信息流

Legislative hearings are distinct communication events with their own human recognizable structure and logic, yet they have been analyzed as distinct phenomena beyond their transcripts. Understanding legislative hearings at a deeper level is a major step in analyzing the work and power relationships within democratic bodies. In this work, we present a novel end-to-end pipeline for parsing rhetorical structures from legislative committee hearings with the goal of improving legislative discourse analysis and democratic accountability. We also propose a rhetorical structure grammar and find that 55.21% of California state bill discussions in our dataset can be parsed with this grammar. We evaluate our pipeline with two utterance section classifiers on a random subset of over 2,000 annotated utterances corresponding to 100 hearings. We find that a logistic regression-based classifier with grammar-based output masking and utterance tag smoothing outperforms more complex models, achieving an F1 score of 52.8% against our ground truth discourse section labels.

Researchers often need web corpora that are not answers to a single query, but reusable collections of documents spanning many decentralized sources. We refer to this problem as thematic web data collection: assembling semantically relevant documents that share a common theme but are scattered across heterogeneous and structurally disconnected regions of the web. Existing approaches, including traversal-based methods and LLM-driven web agents, are limited by local exploration or short-horizon stopping, preventing comprehensive collection. We formulate this task as a long-horizon information foraging problem over query-induced semantic patches, where the central challenge is deciding when to continue local exploitation and when to switch to new regions. We propose PatchScout, a multi-agent framework that instantiates a class of patch-switching policies to balance local exploitation and adaptive patch switching under a fixed budget. Experiments on four live-web topics show that PatchScout substantially improves yield and domain coverage over traversal-based and search-augmented reasoning agent baselines. In an applied social science setting with an incomplete prior collection, PatchScout discovers 102 previously unidentified entities, illustrating its ability to expand existing thematic datasets.

LLM agent 难以自动化复杂的企业流程，因为静态的、预先构建的 harness 无法捕捉专有策略、未文档化的行为，或只有在执行过程中才会浮现的、依赖状态的恢复路径。从 agent 经验中积累的程序性记忆提供了一条无需训练的路径来弥合这一差距——但与情景记录或语义事实不同，它必须将重复出现的成功和经过诊断的失败，整合为可复用的动作与恢复策略。我们提出 WASP-Mem（Wiki-based Agent-Synthesized Procedural Memory），这是一种轻巧而有效的方法，将执行轨迹与反馈提炼为持久的、相互链接的 wiki 页面，从而实现 agent 的持续学习。在 TheAgentCompany benchmark 上，WASP-Mem 在我们比较的三个模型中，相较于最强基线将 checkpoint 通过率提高了 4.6-6.1%，同时所需存储减少 57-61%。我们进一步提出了一项系统性对照研究，涵盖整合方法、反馈类型、轨迹摄取范围和 refinement 轮次。我们的结果表明，诸如人工定义 schema 和显式去重规则这类规定性知识组织方式会损害性能，而不是弥合程序性知识差距。相反，人类诊断将失败轨迹转化为纠正性知识，并推动持续改进：单轮人类诊断比四轮自我判断高出 7.0-10.9%。这些发现指向一种简单的分工：让 agent 组织程序性

Sli2Vol通过相邻切片之间的亲和力传播二值掩码，但误差可能持续存在，因为每个预测都会成为下一步的输入。三个正交标注减少了对任何单一链的依赖，但它们提供的不仅仅是三个起始掩码：当一条链与另一个标注平面相交时，正确的标签已经已知。这些标签可以在传播过程中纠正漂移，并在融合时指示对每条链的信任程度。在标注平面之间，链所携带的信息也可能很重要。二值掩码记录前景成员关系，但丢弃了到边界的距离，这一线索可能有助于跟随切片之间的渐变形状变化。我们使用固定的预训练亲和力骨干网络来解决这两个误差来源。我们通过在每条链与另一个标注平面相交处校正其预测来锚定该链，然后利用其与这些标签的一致性来加权融合。我们将这种锚定掩码与两种传输边界几何的方式进行比较：一种是通过稠密亲和力平均传播的截断符号距离场（SDF），另一种是其顶点遵循来自同一亲和力的运动线索的轮廓。轮廓跟踪器从手工设计的规则开始，并从506个CT体积中的超体素伪对象学习运动校正，无需手动器官标签。在来自约403次CT扫描、涵盖17个腹部结构的2,091个留出结构实例上，锚定将平均Dice从正交掩码传播的0.737提高到0.7

Most memory systems for LLM agents are retrospective, preserving and retrieving past information to inform current decisions. Proactive agents, however, also require prospective memory: retaining future-directed intentions and realizing them when appropriate opportunities arise, even without explicit requests. Crucially, remembering the right intention does not guarantee acting at the right time: a related observation may call for waiting, while a later cue may make the same intention actionable. Inspired by cognitive theories of human prospective memory, we introduce Pro-Memory, a memory framework that separates maintaining future intentions from deciding when and how to realize them. Pro-Memory preserves pending objectives linked to their supporting evidence, assesses new observations against their original goals and activation conditions, and coordinates appropriate contributions with the agent's ongoing actions. It further updates intention states based on actions actually issued or evidence of resolution, while an experience memory distills feedback from completed episodes into guidance for future decisions. Experiments on two benchmarks spanning delayed-intention tasks and open-ended proactive assistance demonstrate consistent improvements in proactive task performance over representative memory approaches across multiple LLM backbones, while limiting unnecessary interventions.

具身智能体正越来越多地部署在开放世界环境中，但当前的鲁棒性评估仍偏向于危险指令和越狱漏洞。一个更根本的问题在很大程度上被忽视：智能体很少在语义充分但上下文极简的指令（即现实人类交流中常见的情况）下接受测试，这类指令缺少辅助性上下文线索。我们首次识别出一种称为“环境上下文幻觉”（Ambient Context Illusion）的新现象：当省略非必要的上下文细节时，尽管任务语义未变，智能体仍无法正确执行任务。为系统研究这一现象，我们建立了 VagueBench，这是一个基于 AI2-THOR 交互式仿真环境构建的基准，用于评估具身智能体的上下文依赖性。我们系统考虑指令的关键语义组件（实体、流程、约束和先验知识），并推导出一个包含四大类、九个细类的统一分类体系，以减少辅助信息。VagueBench 包含跨多种室内环境的 1,303 对指令，通过人机协同生成框架构建，该框架集成了语义冗余估计、分类体系引导的转换以及严格的语义不变性验证。通过对 6 个最先进的具身智能体进行广泛评估，我们观察到在 VagueBench 上任务性能和可执行性均显著下降（任务成功率相对下降 -19.65% a

由于应用场景广泛，手部定位问题长期以来一直是研究重点。该任务将手部建模为单个点，并确定其在给定坐标系中的位置。然而，受数据模态限制，现有手部定位技术面临若干挑战。例如，基于视觉的定位会引发隐私问题，而基于可穿戴设备的方法会降低用户舒适度。本文提出 mmHand，一种全新的非接触式、隐私保护的动态手部定位系统，利用单台商用毫米波设备实现像素级精度。我们首先提出 mmImage 生成工具，从原始毫米波数据中充分提取空间信息，并引入一种新颖的毫米波数据二维图像格式表示。接下来，我们设计了一个框架，为毫米波数据提供新的质量评估方法和像素空间标注。最后，我们提出一种跨模态空间特征增强模型，具备强大的空间特征提取能力，能够在毫米波雷达 U-V 像素坐标系中以像素级精度准确定位手部位置。我们在三种场景下对 12 名受试者进行实验评估，四项指标的结果证明了我们手部定位系统的有效性。

训练分类器意味着让验证指标决定保留哪个模型以及何时停止训练。对该指标的一个自然要求是优越性：每一个被正确分类的预测都应比任何被错误分类的预测得分更高，无论底层概率多么接近。Brier 分数和对数分数本身都不满足这一点。我们表明，优越性由两个端点之间的间隔决定——正确预测上得分的上确界与错误预测上得分的下确界——并对每种分数以闭式推导出二者。它们的差是最小惩罚，即恢复优越性所需惩罚的最大下界。先前工作仅根据第一个端点校准惩罚来恢复优越性；对于 Brier 基础，我们的最小惩罚比已发表的惩罚更小，缩小因子从三类时的 4.00 到二十类时的 2.11。我们进一步将已发表惩罚和最小惩罚嵌入一个单参数族，并表明在单值平局约定下，该族中的每个成员都是严格恰当的。在一项预先设定的双数据集案例研究中，所选模型的平均 macro-F1 在报告的十类情形中随惩罚增大而上升，不过在考虑种子/选择模式嵌套后无法与噪声区分；四十七类的第五种情形甚至无法复现这一描述性模式。预期的无优越性基线在代数上是一个重缩放 o

