--$40.0$pp over plain-English diagnoses on Anthropic models (Fisher's exact $p \le 0.0022$), at $1.8$--$2.2\times$ better per-success token efficiency. The lift is not significant on gpt-4o-mini ($p{=}0.435$); a second-domain replication on a billing API confirms the pattern. The comparison only holds after auditing two undocumented classes of answer leakage in LLM benchmarks. We shipaudit\_prompt\_leakage.py as reusable CI infrastructure. Code and data: https://github.com/arquicanedo/self-reflective-apis.

评分理由: The paper focuses on API design for AI agent error recovery (structured feedback vs. natural language) and auditing prompt leakage. It does not discuss model unification, tokenization mechanisms, visual encoders, world model learning, multimodal architectures, or model-based reinforcement learning. Thus, relevance to all specific technical keywords defined in the scoring criteria is negligible.

Self-Reflective APIs, AI Agent Recovery, Structured Feedback, Validation Error, Token Efficiency, Prompt Leakage, API Recovery

多方对话是研究协作推理与决策的关键场景，然而现有数据集很少聚焦于结构化、深入的复杂推理任务。我们引入 DeliChess（一种新颖的群体审议对话数据集），在该数据集中，参与者协作解决多选式国际象棋谜题。每个小组首先独立完成谜题，随后进行多方讨论，最后提交修订后的集体答案。该数据集包含 107 条对话，附有完整转录文本、讨论前后的选择项，以及关于谜题难度和走棋质量的元数据。我们基于棋引擎评估的三个指标来评估性能，发现审议过程显著提高了群体的准确率。我们进一步利用基于先前审议数据训练的 classifier，分析探查性话语（即引发提议、论证或战略反思的消息）的作用。尽管探查性话语使得讨论后群体表现的变异性增加，但它并不总是能带来更好的表现。我们的数据集为建模群体推理、对话动态以及解决明确战略领域中不同观点和意见的分歧提供了丰富的 testbed。

Multi-party dialogue is a critical setting for studying collaborative reasoning and decision-making, yet existing datasets rarely focus on structured, in-depth complex reasoning tasks. We introduce DeliChess, a novel dataset of group deliberation dialogues in which participants collaboratively solve multiple-choice chess puzzles. Each group first completes the puzzle individually, then engages in a multi-party discussion before submitting a revised collective answer. The dataset includes 107 dialogues with full transcripts, pre- and post-discussion choices, and metadata on puzzle difficulty and move quality. We evaluate performance using three metrics based on chess engine evaluations, and find that deliberation significantly improves group accuracy. We further analyse the role of probing utterances (i.e., messages that elicit proposals, justifications, or strategic reflection) using a classifier trained on prior deliberation data. While probing makes group performance more variable after discussion, it does not consistently lead to better performance. Our dataset offers a rich testbed for modelling group reasoning, dialogue dynamics, and the resolution of differing perspectives and opinions in a well-defined strategic domain.

评分理由: 该论文主要贡献在于构建了一个多主体对话数据集（DeliChess），用于研究棋类谜题求解中的协作推理和对话动态。提供的关键词涉及多模态大模型架构（Unify Models, Tokenizer, Visual Encoder, MLLM, MultiModal）、世界模型（World Models）及强化学习（model-based RL），这些内容在论文摘要和标题中均未提及。论文焦点在于对话分析而非模型架构或强化学习，因此所有关键词的相关度均为 0。

Multi-party Dialogue, Chess Puzzle Solving, Collaborative Reasoning, Group Deliberation, Dialogue Dataset, Strategic Domain, Probing Utterances, Collective Decision Making

尽管学界普遍认为 AI 生成文本带来广泛的社会风险，但在 AI 生成文本检测领域，对于何为有害使用尚未达成共识。相反，现有数据集与方法往往自行定义标准并做出假设，有时隐含于此，且通常与现实世界的需求和应用关联松散。为填补这一空白，我们在此系统性地定义了各类 AI 生成文本的概念及其特征。为此，我们收集了 AITDNA——一个新的基准数据集，包含人机协作构建的文本，并标注了详细的起源信息，包括完整的编辑历史和 AI 交互历史。我们评估了多种机器生成文本检测器，发现它们往往仅在特定概念下表现良好，却难以充当通用检测器。我们公开发布了代码和数据。