推测解码（SD）可加速大语言模型（LLM）服务，但其收益取决于工作负载的形态（如批大小和上下文长度）以及内容。基于模型的草稿器可针对任意内容提出短块，而无模型草稿器（如 n-gram 查找）在输出重复其输入时（例如代码编辑）能以几乎零成本提出长复制。自适应验证将整个批次的验证预算花在最可能被接受的草稿 token 上，但它仅支持一个基于模型的草稿器。无模型草稿器不提供逐 token 置信度，因此无法参与预算分配，部署时必须固定一个草稿器或在草稿器之间拆分 GPU。我们提出 \sys，一个在每一步同时运行基于模型和无模型草稿器的框架，并在自适应验证内、完全在 GPU 上为每个请求提供两者链中更优者，从而保持异步调度不受影响。关键思想是每个已验证草稿 token 的价格：该槽位在批次其他位置本可获得的有效吞吐量。一旦无模型草稿器的 token 携带校准后的接受概率估计（在通用文本上一次性拟合），这一价格就将两个草稿器的链放在同一尺度上。在 vLLM 中以 \dspark 和 n-gram 草稿器实例化后，\sys 在代码编辑、批大小为 4 时输出吞吐量达到 \dspark 的 1.10--1.62 倍，批大小为 128 时为 1.04--1.27 倍，而在温度为 0 的聊天场景中保持为其 0.97--1.01 倍。单一部署

在线 3D Gaussian 映射能够实现快速、高保真的新视角合成，在增强现实、机器人学和自主系统中具有广泛应用。然而，当前许多方法假设场景为静态，或将动态部分作为离群值过滤掉。本文研究视野内的场景变化，即在单次在线序列中物体在相机前方被操作移动；这与先前在相机视野之外或不同次采集之间检测变化的工作不同。这种设定将瞬态动态与持久动态纠缠在一起，因为物体在移动时可能成为干扰物，而一旦稳定下来又会变成持久物体。尽管近期研究开始通过移除干扰物以获得干净的静态背景来处理它们，但视野内场景变化仍缺乏充分探索。为解决这一问题，我们提出 SettleSplat，一个单目映射系统，可在这种变化下维护最新的 3D Gaussian 地图。具体而言，我们将粗略的逐像素不确定性提升为 Gaussian 动态性场，然后剪除瞬态部分。对于持久变化，我们推导出由无参数渲染深度不确定性加权的几何矛盾，以将真实变化与优化不足的区域区分开。在 Bonn 和 Co-Fusion 上，SettleSplat 实现了一致的 3D Gaussian 地图，在完整图像和变化区域上均优于当前基线，同时保持相当的速度。

Large language model (LLM) assistants increasingly advise backers on which crowdfunding campaigns to support. LLMs have known race and gender problems, and it is unsettled whether they reproduce, erase, or reverse the discrimination documented in entrepreneurial finance. We ask when, and through which decision process, LLM assistants use founders' gender and ethnicity. In a pre-specified choice-based conjoint experiment, seven LLMs chose among five synthetic campaigns with independently randomized quality signals and founder names, under three backer personas and a persona-free control (40,248 responses). By default, founder names barely matter: under budget-conscious and delivery-focused personas, gender and Black-name effects are statistically equivalent to zero. When the backer wants to support underrepresented creators, the models favor Black-signaled (+8.1 percentage points; positive in all seven models) and female-signaled (+5.1) founders. They do so most strongly among near-tied campaigns and for the founder who alone carries the signal, and they almost never name gender or ethnicity in their reasons. The results fit an agency account of algorithmic discrimination: an aligned assistant takes its demographic preferences from its user rather than from its training data.

Graph-aware calibrators make graph neural networks (GNNs) more trustworthy by predicting a temperature for each node, usually with auxiliary networks whose decisions are hard to inspect. We propose Interpretable Structural Temperature Scaling (ISTS), a twelve-parameter linear model that sets each node's temperature from named descriptors of its confidence, neighbourhood and closeness to labelled nodes, so every temperature traces back to descriptor contributions. ISTS preserves predictions, and cross-validated regularisation lets the calibration data decide how far the model departs from temperature scaling (TS). Its explanations are faithful at the level of node groups: on test nodes, ISTS softens groups that a global temperature leaves over-confident and sharpens under-confident ones (Spearman correlation 0.67 over 246 groups). On seven graphs with two backbones and eleven baselines, ISTS has the best average rank in negative log-likelihood (NLL) and Brier score and lowers NLL relative to TS by 1.8%, more than twice the gain of an identically fitted entropy-based temperature. It ranked first again on four held-out graphs under a protocol fixed in advance. From about 100 calibration nodes on, ISTS is a cheap and inspectable choice, and on point-cloud segmentation it significantly improves NLL, Brier score and the ranking of points for rejection.

Deep Research（DR）系统将大语言模型扩展为自主智能体，这些智能体能够分解复杂的信息需求、迭代搜索网络并综合证据以回答具有挑战性的问题。然而，其多阶段执行可能引入鲁棒性风险，因为用户问题或中间搜索输入的变化可能跨越研究阶段传播，并导致错误答案。现有评估侧重于任务解决能力或对外部证据扰动的鲁棒性，导致DR执行中的内在弱点未得到充分探索。因此，我们对来自三个代表性DR系统的1,932条执行轨迹进行了实证研究，并在子问题构建、在线搜索和最终推理中识别出九类反复出现的缺陷。在这些发现的指导下，我们提出了EMBERS，一个经验驱动、缺陷引导的模糊测试框架，用于DR系统的鲁棒性测试。从正确回答的种子问题开始，EMBERS使用缺陷引导变异、网络搜索辅助的保真度保证和多样性感知的种子维护来生成面向阶段的测试输入。在16个框架--模型--数据集设置上的实验表明，在相同测试预算下，EMBERS的失败发现吞吐量达到最强基线的3.21倍。人工和LLM辅助评估表明，大多数生成的输入保留了原始信息需求和必要的上下文约束。此外，模型 f

Why do language models fail on long inputs? The common answer is length. We show that, for retrieval, length is cheap and competition is not. On five open models (0.5B to 3B parameters, four families), neutral filler costs nothing: accuracy stays at 0.97 to 1.00, up to 32K tokens on Qwen2.5-1.5B. Statements in the format of the evidence, placed at the same positions, drop it to 0.16 to 0.33. A single statement that gives the queried key a rival value multiplies the loss by 9 to 110, and every additional rival adds to it. Inside the model, one quantity tracks the failure: the attention mass left on the evidence. Across 126,473 trials it explains up to 58% of the variance in loss; log length explains at most 8%. The masked-filler protocol behind the claim that length alone hurts reproduces on MMLU (11 to 18 points lost at 30K tokens), but fails a visible control in retrieval: masking costs 45 accuracy points at 32K tokens, while the same filler left visible costs none. In retrieval, context rot is a competition for attention. We release the probe and all 126K trial-level measurements.

基于扩散的数据增强已成为应对数据稀缺图像分类的一种强大策略。近期流程通常先将预训练的文本到图像扩散模型适配到有限的类别样本上，随后使用多样化的文本提示采样合成图像。然而，在极端数据稀缺场景（如长尾单样本设置）中，这种方法存在双重失败模式：生成的样本往往缺乏对分类器训练至关重要的判别性前景细节，且适配后的类别先验可能使早期去噪过程产生偏差，导致输出图像尽管文本提示不同仍呈现相似模式。为克服这些限制，我们提出一个采样侧框架，由两个互补组件组成，同时增强细节和多样性。首先，Mask-Guided Reference (MGR) Attention 通过零样本分割仅用参考特征更新前景 value tokens，将参考影响限制在目标对象上，并缓解过度的全局特征注入。其次，Repulsive Noise Optimization (RNO) 促使初始噪声轨迹偏离适配模型的主导生成模式，使目标提示能够表达更广泛的上下文变化。大量评估显示一致的改进，在 CUB-LT IF100 的长尾单样本设置中最大提升达 2.50 和 2.94 个点，并在少样本和背景偏移场景中具有进一步鲁棒性

工具描述告诉智能体某个操作能做什么，但未必能揭示该操作在使用当下是否健康或合适。执行框架通常通过监控器、追踪和近期故障掌握这些信息，而选择下一步动作的模型却并不掌握。我们将这一缺口称为操作盲区。我们提出 Agent Harness Awareness，一种信任分离设计：由主机在关键动作前提供一个小型、新鲜、经过验证的状态快照，同时保留测量、执行、评估和最终分派权限。我们使用 ToolRoute 测试其工具与网络实例，这是一个受控隔离基准，保持任务和候选语义不变，并改变模型可用的操作信息。在一项覆盖 GPT-5.6 Sol、Claude Sonnet 5 和 Gemini 3.7 Flash 的 216 次决策消融实验中，经过验证的工具健康状态将平均可观测策略遗憾较同形态中性对照降低 96.9%（从 4,252 ms 降至 131 ms），并将单次分派完成率从 69.4% 提升至 95.8%。在另一项包含实时本地 TCP/HTTP 执行、OpenTelemetry 客户端 span 和 Toxiproxy 注入故障的 27 次决策端到端复现中，经过验证的状态将平均遗憾从 8,956 ms 降至 0 ms（零额外可观测运营成本），并在 9/9 个模型×故障案例中完成实时动作，而中性对照为 5/9。这些研究确立了 Agent Harness Awareness 的工具与网络实例：选定的 hos

Deploying small language models on resource-constrained hardware requires compression, yet practitioners lack systematic guidance on how quantization, pruning, and post-compression recovery interact when combined. We present a factorial ablation study evaluating 72 compression configurations across four SLM architectures (2.7B-7B parameters), extended with a 48-configuration ordering analysis, benchmark accuracy validation on 30 compressed models, and robustness checks including multi-seed validation and full-corpus perplexity evaluation. INT8 quantization is effectively lossless (<0.5% accuracy degradation), while NF4 quantization combined with L1 magnitude pruning causes model collapse in all 16 tested combinations. LoRA-based fine-tuning reduces degraded perplexity by up to 97% after aggressive pruning; an ablation against standard fine-tuning without a teacher model shows the recovery is driven by LoRA adaptation itself rather than the distillation signal. Compression ordering shows no significant effect among viable configurations (ANOVA p = 0.96) but determines viability for NF4+pruning combinations. INT8 quantization alone provides 40-45% memory reduction with no quality loss. On-device CPU deployment via llama.cpp confirms these findings transfer off the GPU: 4-bit quantization runs 46-54% faster than 8-bit for a perplexity difference under 0.5 points, while un-structured pruning yields no CPU latency or memory benefit, consistent with its dense storage format.

Urban flood resilience emerges from complex interactions among natural, built, and socio-economic systems, yet most recovery policies implicitly assume that resilience mechanisms are spatially uniform. To address this gap, this study develops a spatially explainable machine learning framework that integrates geographically weighted random forests with a novel Local Spatial TreeSHAP algorithm, enabling nonlinear resilience mechanisms to be identified and interpreted at the community scale. The framework is applied to 779 communities affected by the July 2023 extreme rainstorm in Beijing. Results reveal that resilience mechanisms exhibit substantially stronger spatial heterogeneity than the underlying physical and socio-economic conditions themselves, with SHAP-value spatial autocorrelation decreasing by 66.2% on average relative to predictor variables. Three distinct resilience regimes emerge across the urbanization gradient. Highly urbanized communities benefit from synergistic interactions among commercial infrastructure and accommodation services; transitional medium urbanized communities are constrained by infrastructure-development mismatches that create facility gaps despite continued urban expansion; and low urbanized communities are dominated by ecological and topographic thresholds, where adverse interactions between transportation and river-network structures undermine recovery capacity. Field investigations and interviews with 12 stakeholders corroborate these patterns. By linking geographically weighted machine learning with locally grounded attribution analysis, the proposed framework provides a new pathway for uncovering spatially varying resilience mechanisms and translating them into place-based recovery strategies under increasing climate extremes.

科学发现建立在通过领域知识与推理来解读的准确观察之上。对于科学 3D 数据，这要求理解结构与空间关系，从蛋白质接触到解剖组织。现有多模态基准主要评估文本和 2D 图像，而科学 3D 评估在各学科间的表示形式和协议各不相同，限制了跨学科比较。为评估这些基础能力，我们推出 ScienceVision，一个通过专家级视觉问答（VQA）来评估科学 3D 理解与推理的跨领域基准。我们构建了 250 个问题，分别来自生物学、化学、医学影像、机械 CAD 和建筑学各 50 个，采用合成问题生成与领域专家审核和标注相结合的流程。每个任务将一个原生 3D 制品与一个科学问题、指定的回答格式以及任务特定的参考答案配对，要求将领域知识应用于几何或拓扑证据。我们在三种设置下用相同问题评估了 12 个模型：通过智能体框架的文件与代码执行工具访问原生 3D 制品的视觉问答；在相同工作流基础上，增加通过 Model Context Protocol (MCP) 暴露的、受专家启发的查看器检查与测量工具；以及使用智能体框架基于规定渲染的 2D 图像进行视觉问答。所有设置均使用相同的……

近期人工智能辅助数学的突破表明，开放式机器搜索能够超越人类数学家，但要将这一范式迁移到物理学，需要重构研究工作流。我们提出Simjecture——一个开源框架，其中语言模型智能体围绕人类定义的物理猜想开展研究。该框架解决了这一迁移中的两个核心挑战：第一，由于物理学是经验性的且没有终极模型，它将探究结构化为反例驱动的证伪与前瞻性假设修复；第二，由于物理理论必须与现实世界交互，它将计算物理学作为可执行的试验台，同时将数值工具的资格认定与科学主张审查分离，使求解器和诊断工具能够在有记录的调查过程中被调整。我们在一次历时82.6小时的调查中应用该框架，研究Z-pinch中辐射能超过内爆壳层动能这一长期存在的问题。调查揭示了能量转换、保留与释放时序之间相互竞争的效果。更强的相干螺旋变形会增加电阻性转换，同时减少辐射并增加保留能量。更宽的拖尾等离子体分布会延迟后期的辐射峰值，并弥补其早期量热观测中的大部分亏缺，其光谱响应在脉冲期间发生变化。这些结果为高能量密度下的能量转换提供了新的见解。

Neural controlled differential equations（Neural CDEs）沿通过插值离散观测构造的连续路径演化隐藏状态。稀疏测量使观测时刻之间的输入无法确定，而符合同一数据的可容许连续路径可能产生不同的终端状态。我们针对速度有界的平面驱动项刻画这种歧义。相容终端状态集合允许一个首阶 zonotopic 近似，该近似由可行有符号面积和传输后的 Lie brackets 决定。其到标量预测上的投影产生一个可计算、针对特定目标的歧义证书。我们还研究减少剩余不确定性所需的测量次数。对于非退化平面 CDEs，在速度有界时，确定性自适应采样策略下的最优最坏半径按 $N^{-1}$ 缩放；当速度和加速度均有界时按 $N^{-2}$ 缩放，其中 $N$ 是观测区间数。合成实验使用可容许连续补全评估该证书，并考察路径正则性的作用。

Tokenization is a fundamental component of modern language models: before a model sees data, its tokenizer has already determined how the data appears. We identify and challenge three myths in the tokenization research literature. Myth #1: Bits-Per-Byte (BPB) is agnostic to the tokenizer and/or the vocabulary size: BPB is tokens per byte times bits per token, and the two factors do not cancel; using untrained models, we show that BPB prefers some tokenizers over others: the model has learned nothing, but the tokenizer--model pair has. Myth #2: lower BPB means a better model: On untrained models, BPB strongly prefers some vocabulary sizes, even though accuracy remains at chance, and on real models trained on the same data, BPB correlates weakly with downstream performance. Myth #3: different "tokenizer-agnostic" metrics yield consistent results: The "unigram-normalized loss" is not tokenizer-agnostic; its reported correlation of 0.989 with BPB comes from a confounder and falls to 0.057 once fixed; from the same data the two metrics select different vocabularies at each model size. We present a tokenizer trilemma: no metric can score only the true distribution best, give the same predictions the same score, and tie all untrained models, so choosing a metric is choosing a research question. We conclude with a caution for reading the tokenization literature and a conjecture for future work.

Bias in transformer-based models is a growing concern as these architectures become more prevalent in automated decision-making systems, such as prediction models for hiring, lending, and clinical treatment recommendation. Mechanistic interpretability (MI) is a growing body of work that measures meaningful causal effects of activations on model output by identifying human-interpretable features encoded in model activations. When such features correspond to sensitive attributes, they provide a direct way to quantify the bias those attributes introduce and to causally intervene to mitigate it. However, counterfactual fairness (CF) methods have not been systematically adapted to this setting of intervening on interpretable mechanisms in transformers. We propose a framework, Mechanistic Interpretability for Counterfactual Fairness (MICF), that utilizes MI tools to reformulate CF constraints in terms of transformer activations, unifying former approaches into a single pipeline for fairness analysis in transformers. MICF does not require a prior causal model or specific transformer architecture. We evaluate MICF on real-world tabular data and show that CF can be improved across different prediction tasks.