Although it is generally agreed that AI-generated text poses a broad societal risk, there is no common understanding in the AI-generated text detection literature on what constitutes harmful use. Rather, existing datasets and approaches often define their own criteria and make their own assumptions, sometimes implicitly, and often only loosely related to real-world needs and applications. To address this gap, we here systematically define various notions of AI-generated text and their characteristics. To study these, we collect AITDNA - a new benchmark of human-machine co-constructed texts that is annotated with detailed genesis information, such as the entire edit and AI-interaction history. We benchmark various machine-generated text detectors and find that they often only perform well for specific notions but not as broad detectors. We release code and data publicly.

评分理由: 该论文专注于 AI 生成文本检测及基于现实假设的基准测试（AITDNA），涉及人机协作构建文本。提供的关键词涉及多模态大模型、世界模型、视觉编码器和基于模型的强化学习。论文内容（文本检测、NLP）与关键词（多模态架构、世界建模、强化学习）无技术重叠，因此所有关键词评分为 0。作者列表中未包含指定的专家（Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang）。

AI-generated text detection, human-machine co-constructed texts, AITDNA benchmark, realistic assumptions, machine-generated text detectors, genesis information, edit history

基于表达性逻辑的证明助手在证明搜索方面的自动化能力受限，导致基于证明助手的形式化验证成本上升。我们通过引入 Isabelle/HOL 的 Abduction Prover 来解决这一问题。面对具有挑战性的证明目标，Abduction Prover 通过溯因推理识别有用的猜想，从而为该目标构建证明脚本。

Proof assistants based on expressive logics suffer limited automation for proof search, raising the cost of formal verification based on proof assistants. We address this problem by introducing the Abduction Prover for Isabelle/HOL. Given a challenging proof goal, the Abduction Prover constructs a proof script for the goal by identifying useful conjectures using abductive reasoning.

评分理由: The paper focuses on formal verification and automated theorem proving using abductive reasoning within the Isabelle/HOL framework. The provided keywords pertain to multimodal large language models, world models, and reinforcement learning, which are unrelated to the logical and formal methods domain of this work. Hence, all relevance scores are 0. None of the listed expert authors are present in the author list.

Abduction Prover, Isabelle/HOL, Proof Assistants, Proof Search, Abductive Reasoning, Formal Verification, Automated Theorem Proving

随着 Replika 和 Character.AI 等 AI 伴侣平台的迅速扩张，关于不安全人机交互的担忧日益加剧。本研究介绍了 AICompanionBench，据我们所知，这是首个公开的、包含细粒度安全风险类别标注的人机伴侣对话基准数据集。该数据集包含从 Reddit 收集的 2,123 个真实世界的 Replika 对话，并通过人机协作标注了九个类别：性行为、反社会行为、身体攻击、言语攻击、物质滥用、自残与自杀、控制、操纵以及无害。基于此基准，我们在 LLM 作为裁判框架下评估了 20 个最先进的开源和闭源大语言模型（LLMs），以检测不安全交互。结果显示模型性能存在显著差异，尽管较强的模型实现了高整体准确率，但在操纵等细微类别上仍存在困难，且会将良性对话错误地识别为有害内容。我们的发现表明，尽管当前大语言模型（LLMs）能有效检测显式有害内容，但在识别隐式不安全交互方面仍存在局限。总体而言，本研究为 AI 伴侣安全研究贡献了一个新的基准数据集，并提供了利用大语言模型（LLMs）监控 AI 伴侣系统的见解。该数据集公开可用，地址为：https://github.com/anonymousresearcher2026/AICompanionBench/blob/main/AICompanionBench.xlsx