Mixture-of-experts (MoE) models expand model capacity by activating only a subset of experts for each token. On accelerators with static execution plans, expert capacities and placement are fixed before routing reveals each expert's load. Uniform worst-case capacities waste work on padding; smaller capacities tied to expert identities can overflow. We present RankLane, which ranks each device's experts by current load and assigns them to fixed-capacity lanes. Runtime indices select experts and token rows while tier populations, shapes and device placement remain fixed. An aggregation graph preserves device ownership and reduces each token's local contributions on one core. Routing calibration and measured execution costs guide tier selection; overflowing chunks rerun at full capacity. We evaluate full-model prefill of Qwen3-30B-A3B and Gemma 4 26B-A4B with MXFP6 weights on four Cloud AI 100 Ultra SoCs, using 128-2048-token chunks and including overflow detection and in-program recovery support. Against a dense baseline, RankLane achieves end-to-end prefill speedups of 1.27-4.18x on Qwen3 and 1.93-3.46x on Gemma.

Large language model (LLM) agents increasingly rely on safeguards to prevent unauthorized actions while preserving legitimate service. However, existing risk assessments often focus on individual safeguard performance while overlooking dependence among safeguard decisions. This paper develops a dependence-aware framework for risk assessment and optimal design of multi-component AI safeguard systems, jointly selecting configurations and operating policies for sequential layered screening and selective verification. We establish stochastic comparisons of layered-screening error probabilities and invocation costs under lower-orthant ordering. For selective verification, we characterize exactly which policies have error probabilities and review costs that respond monotonically to stronger dependence under supermodular ordering. For a fixed first stage and reviewer, we show that the optimal routing map is a two-threshold rule in the first-stage likelihood ratio. An accompanying open-source software package, Dependence-Aware Policy Optimization Algorithms (DAPOA), facilitates the practical implementation of the proposed framework. Numerical experiments show that ignoring dependence can lead to inaccurate risk assessments and infeasible safeguard designs, with independence-based selections violating security requirements in six of eight scenarios. An empirical study using four LLM safeguards on benchmark-derived agent actions further demonstrates the practical implications of dependence modeling. Under the principal operating setting, dependence-aware selection lowers held-out total cost by 39.6% relative to independence-based selection.

In natural human-AI dialogue, users' corrections, requests, and self-descriptions carry reusable signals about how they want work to be done. These signals reveal individual process preferences, user-specific standards for evidence, depth, analytical focus, procedure, and delivery. We ask whether a preference revealed in one dialogue can shape an assistant's later answer when the task does not restate it. ProcessPrefBench uses three months of natural dialogues from two knowledge workers to build 66 source-target cases separately judged for source evidence and target applicability. Sources are preselected, so the benchmark tests application rather than retrieval. We test three models with target-only, full-history, and distilled-preference prompts. Target-specific 0-2 rubrics score preference realization, and a task gate flags severe failures. Compared with target-only prompting, full-history and imperative distilled-preference prompting raise preference realization by 0.204 and 0.244 points on the 0-2 scale, respectively, with no reliable advantage from distillation. Fabricated-evidence failures rise from 5.8% under target-only prompting to 10.1% under full-history and 10.4% under imperative prompting. ProcessPrefBench highlights a tension in learning from dialogue, where drawing on prior interactions can improve preference realization yet also invite fabricated evidence. We hope it helps build assistants that honor a user's process preferences without sacrificing factual reliability.

LLMs are increasingly used to design heuristics for computer systems. However, most synthesis frameworks evaluate candidates by their average performance, which can mask severe regressions on specific workloads. Even adding adversarial training inputs doesn't eliminate these regressions, and percentile objectives (e.g., optimizing for the 95th percentile) overlook how severe regressions can be beyond the selected threshold. We present TAILOR, a framework that guides LLM synthesis using the Conditional Value at Risk (CVaR) of the cost difference between each candidate and an existing baseline on the same workload. This paired CVaR measures the average difference over the worst fraction of workloads. We show that a negative paired CVaR guarantees improvements in both average performance and tail optimality gap relative to the unknown optimum, without ever computing that optimum. We find that TAILOR achieves the best mean--tail trade-off among the synthesis objectives we evaluate across three case studies: Azure GPU batch admission, Azure VM packing, and server placement. Tested across seven diverse optimization benchmarks, TAILOR produces heuristics that are certified with 99% confidence to improve both average performance and tail optimality gaps over the baseline in six of them. On these six benchmarks, TAILOR also outperforms known robust, human-designed alternatives in the tail while achieving better or comparable average-performance improvement over the baseline.

Reinforcement learning (RL) policies often outperform classical controllers in simulation, yet rarely reach production in safety-critical processes. The barrier is that there is no principled way to answer "Is this policy safe to deploy?" with statistical guarantees. We introduce DEPLOY-RL, a post-training certification framework built on one key insight: deployment certification requires discovering where failures occur (boundary discovery), not measuring how much everywhere (uniform reconstruction). Our contributions: (1) a contract-coupled acquisition function that concentrates sampling on certification-critical boundaries, achieving approximately 2x sample efficiency with a semi-empirical ambiguity reduction bound (domain-calibrated convergence guarantee); (2) conformal risk control providing finite-sample false-go guarantees ( 85% coverage among 14 baselines under our evaluation protocol.

AI systems embed majority-group defaults into training data, evaluation metrics, and category definitions, producing documented harms for queer communities including erasure, misclassification, and discrimination. Standard technical audits often rely on aggregate measures and cannot detect harms that be come visible only through the lived experience of affected communities. We conducted a participatory auditing workshop at EurIPS 2025 where 16 queer community members audited four case studies using the 4Cs harm taxonomy (Content, Conduct, Contact, Contract) applied across the AI lifecycle. Participants used structured worksheets and plenary synthesis to classify harms and trace them to their origins in the development pipeline. Across all four cases, participants traced harms to problem definition and data collection, and they identified contractual structures that extract value from vulnerable populations while providing minimal recourse. These findings illustrate that community-informed auditing surfaces concrete, identity-specific harms that aggregate evaluation methods risk overlooking.

Depth pruning removes entire transformer blocks and therefore depends on a reliable criterion for identifying which blocks can be removed. Block Influence (BI) ranks blocks by how minimally they rotate the hidden state. However, local redundancy does not necessarily imply downstream irrelevance. We address this limitation using the Jacobian Lens (J-Lens) neural artifact, a map transporting perturbations at a layer to the final residual stream. At a common sparsity level, pruning scores applying J-Lens transportation to $BI$, what we call $J-BI$, were found to improve post-pruning accuracy by up to 19 points over $BI$ while reducing damage to perplexity for Qwen3-8B, Llama-3.1-8B-IT, and Gemma-3-12B-IT. We also perform a single-block ablation study and use J-Lens probing to explain why the proposed scores improved performance.

Most mixing-based augmentation methods improve generalization through a simple idea: combine two training samples with a fixed global ratio or a hard spatial replacement. While effective, this design leaves little room for local variation, which becomes a more visible limitation in long-tailed recognition, where minority classes are especially sensitive to the quality of augmented samples. We propose TopoMix, a training-free augmentation method that replaces fixed mixing with continuous spatial mixing. TopoMix constructs a smooth spatially varying mask from a low-resolution random control grid, bicubic interpolation, and stochastic non-linear shaping, and uses this mask for element-wise image mixing and label reweighting. Without introducing trainable parameters or auxiliary networks, TopoMix offers a simple way to make mixing more flexible in the input space. Experiments on standard long-tailed benchmarks show that it consistently improves over representative mixing-based baselines, indicating that smooth spatial mixing is an effective augmentation strategy for long-tailed recognition.

Next-visit diagnosis prediction forecasts the diagnoses documented at a patient's next encounter, turning retrospective electronic health records into prospective clinical prediction. However, conventional aggregate evaluation obscures its multifaceted nature: distributionally, models must handle variable-size diagnosis sets and long-tailed labels; temporally, they must distinguish persistent, newly recorded, and recurrent diagnoses; and both vary with observed history depth. We introduce an evaluation framework along these dimensions and apply it to rule-based heuristics, LLM-based prompting, retrieval, memory, and multi-agent approaches. This reveals that repeating the last visit outperforms most LLM-based approaches overall despite predicting no new or recurrent diagnoses, and that retrieval and memory exhibit distinct profiles whose relative strengths vary across temporal transitions and backbones. Motivated by this, we propose LONGINUS, a retrieval-memory multi-agent adjudication framework combining evidence retrieved from patient records with structured memory of the patient trajectory. Using cross-view agreement as a reliability heuristic, it preserves shared predictions and adjudicates disagreements without revealing their source. Across MIMIC-III and MIMIC-IV with four LLM backbones, LONGINUS achieves the best transition-specific performance among LLM-based methods in nearly all settings while remaining competitive with them on aggregate metrics.

几何证书会对指定的估计器对象作出数学断言，而机器人动作需要物理变换、服务决策和及时的证明支持。我们提出 CARD，一种携带证明的接口，它使这些对象保持彼此区分，并记录它们如何组合。其提升纤维表述将语义运动与见证修复分离，并刻画投影成本与携带证明成本何时一致。其可行动性结果给出一个精确的有用服务定律：无放回均匀采样产生超几何产出，而在观察到的实时确认条件下，从合格总体中采样每次请求给出一个有用服务。主要研究包含来自 AA15-X500 机载测量系统 20 架次的 404 个 Full16 案例。所有 404 个经检查器确认的优化器点均与参考不一致，旋转和平移差异中位数分别为 1.9891 度和 21.6322 m。配对的因果导航变换在所有 404 个案例中均与参考一致，中位数分别为 0.0036566 度和 0.0104519 m。在五个匹配的请求预算下，超几何预测与观测到的普通随机均值之间的最大差异为 0.8412 个服务。普通匹配随机达到 0/50 目标成功，而可行动性匹配随机达到 50/50，并与 CARD 策略匹配。在 11 个 KITTI、EuRoC 和 TUM SLAM 案例上的公共下游评估测得，对于 random v，优化器墙钟时间比率为 1.128x-67.65x

World models predict how scenes evolve, but they are difficult to interrogate: given a model that forecasts a video, one cannot easily ask what would have happened had a particular interaction between two objects been different. We present Factorized Object Relations for Counterfactual Editing (FORCE), a world model that infers pairwise interactions from raw video without supervision and represents each as an explicit latent per object pair, so that a relation can be removed, changed, transferred or introduced directly. Objects are extracted by slot attention over frozen self-supervised features; a relation encoder infers each pair latent from a short observation window, and an interaction head turns it into a force. The latent is held fixed across the rollout, so an edit persists rather than being recomputed, and every edit acts on the model's own quantities with no mapping to simulator parameters. We evaluate in a simulated environment where the true outcome of an edit is available, scoring the change each edit induces against the change the simulator produces under the same operation. On held-out clips, FORCE's edits made from video agree with the simulator's response at 0.19 to 0.37 on our effect-based score; trained and scored on exact object positions instead, the same architecture reaches 0.57 to 0.85. The gap therefore lies not in the learned relational dynamics but in perception: what bounds an edit is how precisely objects can be located from video, and the precision editing needs lies just beyond what current object-centric perception delivers. Closing that gap is a concrete target for making counterfactual world models work from video.

患者与初级保健提供者之间的高质量匹配可改善下游健康结果，但提供者流动会破坏这些匹配并迫使患者被重新分配。面临这一问题的卫生系统不仅必须决定推荐哪些提供者，还必须决定提供多少选择。直接分配避免了对稀缺提供者能力的竞争，但依赖于对患者偏好的噪声估计；更广泛的选择使患者能够利用私人信息纠正这些估计，但可能造成代价高昂的协调失败。我们通过设计固定的个性化菜单来研究这一权衡，患者从菜单中选择提供者。我们证明构建最优菜单是 NP-hard 的，并建立了一个信息论下界。我们发现，为每位患者提供一个单一提供者可在常数因子范围内达到该下界，而无限制选择可能表现任意差。然而，我们表明当系统无法可靠区分患者的各个备选方案时，额外选择会优于单一提供者分配。我们通过两种损失形式来形式化这一权衡：菜单损失（当菜单提供的选项过少时产生）和协调损失（当稀缺能力暴露于过多选择时产生）。受这一分解启发，我们开发了 Scenario-Averaged Marginals (SAM) 算法，该算法基于提供者影子价格定制菜单，从而仅在效用超过能力成本时提供选择。我们识别出 suffi

How much phonological knowledge do language models (LMs) possess about Chinese characters, and how is this knowledge represented and used? We introduce PinyinBench, a controlled benchmark covering homophones, phonological features (initials, finals, and tones), and polyphonic characters. We investigate 17 text-only and multimodal LMs through three complementary approaches: behavioral evaluation, layer-wise probing, and activation steering. Behaviorally, models demonstrate substantial knowledge of initials and finals, with a frontier model (DS-v4-flash) scoring 90%+ in homophone judgment, when two characters share no phonological features. However, its accuracy drops sharply to 72% when two characters differ only in lexical tone. Probing reveals that tone is highly decodable from embeddings and improves modestly across layers, whereas segmental information starts less decodable and improves substantially with depth. Activation steering shows that segmental representations can systematically influence generation, while tone exhibits much weaker effects. Furthermore, audio-bearing training improves phonological decodability, including tone, without corresponding improvements in steering effectiveness. Together, these findings reveal distinct representational and functional profiles for segmental and tonal information, with tonal information much harder to access and manipulate, demonstrating the necessity for combined evaluation, probing, and causal intervention for understanding linguistic knowledge in LMs.

Dimension reduction methods are chosen by a common claim: t-SNE and UMAP preserve local structure, and PCA preserves global structure. Checking it on given data takes local and global scores on one scale and a test that compares two maps. Both scores come from one curve. The locality profile is the chance-corrected share of each point's k nearest neighbors that a map keeps, over all neighborhood sizes. Its value at k = 10 is the local fidelity L, and its mean over 30 to 50 percent of the data is the global fidelity G. We prove that at the midpoint the profile is the agreement on which half of the data is near each point, a rank correlation of the distances. Because L and G are one statistic, two maps can be tested against each other. The LG reversal test applies to any two dimension reduction methods. It decides, with a block-jackknife p-value, whether one map is more local and the other more global, and returns the scale at which to change maps. On 156 maps of twelve datasets with up to 1.38 million points, G agrees with established global scores at 0.89 and 0.96, and 2,000 points estimate it. The test confirms the usual claim on seven datasets. On both retail datasets and both rolled-up manifolds, it finds the neighbor embeddings more global. In an agent memory of one million passages, the two maps cross at 150,000 passages (p = 2.6e-5). Applications span pathology, single cells, supply chains and AI agent memory.

Forecasting emotion in conversation is increasingly framed as predicting the next emotional state's value. We show this targets a hard quantity in a state-observed testbed and propose a learnable, deployable reformulation instead. On valence-arousal-dominance trajectories from a dyadic corpus, repeating a speaker's own previous state sets a persistence baseline that no frozen-encoder correction meaningfully improves upon and a state-space filter exceeds; the floor requires the historical state to be observed, since an estimated anchor collapses persistence below a direct forecaster. What survives is the timing of a departure from persistence, learnable on three of four corpora and near chance on the highest-base-rate corpus, a negative result we report. We recast the task as selective departure forecasting: a lead-time-weighted metric with a leakage protocol, and commitment policies whose best choice switches with the departure rate. A dialogue-level conformal risk control bounds the expected loss of committed predictions, meeting the target in all twelve configurations tested, though the guarantee is bought through abstention, with coverage and alarm recall falling sharply on the hardest corpus. One tested language model's confidence fails as a ranking signal where another succeeds, a model-dependent failure not yet attributable to family versus scale. Code and result logs are available from the corresponding author upon request.

计算药物发现提供了越来越多用于提出和评估分子的工具，但将这些工具组合起来并决定下一步该做什么仍然需要大量专家投入。在此，我们介绍 El Agente F'armaco，一个 AI 科学家，它利用现代大语言模型（LLM）技术，将用日常语言表达的发现目标转化为复杂、可扩展的计算筛选工作流。该智能体框架会选择科学工具、审查其结果，并根据目标和用户需求调整后续步骤。我们的方法以一个共享科学记忆为核心：F'armaco Atlas，它将分子、结构和结果连同其来源——即获得方式的记录——一起存储。智能体使用有针对性的数据库查询来检索每项决策所需的证据。单个请求可以为整个分子集合启动计算，而无需为每个化合物单独进行 LLM 决策。这种设计限制了随着筛选集合增长模型必须处理的信息量，同时保持底层结果可供检查与进一步分析。集成工具支持分子生成、结合与分子性质预测，以及搜索可购买化合物。在与专家主导的研究活动进行的回顾性比较中，智能体提出了化学上不同的候选分子，具有有竞争力的对接评分，并在明确施加的 s 条件下具有更高的计算药物相似性。

When a language model decides refund or claim requests under written policies, the system can show it similar past cases. Each past case has two parts: its facts and the decision recorded for it. We separate the two. A model sees a policy, a request and three past cases, and we change either the recorded decisions or the facts, keeping the request fixed. We select requests that are close calls: deciding without past cases, DeepSeek V4.1 Flash approved each in some of four tries and denied it in others. When the past cases are near-copies of the request (its facts with a few details changed), switching their recorded decisions from DENY to APPROVE raises approval by about 30 to 55 percentage points in DeepSeek V4.1 Flash, GPT-5.6 Luna and Claude Haiku 5.5 on policies written for the study, and DeepSeek and Luna show the same on nine public policies paraphrased from real ones. The facts push the other way: shown without their recorded decisions, past cases whose facts support approval make these two models approve less often than past cases whose facts support denial. One reading is that the request looks weaker next to them. This push is not just a correction toward the right answer: in an additional DeepSeek analysis, it was of similar estimated size on requests that human annotators judged approvable and on those they judged deniable. In a separate DeepSeek run on the public policies, when the past cases are near-copies of a different request, switching their decisions moves approval by 9 points instead of 43, while facts that support approval still lower it by 22 points against facts that support denial. Systems that retrieve past decisions should test the two parts separately.