As AI companion platforms such as Replika and Character.AI rapidly grow, concerns about unsafe human-AI interactions have intensified. This study introduces AICompanionBench, to our knowledge the first publicly available benchmark dataset of human-AI companion conversations annotated with fine-grained safety risk categories. The dataset contains 2,123 real-world Replika conversations collected from Reddit and annotated through human-AI collaboration across nine categories: sexual behavior, antisocial behavior, physical aggression, verbal aggression, substance abuse, self-harm and suicide, control, manipulation, and no-harm. Using this benchmark, we evaluate 20 state-of-the-art open-source and closed-source LLMs under an LLM-as-judge framework for detecting unsafe interactions. Results show substantial variation in model performance, with stronger models achieving high overall accuracy but still struggling with nuanced categories such as manipulation, as well as benign conversations that are incorrectly identified as harmful. Our findings suggest that while current LLMs can effectively detect explicit harmful content, they remain limited in identifying implicit unsafe interactions. Overall, our work contributes a new benchmark dataset for AI companionship safety research and offers insights into monitoring AI companion systems using LLMs. The dataset is publicly available at: https://github.com/anonymousresearcher2026/AICompanionBench/blob/main/AICompanionBench.xlsx

评分理由: The paper focuses on safety benchmarking for AI companions using LLMs-as-judges, which does not involve model unification, tokenizers, visual encoders, world models, multimodal architectures (MLLM/MultiModal), or model-based reinforcement learning. Thus, all keyword scores are 0.

AICompanionBench, LLMs-as-Judges, AI Companion Safety, Human-AI Interaction, Safety Risk Categories, Benchmark Dataset, Unsafe Interactions

库普曼（Koopman）理论将非线性动力学转化为线性谱问题。然而，在计算中，一切均取决于一个困难的有限维选择：可观函数（observables）必须具有表达能力，在动力学作用下几乎不变，且理想情况下与复合运算兼容。深库普曼（Deep Koopman）方法学习灵活的坐标，而保结构方法则在固定字典上强制算子恒等式。我们通过引入深嵌入乘性动态模态分解（DeepMDMD）来结合这些思想，该方法学习一个潜在空间及其划分，同时将库普曼乘积法则（Koopman product rule）作为精确代数约束强制执行。训练过程在精确的乘性算子更新与促进库普曼闭包（Koopman closure）的可微潜在聚类步骤之间交替进行。最终得到的是一个在学习到的潜在单元（latent cells）上的有限转移映射。其非零谱位于单位圆上，其字典由动力学塑造而非环境几何塑造，预测先在潜在坐标中进行，随后解码至物理空间。在哈密顿（Hamiltonian）、混沌及流体示例中，DeepMDMD 学习到的字典比几何 MDMD 划分产生的字典更为紧凑且动力学相干性更好。该方法减少了谱污染，揭示了更丰富的连续谱结构，并在强噪声下仍能给出稳定预测。在高维流动中，包括 158,624 维圆柱尾流和噪声雷诺数 Re=20,000 的顶盖驱动腔（lid-driven cavity），它在状态空间 MDMD 失效之处仍能保持相干结构及长时间谱统计特性。这些结果暗示了库普曼学习的一个实用规则：学习坐标，约束代数。

Koopman theory turns nonlinear dynamics into a linear spectral problem. In computation, however, everything depends on a hard finite-dimensional choice: the observables must be expressive, nearly invariant under the dynamics, and, ideally, compatible with composition. Deep Koopman methods learn flexible coordinates, whereas structure-preserving methods enforce operator identities on fixed dictionaries. We combine these ideas by introducing Deep Embedded Multiplicative Dynamic Mode Decomposition (DeepMDMD), a method that learns a latent space and a partition of it, while enforcing the Koopman product rule as an exact algebraic constraint. Training alternates between an exact multiplicative operator update and a differentiable latent-clustering step that promotes Koopman closure. The result is a finite transition map on learned latent cells. Its nonzero spectrum lies on the unit circle, its dictionary is shaped by the dynamics rather than by ambient geometry, and forecasts are made in latent coordinates before being decoded to physical space. Across Hamiltonian, chaotic, and fluid examples, DeepMDMD learns dictionaries that are far more compact and dynamically coherent than those produced by geometric MDMD partitions. It reduces spectral pollution, reveals richer continuous-spectrum structure, and gives stable forecasts under severe noise. In high-dimensional flows, including a 158,624-dimensional cylinder wake and a noisy $Re=20,000$ lid-driven cavity, it preserves coherent structures and long-time spectral statistics where state-space MDMD fails. These results suggest a practical rule for Koopman learning: learn the coordinates, constrain the algebra.