由于对海量医学图像进行人工标注的成本高昂，自监督学习（SSL）受到了广泛关注，并在多种医学影像任务中展现出前景。在 SSL 方法中，对比学习已成为一种突出的方法，鼓励模型编码在不同增强视图之间保持不变性的语义信息。然而，对于医学图像而言，这种不变性可能是虚假的，因为在增强过程中细微但关键的医学特征可能被抑制。本文通过将学习到的特征嵌入编码为概率分布而非固定向量来缓解该问题，从而捕获更丰富、更细致的信息谱。此外，我们提出了一个统一概率框架，无缝集成多种 SSL 方法。我们的方法在多个数据集上的多种下游任务（包括医学图像分割和分类）中优于最先进的 SSL 方法。我们进一步通过针对更强增强的鲁棒性、有限微调标签下的标签效率以及计算权衡的额外分析来验证该方法。代码发布于 https://github.com/wwzjer/ProbSSL。

A single camera cannot distinguish between a slowly moving object nearby and a rapidly moving object in the distance without additional information. For a shuttlecock, air drag provides a strong deceleration cue that can be combined with known court geometry. We fit a gravity-and-drag trajectory to smartphone video and compare a free monocular fit with a fit constrained to cross the net plane. Camera calibration uses court lines, and the trajectory is fitted to observed shuttle positions and frame timestamps by robust nonlinear least squares. We evaluated 51 badminton shots from two recording sessions, spanning drop shots, drives, half-smashes and smashes. The net-plane fit reduced the median absolute speed difference from a drag-based stereo reference from 13.3% to 8.6%. A second reference, obtained by differentiating triangulated positions without a drag model, also gave a smaller median difference for the net-plane fit. The results support combining court geometry with flight dynamics for badminton speed estimation from a single camera.

人类能够在假设的运动下想象场景变化，这一能力是空间推理和机器人规划的核心。然而，这一能力在视觉语言模型和世界模型中仍未得到充分探索。我们将这一能力形式化为反事实空间想象（Counterfactual Spatial Imagination, CSI），并推出 FlowBench——一个用于训练和评估上述两类模型的资源，包含 42.8 万个训练样本，覆盖仿真与真实场景、家庭环境与桌面操作，以及多样的问题类型。一个可扩展的全自动流水线为每个假设运动配以空间问题，并在仿真中提供精确的结果帧和光流。对于 VLM，专有模型表现出“想象鸿沟”：在需要推断运动结果时，其表现比直接观察时更差，而开源模型的表现接近随机水平；在 FlowBench-Sim-Train 上微调 Qwen3-VL-8B 后，其在仿真中的准确率提升了 29.5 个百分点，在真实图像上提升了 16.9 个百分点，展现出强大的仿真到现实迁移能力。对于世界模型，在 FlowBench-Sim 上训练可改善我们测试的全部四个模型系列的文本到光流预测，将 Psi-0.6 的准确率从 52.1% 提升至 61.7%。因此，FlowBench 可作为这两类模型的训练与评估资源。代码和数据集已在 https://github.com/liuyifan22/FlowBench 发布。

Existing Graphical User Interface (GUI) benchmarks predominantly rely on precise and stateless user instructions. In everyday usage, however, instructions are often \emph{vague and history-dependent}: users omit the exact details of their past operations because they no longer remember them, leaving the agent to recover the missing context from long and noisy interaction histories. To bridge this evaluation gap, we introduce \textbf{MemU-Bench}, a mobile benchmark of \textbf{118 vague tasks} grounded in 29 long-horizon user trajectories, where agents must distill raw interaction streams into memory and retrieve the missing context on demand, with task outcomes verified against serialized application databases. Alongside Success Rate, we incorporate Interaction Efficiency (IE) and Success weighted by Path Length (SPL) to expose agents that succeed by exhaustively searching across different applications rather than by \emph{genuinely recalling} the history. Evaluations show that without user history, state-of-the-art systems struggle under such ambiguity: the best-performing Gemini-3.1-Pro achieves an SPL of only 16.4\%, and end-to-end GUI agents fall below 5\%. To address this, we further propose \textbf{AppMem}, a training-free framework that distills user history into a \emph{structured, app-centric} memory bank, tracks interrupted intents across application switches, and selectively retrieves relevant memories via target-app inference. AppMem raises the SPL of GPT-5.5 from 15.7\% to 66.9\% (\textbf{+51.2\%}), outperforms a naive summary baseline by 9.3\%, and delivers consistent gains across evaluated backbones.

Temporal sentence grounding requires a video representation that preserves query-relevant evidence and a sentence representation that reflects the visual context in which it is interpreted. Prior end-to-end grounding addresses the first requirement through sentence-conditioned visual adaptation, but leaves the language representation independent of the video. This work proposes Multimodal Cross-Condition Learning (MCCL). Its Video Prefix Adapter (VPA) converts pooled visual context into a sequence of continuous tokens that condition a frozen DistilBERT encoder. Combined with the Sentence Conditioned Adapter (SCADA), VPA supports an ordered exchange from vision to language and back to vision. A comprehensive study examines two video backbones, LSTM and Mamba temporal modules, and four conditioning modes on two datasets. Bidirectional conditioning exceeds the stronger one-way mode by 1.6--2.3 mIoU points in all four Charades-STA settings. ActivityNet Captions exhibits a different pattern: bidirectionality helps I3D--Mamba, whereas video-to-text conditioning is stronger with C3D. The study extends one-way visual adaptation into a configurable multimodal framework and reports descriptive gains over the corresponding baselines.

在评估 LLM 生成的代码时，“已验证”究竟应当意味着什么？许多现有流水线只报告一个单一的验证数字。我们认为，这种做法掩盖了在条件审计序列中测得的三个不同属性：验证器接受（Surface A）、\emph{语法规范绑定（Surface B：证明义务是否绑定到规范？）}，以及 \emph{语义规范保真度}（Surface C：规范是否捕捉到了非形式化需求？）。两个实证演示表明，这些层面是可分离的。\textbf{第一（$A \neq B$），通过 Li@Opus（在 Claude Opus 4.7 下对 \citet{li2025dafny} 的 Dafny-as-intermediate-representation (IR) 流水线进行的同模型复现）在三个 Python 基准上}：它在 HumanEval 上达到 \textbf{99.1\% 的验证器接受率}，而只有 \textbf{1.9\%} 携带绑定到规范定义的谓词或函数的证书。\textbf{第二（$B \neq C$），通过 VCG 在两个 Rust 基准上}：我们在 Rust 上使用 VCG（这是我们自研的 Dafny-as-intermediate 流水线），因为 Rust 目标让我们能够强制产出完全可验证的输出（每个组件都以 Dafny 证书形式交付，通过构造即确保绑定门通过）。即使在这一构造内部，\textbf{11.48\%（MBPP-Rust）和 14.82\%（VeriContest LC 切片）的结构化验证样本仍然无法通过隐藏测试}。在 VeriContest 的 LC 切片上，\textbf{VCG 报告的 End2End 率为 7.79\%（535/6870）。}引用的外部比率为 5.51\%（GPT-5.5）和 1.88\%（单次 Opus 4.7）。这些给出了描述

在不断变化的环境中，基础模型可能会遇到其原始训练未覆盖的陌生领域知识。然而，新领域很少提供大量由专家创建的训练数据或可靠的监督信号来进一步训练。我们提出 Recursa，这是一种基础模型系统，无需人类专家或更强的教师模型，就能通过从自身经验中学习来适应新领域。Recursa 使用知识库来存储所学内容：领域特定技能以及何时应用这些技能的指令，还有过往任务经验的记忆以及关于改进方向的反思。Recursa 从领域材料中自举构建知识库，然后通过递归自我改进对其进行精炼。它使用并行学习器执行任务并合成局部改进，然后验证这些改进并将其合并到知识库中。在 Harvey LAB 的 Firm Knowledge 基准上，Recursa 使用 DeepSeek-V4.1-Flash 将留出测试任务的全通过率从 30.0% 提升至 55.0%，使用 Qwen3.8-Flash 从 35.0% 提升至 46.0%。

多模态脑MRI可提供互补的解剖、组织和病理信息，以支持下游人工智能（AI）应用。然而，在公开和汇总的神经影像数据集中，各模态的可用性并不均衡：T1加权（T1w）MRI广泛存在，T2加权（T2w）和液体衰减反转恢复（FLAIR）MRI的可用性相对较低，而钆增强T1加权（T1Gd）MRI则较为稀缺。因此，汇总数据集包含具有不同已采集MRI对比组合的受试者，这反映了各贡献研究在研究设计、采集协议以及临床或研究目标上的差异。许多缺失模态基准通过从完整的四模态病例中扣留对比来模拟不完整MRI，这可能无法代表汇总数据集中观察到的可用性模式。为弥补这一空白，我们提出C-WFDM，这是我们的Wavelet-Fusion Diffusion Model（WFDM）的ControlNet扩展，用于受试者特定的缺失模态补全。C-WFDM构建于早期BrainScape快照之上，该快照包含157个公开MRI数据集和45,880次预处理扫描。预训练的WFDM主干保留目标模态和元数据条件，而目标特定的ControlNet适配器使用同一受试者的MRI对比来指导合成。C-WFDM从T1w合成T2w，从T1w和T2w合成FLAIR，并从T1w、T2w和FLAIR合成T1Gd。已采集的条件对比在可用时会被使用。

Speech transcription errors can be difficult to recognize when competing word hypotheses are unfamiliar or visually similar. We describe an interface that combines alternative transcripts, replay and contextually checked pictograms for selected uncertain concepts. The system will initially support French and Spanish functional utterances. A counterbalanced study will assess whether pictorial feedback helps listeners recover the speaker's intended action beyond text alternatives alone. We will also quantify semantic mismatch and additional interaction time, since an inappropriate image can reinforce the wrong interpretation rather than repair it.

Tell an LLM negotiator to be "headstrong" and it should hold firm on price, but does that instruction retain the same behavior in every language? To test this, we use identical buyer and seller instances of Gemini 3.6 Flash to negotiate the same 12 CraigslistBargain scenarios in English and Korean. The pilot-scale 2 x 2 x 3 design crosses interaction language, seller urgency (none or strong), and buyer persona (cooperative, neutral, or headstrong), yielding 432 attempted negotiations. We call this property cross-lingual social-policy fidelity: whether a prompt-defined social policy produces the same behavioral effect regardless of interaction language. It often does not. Korean negotiations reached agreement far more often than English (77.3% vs. 52.5%, a 24.8-point gap, Holm-adjusted p = .0039), but only for neutral and headstrong buyers; cooperative buyers were already near-ceiling in both languages. Buyers' opening offers barely differed by language; the gap instead emerged later, through larger seller concessions and a smaller final gap in Korean, and agreed prices were also lower in Korean. Repeating the full design with Qwen3.7 Flash as both agents gives directional support: agreement was again higher in Korean (+13.9 points vs. Gemini's +24.8, with an interval whose lower bound is near zero), and sellers again conceded more. Pragmatic analysis reveals why: while both languages frequently featured face-threatening acts like rejections and counter-offers, Korean dialogue softened these threats with substantially more politeness mitigation, especially among sellers (11.7% vs. 3.6%). These results show that such a policy can produce different simulated negotiation behavior, not just different wording, when only the interaction language changes.

开放集目标检测（OSOD）在密集候选框机制下存在一个根本性局限：与同一对象相关的高度重叠候选框往往被赋予相互矛盾的已知/未知预测。现有方法对每个候选框独立进行开放集推理，而忽略了相关候选框之间共享的语义，导致检测碎片化和开放集决策不稳定。本文将 OSOD 重新表述为以对象为中心的语义推理问题，其中相关候选框共同支持一致的区域级语义决策。基于这一视角，我们提出 OCSI-Net，一个统一框架，将以对象为中心的语义一致性与几何感知的不确定性稳定化相结合。具体而言，OCSI-Net 将相关候选框聚合为结构化区域簇，并从聚合的语义证据中推断区域级语义，从而在同一对象区域内强制执行一致的开放集预测。此外，引入了几何感知的不确定性机制，以稳定开放集过渡边界附近的优化。在多个基准上的大量实验表明，OCSI-Net 始终达到最先进性能，相比第二优方法，已知和未知检测的调和平均值平均提高 6.83%。

We introduce Dory1.0-4B, a 4-billion-parameter looped (latent-recurrent) language model with elastic depth: a single checkpoint serves different inference compute budgets. Dory1.0-4B is built on a normalized Transformer (nGPT) backbone with interleaved sliding-window and full-attention layers, organized into an input block, a shared recurrent block, and an output block. At inference, the recurrent block can be applied a variable number of times. We pre-trained Dory1.0-4B on 6 trillion tokens. The first 4.5 trillion tokens were trained in NVFP4, and additional pre-training then extended the context length to 256K tokens. Next, we post-trained the model using Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR), with a difficulty-aware loop curriculum that routes harder prompts to deeper loops. Increasing the loop count improves accuracy, with the largest gains on coding, agentic, and long-context benchmarks. With two or more loops, Dory1.0-4B outperforms larger models such as Qwen3.5-9B, Gemma 4 12B, and Gemma 4 26B-A4B on average across 12 benchmarks. Looping also makes Dory1.0-4B more token-efficient on agentic benchmarks: with four loops, it generates 1.3-2.5x fewer tokens than with one loop, as the extra computation is spent in latent space rather than on longer outputs. To reduce memory during looping, Dory1.0-4B caches keys and values only from the last loop, keeping its KV-cache footprint equal to that of a non-looped model at any loop count while retaining most of the accuracy gains. We open-source the BF16, MXFP8, NVFP4, and GGUF quantized checkpoints.

Attribution graphs, an emerging tool in mechanistic interpretability, use transcoders to decompose language model computations into sparse interpretable features connected by causal edges. However, turning a graph into a safety-relevant insight requires hours of manual analysis by experts. We introduce Circuit Oracle, an off-the-shelf LLM equipped with a multi-agent harness and task-specific authored skills, which answers natural-language questions about a target model by reading the attribution graph through tool calls. We evaluate it on three safety-relevant proxy tasks: detecting spurious features in probe circuits, eliciting hidden knowledge from taboo-finetuned models, and suppression jailbreaking via feature-level interventions. Beyond task scores, the oracle produces pruned subgraphs and context-specific feature explanations that we assess qualitatively. Results are mixed. On spurious-feature detection, Circuit Oracle reaches 78.2 $\pm$ 2.4% accuracy, outperforming single-layer feature-ranking baselines. On secret elicitation, trained self-explainers like Activation Oracles remain the stronger baseline. On refusal suppression over 50 prompts, its single committed intervention is slightly below a diff-in-means baseline, and only exceeds it when allowed the best of five attempts. Results are stable across five runs and across model choices for the agents. We also release the pruned subgraphs and context-specific feature readings that revise existing global labels, and invite future work to develop evaluations that can measure their usefulness directly.

Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles. Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy. In this paper, we present TileWeave, which partitions sparse attention by this cross-query reuse: highly shared key blocks run in coarse NPU tiles, while low-reuse blocks run on the GPU only for the query blocks that selected them. In addition, we design an NPU-native pooled-key selector that keeps selection inexpensive, and a runtime that publishes GPU partials in the order the NPU consumes them and merges them in its existing output store. On RULER, HeteroSparse matches XAttention's accuracy with Qwen3-1.7B and improves it with Llama-3.2-3B-Instruct; its mean LongBench scores match XAttention as well. On Snapdragon SM8750 and SM8850 phones, it achieves 1.35--3.23$\times$ faster full-model prefill than XAttention over 1k--16k tokens and up to 1.58$\times$ faster than dense attention.

Hundreds of languages have written text but almost no translations, and whether a model learning one from untranslated text builds on a related language or familiar letters is unclear: they usually coincide. Here we show that familiar letters, not a related language, let an unseen language into a shared space of meaning. GlyphMind, a tokenizer-free vision transformer that reads text as pixels and lets any sentence be shown in any script, was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, a language and script it had never seen, it retrieved English translations seven times as often as the strongest simple baseline. Rewritten in unfamiliar letters, Lao did not enter although its close relative Thai was in pretraining; Santali, which had none, entered in Devanagari, familiar from Hindi, but not in its own script. Transliteration into familiar scripts may thus open multilingual models to languages without translations.