评分理由: The paper focuses on Koopman operator theory and Dynamic Mode Decomposition for dynamical systems (fluids, Hamiltonian), whereas the provided keywords pertain to Multimodal Large Language Models (MLLM), Tokenization, and Reinforcement Learning. There is no overlap in methodology, domain, or terminology between the paper's content and the evaluation keywords.

Koopman theory, Dynamic Mode Decomposition, Algebraic constraints, Latent space, Fluid dynamics, Hamiltonian systems, Spectral analysis, Deep learning

保持数据隐私是结构化数据管理和数据挖掘领域的一个重要议题。然而，分布式因果结构学习中的隐私泄露问题是一个持续的挑战，尤其是在涉及数据传输和计算的场景中。本文提出了一种基于全同态加密（FHE）的方法，该方法在密文上执行计算，确保数据在传输和计算过程中始终处于加密状态。然而，由于计算成本高昂以及 FHE 对除法和有限支持的对数运算支持有限，将 FHE 应用于因果结构学习面临挑战。为应对这一挑战，我们提出了一系列新颖的技术，包括：（i）电路简化以提高效率；（ii）通过牛顿 - 拉夫逊倒数法和泰勒展开近似除法和对数运算；以及（iii）一种带有 SIMD 加速的批处理技术，以增强整个学习过程。此外，该方法具有良好的可移植性，易于扩展至 FHE 之外以支持差分隐私。实证结果表明，我们的方法在测试数据集上实现了高一致性且因果结构与明文版本相当。最后，该方法高效且实用，即使在 FHE 的隐私保护下，也能在数十分钟内完成因果结构的学习。

Preserving data privacy is an important topic in structural data management and data mining. However, the issue of privacy leakage in distributed causal structure learning is a persistent challenge, especially in cases where data transmission and computation are required. In this paper, we propose a method based on fully homomorphic encryption (FHE) that performs calculations on ciphertexts, keeping data encrypted in transition and computation. Nevertheless, adopting FHE to causal structure learning is challenging due to the high computation cost and limited support on division as well as logarithm operations in FHE. To tackle this challenge, we propose a series of novel techniques including (i) circuit simplification for better efficiency, (ii) approximation of division and logarithm through Newton-Raphson Reciprocal and Taylor expansion, and (iii) a batching technique with SIMD-acceleration to enhance the whole learning process. Additionally, our method can be easily extended beyond FHE by demonstration of its portability to support differential privacy. Empirical results show that our method achieves high consistency and comparable causal structure with the plaintext version in the datasets tested. Last, our method is efficient and practical to complete learning causal structures in tens of minutes even under the privacy protection of FHE.

评分理由: The paper focuses on causal structure learning using Fully Homomorphic Encryption (FHE) for privacy preservation, which is unrelated to the provided keywords concerning Multimodal Large Language Models (MLLM), World Models, Tokenizers, Visual Encoders, and Model-Based Reinforcement Learning. None of the specified expert authors (Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang) are present in the author list.

Fully Homomorphic Encryption, Causal Structure Learning, Data Privacy, Distributed Learning, Circuit Simplification, Approximation Techniques, SIMD-acceleration

南希·格雷斯·罗曼太空望远镜（Roman）计划于 2026 年 9 月或更早发射，将开展具有前所未有的空间分辨率和采样频率的宽视场红外成像巡天，从而能够发现数百万个天文瞬变源。因此，有必要建立自动生成警报的自动化处理流程，以便望远镜在发射后不久即可开始发现可靠的瞬变源和变源。然而，目前尚无真实的罗曼望远镜数据，这使得此类处理流程的开发变得困难。在这项工作中，我们提出了一种机器学习模型 RuBR 以及一种通用方法，用于在 RAPID 处理流程中区分真实的瞬变源和变源探测与虚假（伪造）探测。具体而言，我们提出了基于此方法的三种模型：RuBR_comb，其在组合的本地注入瞬变源和 OpenUniverse2024 瞬变源上进行训练和测试；RuBR_loc，其在本地注入瞬变源上训练并在 OpenUniverse2024 瞬变源上测试；以及 RuBR_DA，其在域适应模式下将本地注入瞬变源与一部分 OpenUniverse2024 瞬变源结合用于训练。这为在罗曼任务早期阶段缺乏真实标签的情况下，将 RuBR_comb 模型适应到真实观测的策略铺平了道路。尽管图像差分处理流程仍在不断改进，但我们的实验结果表明了所提出方法的有效性，以及其在罗曼时代实现稳健真实 - 虚假分类的前景。

The Nancy Grace Roman Space Telescope (Roman), set for launch as early as September 2026, will conduct wide-field infrared imaging surveys with unprecedented spatial resolution and cadence, enabling the discovery of millions of astronomical transients. Hence, it is necessary to have automated pipelines for generating alerts in place so that the telescope can begin discovering reliable transients and variable objects soon after it is launched. However, no real Roman data currently exist, making the development of such pipelines difficult. In this work, we present a machine learning model $RuBR$ and a general methodology for distinguishing genuine transient and variable detections from spurious (bogus) detections within the RAPID pipeline. In particular, we present three models using this methodology: $RuBR_{comb}$ trained and tested on combined locally injected and OpenUniverse2024 transients, $RuBR_{loc}$ trained on locally injected transients and tested on OpenUniverse2024 transients, and $RuBR_{DA}$ that combines locally injected transients with a fraction of OpenUniverse2024 transients in domain-adaptation mode for training. This paves the way for strategies to adapt the $RuBR_{comb}$ model to real observations in the absence of any ground-truth labels during the early phases of the Roman mission. While the image differencing pipeline continues to be improved, our experimental results demonstrate the effectiveness of the proposed approach and its promise for robust real-bogus classification in the Roman era.

评分理由: 该论文专注于天文学瞬变源检测，利用模拟数据训练机器学习模型（RuBR）以区分真实与虚假检测，属于天文学图像处理领域。提供的关键词（如 MLLM、World Models、Tokenizer、model-based RL）均属于多模态大模型与强化学习范畴，与本文的研究内容、方法论及术语体系无直接关联，因此所有关键词相关度评分为 0。作者列表中未包含指定的 Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang 专家，故无额外加分。

Roman Space Telescope, RAPID pipeline, Machine learning, Transient detection, Real-bogus classification, Simulated data, Domain adaptation

准确的 T 细胞受体 (TCR) 抗原特异性计算预测将变革 T 细胞生物学的研究并实现可扩展的免疫工程，然而现有模型缺乏足够的灵敏度和特异性以适用于广泛应用。主要限制在于缺乏严格定义的、未见过的基准数据集，这些数据集允许对模型性能和泛化能力进行无偏评估。在此，我们描述了满足这一标准的两类互补数据集，并论证它们既为模型评估提供了稳健框架，也为下一代 TCR-抗原预测算法开发奠定了基础。

Accurate computational prediction of T cell receptor (TCR) antigen specificity would transform the study of T cell biology and enable scalable immune engineering, yet existing models lack sufficient sensitivity and specificity for broad applications. A major limitation is the absence of rigorously defined, unseen benchmark datasets that allow unbiased evaluation of model performance and generalizability. Here, we describe two complementary classes of datasets that meet this criterion and argue that they provide both a robust framework for model assessment and a foundation for next-generation TCR-antigen prediction algorithm development.

评分理由: The paper focuses on computational immunology and TCR-antigen prediction benchmarks, whereas the keywords pertain to Multimodal Large Language Models, World Models, and Reinforcement Learning. There is no thematic or methodological overlap, resulting in zero relevance for all keywords. Additionally, none of the specified expert authors (Yang Shi, Xuanyu Zhu, Yuhao Dong, Saining Xie, Manyuan Zhang) are present in the paper's author list.

TCR Antigenic Epitope, Prediction Models, Generalization Power, Benchmarking, Unseen Datasets, Immune Engineering, Computational Prediction

系统生成的日志构成了安全监控的基础，然而其基于模板的刚性格式既阻碍了自动化分析，也妨碍了人工理解。我们提出 NLLog（自然语言日志），这是一种轻量级管道，它确定性地将解析后的模板重写为 WHO-WHAT-SEVERITY 句子，采用词频 - 逆文档频率（TF-IDF）加权进行聚合，利用树集成对会话进行分类，并使用 TreeSHAP 回溯证据以供分析师审查。在 Hadoop 分布式文件系统（HDFS）和 Blue Gene/L（BGL）语料库上，NLLog 超过了两个复现的匹配协议基线；在 HDFS、BGL 和 AIT 警报数据集上，它保持低误报率，且具有适合安全运营中心（SOC）分诊的通用硬件延迟。覆盖率、稀疏与密集编码对比、忠实性以及对抗性消融实验表明，回退充分性依赖于语料库，注册时的覆盖率检查可在部署前揭示细化需求，而可审计的确定性重写结合轻量级密集编码为日志异常检测和分诊提供了一个可衡量的表示层。

System-generated logs underpin security monitoring, yet their rigid template-based format hinders both automated analysis and human comprehension. We present NLLog (Natural-Language Log), a lightweight pipeline that deterministically rewrites parsed templates into WHO-WHAT-SEVERITY sentences, pools them with term-frequency-inverse-document-frequency weighting, classifies sessions with tree ensembles, and back-projects evidence with TreeSHAP for analyst review. On Hadoop Distributed File System (HDFS) and Blue Gene/L (BGL) corpora, NLLog exceeds two reproduced matched-protocol baselines; across HDFS, BGL, and the AIT Alert Data Set, it sustains low false-positive rates with commodity-hardware latency suitable for security operations center triage. Coverage, sparse-versus-dense, faithfulness, and adversarial ablations show that fallback sufficiency is corpus-dependent, that an enrollment-time coverage check can surface refinement requirements before deployment, and that an auditable deterministic rewrite combined with lightweight dense encoding provides a measurable representation layer for log-anomaly detection and triage.

评分理由: 该论文专注于日志异常检测与自然语言重写，使用集成学习和 TF-IDF 方法。提供的关键词涉及多模态大模型、世界模型和强化学习等领域，与论文内容无直接关联。因此所有关键词相关度均为 0。

NLLog, Log Anomaly Detection, Natural Language Rewriting, Tree Ensembles, Explainability, SOC, TF-IDF

一致性 (Consistency) 是动态次模最大化 (submodular maximization) 中的一个重要属性，它意味着始终维持一个近似最优解 (near-optimal solution)，并在每一步中对解仅进行少量调整。先前工作已针对仅插入 (insertion-only) 的情况探索了这一问题，其中算法面对的是 $n$ 个插入流，并已建立了该问题基数约束 (cardinality-constrained) 版本的上下界。我们在完全动态设置 (fully dynamic setting) 中考虑这一问题，其中操作流可能同时包含插入 (insertions) 和删除 (deletions)。我们开发了一个用于在此设置下设计算法的通用框架 (general framework)，并将其实例化以获得具有次线性一致性 (sublinear consistency) 的首个常数因子近似 (constant-factor approximations)。对于基数约束 (cardinality constraints)，我们提出一个 $\frac 12 - O(\varepsilon)$ 近似，该近似是 $O\left(\frac{1}{\varepsilon^2}\right)$ 一致的。对于秩 -k matroid 约束 (rank-$k$ matroid constraints)，我们构建了一个针对动态最优解 (dynamic optimum) 的 $\frac 14 - O(\varepsilon)$ 近似，该近似是 $O\left(\frac{\log k}{\varepsilon^2}\right)$ 一致的。

Consistency is an important property in dynamic submodular maximization and entails maintaining a near-optimal solution at all times, making only a small number of adjustments to the solution in each step. Prior work has explored this question for the insertion-only case, where the algorithm faces a stream of $n$ insertions, and has established lower and upper bounds for the cardinality-constrained version of the problem. We consider this question in the fully dynamic setting, where the stream of operations may contain both insertions and deletions. We develop a general framework for designing algorithms for this setting, and instantiate it to obtain the first constant-factor approximations with sublinear consistency. For cardinality constraints, we propose a $\frac 12 - O(\varepsilon)$ approximation that is $O\left(\frac{1}{\varepsilon^2}\right)$ consistent. For rank-$k$ matroid constraints, we construct a $\frac 14 - O(\varepsilon)$ approximation to the dynamic optimum that is $O\left(\frac{\log k}{\varepsilon^2}\right)$ consistent.

评分理由: 论文主题属于组合优化与动态算法领域（子模最大化），而评分关键词涉及多模态大模型、世界模型及强化学习领域。两者研究范畴完全不同，无任何技术或主题关联，故所有关键词相关性评分为 0。

Submodular Maximization, Dynamic Setting, Consistency, Approximation, Matroid Constraints, Insertions and Deletions, General Framework

Mean-based 算法（基于均值算法）是一类在线学习算法，它们给平均奖励较低的动作分配较低的概率。近期研究表明，这些算法倾向于收敛到 serially undominated 动作（序列上不被支配的动作），这些动作在经济游戏中近似于 Nash equilibria（纳什均衡）。然而，实证研究也显示，在 bandit-feedback（多臂老虎机反馈）场景下，它们的收敛速度相比已有算法较慢。我们在 time horizon（时间范围）未知且仅可获得 bandit-feedback（多臂老虎机反馈）的情况下研究 Mean-based 算法。在此设置下，我们提供了算法定义序列 $\gamma_t$ 的下界，正式确立了这些算法学习速度的上限。此外，我们提出了两种 Mean-based 算法：一种推广了 $\epsilon$-greedy，另一种将基于均值的 Exp3 扩展到了未知 time horizon。我们的实验表明，Mean-based 算法虽然稍慢，但能与其他 bandit-feedback 算法竞争性地表现。我们进一步分析了与 no-regret（无遗憾）算法的关系。根据 $\gamma_t$ 的选择，与 no-regret 算法的交集是非平凡的，我们证明了存在既是 Mean-based 又是 no-regret 的算法。这为先前贡献所暗示的此类算法的"exploitability"（可被利用性）增添了背景。

Mean-based algorithms are a class of online learning algorithms that assign low probability to actions with low average rewards. Recent work indicates these algorithms converge favorably to serially undominated actions, which approximate Nash equilibria in economic games. However, empirical studies also show slower convergence compared to established algorithms in bandit-feedback scenarios. We study mean-based algorithms when the time horizon is unknown and only bandit feedback is available. In this setting, we provide the first lower bound on the algorithm-defining sequence $γ_t$ that formally establishes a limit on how fast these algorithms can learn. Additionally, we propose two mean-based algorithms: one generalizes $ε$-greedy, and the other extends the mean-based Exp3 to unknown horizons. Our experiments show that mean-based algorithms, although slightly slower, can perform competitively with other bandit-feedback algorithms. We further analyze the relationship to no-regret algorithms. Depending on the choice of $γ_t$, the intersection with no-regret algorithms is non-trivial, and we show that algorithms exist that are both mean-based and no-regret. This adds context to the "exploitability" of this class of algorithms that previous contributions suggest.