虚拟试验利用常规收集的临床数据来评估治疗策略，但将研究方案转化为可执行的分析，需要确定临床概念和研究要求应如何在可用数据中表示，并在整个工作流中一致地应用。这些决策取决于本地数据表示和操作惯例，可能涉及数据源选择、患者和就诊身份、关联以及时间解释。本文介绍 VirtualTrialHarness（VTHarness），这是一个 agent 框架，通过为每个工作流阶段选择可复用的流程性技能，并使用显式契约在各阶段之间保持一致的研究决策，从而协调虚拟试验的可执行化。我们还提出 VTBenchmark，用于评估数据源选择和队列构建。在三个复杂的合成环境中，VTHarness 在需求澄清后成功生成了所有必需的队列输出，并在 95.1% 的工作流中与参考结果匹配，而使用同一流程语料库进行检索的方式为 60.4%。流程性技能在三个语言模型上改善了数据源恢复，并在 13 个 Mayo Clinic 案例中提高了与分析人员代码的数据源重叠度。站点特定适配使八个 MIMIC-IV 任务能够实现与参考匹配的队列执行。在 Mayo Clinic Virtual Trials Program 中，该框架支持研究者对队列构建和终点分析进行审查。这些发现支持协调

一条被存储的经历可能回答某个未来问题，却对另一个问题帮助甚微。我们衡量这种受任务条件影响的效用，而不是给每条记忆分配单一的重要性分数。一个任务盲的写入器在看不到后续任务的情况下存储记忆；随后一个冻结的读取器在有该记忆和无该记忆的条件下回答该任务，并使用完全相同的填充内容。成本账本记录用于形成、存储和暴露该记忆的资源。在基于 LongMemEval-V2 web-agent 记忆构建的单次调用事实回忆任务上，同一记忆在其包含回答所需信息时帮助更大。在这些含有答案的任务对上，其平均配对奖励增益比匹配的对照对高出 $+0.33$（95% 区间 $+0.21$ 至 $+0.46$）。所测试的任务盲摘要和结构化记录产生的增益小于 Raw，包括在固定总上下文长度下；预算扫描的结果与“保留含答案内容对这一差异有贡献”相一致。在随机化记忆位置的新会话中以及第二位读取器下，含答案与对照的对比仍然为正，尽管单个记忆—任务标签会发生变化。通用检索的表现高于随机水平，但漏掉了大多数由筛选定义的含答案记忆。事实回忆增益并未扩展到所测试的软件工程和科学编程基准，在这些基准上，即使原始记忆也没有可测量的效果。我们提供了一个公开、版本化的配对测量与成本数据集 Conditional Memory Utility Dataset。实际结论是

Small language models (<=9B parameters) can support local medical question answering (QA) while reducing privacy and latency concerns, but effective fine-tuning remains challenging. This paper compares two chain-of-thought (CoT) supervised fine-tuning methods: Uniform-across-Sequence (UaS), which applies equal loss weighting to the full CoT-and-answer sequence, and Segment-weighted (SwS), which separately weights CoT and answer losses. We evaluate six general- and medical-domain models on multiple-choice QA (MedQA and EMS-MCQA) and free-text QA (Med-O1). Results show that neither method is universally best: on multiple-choice QA (MedQA and EMS-MCQA), SwS generally performs better, with general-purpose models benefiting most; notably, Qwen3.5-4B improves by up to 44.8% over its zero-shot baseline on MedQA. In contrast, medical-domain models show smaller or even negative gains on MCQs. On free-text QA (Med-O1), UaS performs best for nearly all models and provides consistent improvements across both model types, with the 3B Llama-3.2 model approaching the performance of much larger 7B medical-domain models. We also find that excessively long CoT generation is associated with lower performance in some models on either multiple-choice or free-text QA tasks, suggesting that limiting the CoT length during training may benefit selected models. Overall, the results show that CoT loss weighting should be matched to the task format and the knowledge already encoded in the base model, rather than using a single fine-tuning strategy for all clinical QA tasks.

语言模型常将一个概念分散到许多内部维度上，使其特征难以解释。稀疏自编码器（SAEs）在训练后能恢复更清晰的特征。我们探究模型能否改为直接学习可解释的输出坐标。我们提出 Mixture of Wide Experts（MoWE），这是一种稀疏前馈层，其中被选中的专家读取完整表示，但写入固定且互不相交的输出坐标。在隐藏原因已知的受控任务上，该设计在保持预测性能的同时，比稠密模型和标准混合专家（MoE）模型更清晰地将原因定位。在大致匹配计算量的情况下，我们为每种架构训练两个 19 层运行，各 100 亿 token。MoWE 的评估交叉熵为 2.417，稠密模型为 2.452，而标准 MoE 达到 2.22。我们通过让语言模型评判器根据每个坐标强烈激活的段落来描述该坐标，并在新段落上测试该描述，来衡量原生坐标可读性。对同样适用于其他坐标的描述进行校正后，MoWE 的特异性得分为 0.523，稠密模型为 0.375，标准 MoE 为 0.268。在相同 FFN 更新上训练的 TopK SAEs 在稠密模型上得分为 0.489，在 MoWE 上为 0.556，使原生 MoWE 与稠密 SAE 相当。在第一个案例研究中，我们识别出一条四坐标因果路径，支持模型在提示词 A right angle is 90 之后对度数的预测。随后我们复制七个与角度相关的坐标

Test-time search can improve structured planning, but when methods differ in prompts, post-processing, and selection, it is unclear what an observed gain should be attributed to. We study this question on DisasterBench, where models compose typed multi-tool workflows, by decomposing inference into three observable stages: whether a correct workflow is discovered, whether it survives structural validation, and whether it is finally selected. Comparing independent sampling, a nominal tree-search configuration, and multi-operator search under controlled inference budgets on two model families, we find that (i) the nominal tree-search configuration realizes non-adaptive complete-workflow sampling, and independent sampling with its expansion prompt reproduces its candidate discovery without constructing a tree; (ii) structural validation never removes the last correct workflow in our evaluated candidate pools, whereas majority voting leaves a discovered correct workflow unselected on 10-19% of tasks per pool, with an audit confirming substantive tool-choice or information-flow errors in 57 of 60 examined cases; and (iii) a tool-specification-aware verifier recovers some of these losses but overturns more correct decisions than it fixes. Our results motivate auditing realized rather than nominal inference pipelines, and reporting discovery, validity, and selection separately.

Although self-improvement is becoming increasingly important as tasks become more complex, models struggle to generate reliable supervision for harder, out-of-distribution instances. One way to address this challenge is to exploit compositional structure: models often solve simpler subproblems more reliably than full problems, so their subproblem predictions can be composed into supervision for harder ones. We systematically study when this kind of composed supervision supports self-improvement and when it fails. Our controlled experiments identify three failure modes: unreliable subproblem predictions, errors introduced by composition, and limited supervision coverage. Beyond controlled settings, these findings carry over to more realistic coding and reasoning tasks: composed labels are more accurate than direct self-generated labels and improve performance even on compositions longer than those seen during training, while the failure modes remain visible. Further, although these failure modes often keep compositional self-improvement from fully succeeding on its own, a short SFT warmup with compositional self-improvement can bootstrap reinforcement learning on tasks that require composing skills the model already has, particularly when direct exploration rarely succeeds. Together, these findings characterize when compositional self-improvement can provide sufficiently reliable supervision to expand model capabilities.

Improving situational awareness within the Smart Grid is an effective preventive measure to detect potential anomalies, and avoid their adverse consequences. A wide area measurement system (WAMS) based Phasor Measurement Unit (PMU) can assist in improving visibility within the grid by collecting relevant measurements including magnitude, phase angle, frequency, voltage, and current, with a precise GPS-based timestamp. So far, there are more than 2500 PMUs deployed in the North American power grid, where each of which generates between 30 and 60 samples/s, roughly 1.5 TB/month of streamed data. Thus, processing such large PMU streams requires expensive computational resources and faster algorithms. This paper provides state-of-the-art cyber and physical anomalies in the WAMS and analyses their impact on the integrity and availability of the system. Next, it investigates the current anomaly detection approach in historical and streaming PMU data, including conventional machine learning and hybrid approaches, and assesses their advantages and main drawbacks. Lastly, it sheds light on the importance of handling missing PMU data and concept drifts.

Bubeck 和 Sellke 的鲁棒性普适定律指出，当 $\mathbb{R}^d$ 上的协变量分布满足等周性质时，用具有 $p$ 个实参数的函数类在噪声水平以下拟合 $n$ 个带噪样本，会迫使 Lipschitz 常数达到 $\sqrt{nd/p}$ 的量级：光滑插值需要付出 $d$ 倍的过参数化代价。其假设为高斯型集中、齐次协变量分布、标量输出和独立样本；我们逐一放宽这些假设。指数集中已足够，速率和绝对常数不变，代价是容量预热（capacity burn-in），我们证明某种形式的预热是必要的；由此得到归一化对数凹协变量的定律，相比当前的 KLS 界有 $\log d$ 的损失。对于异质混合，在混合组成的典型性条件下，最坏分量常数被替换为各分量常数的加权平均。对于向量值预测器，Frobenius Lipschitz 尺度下的界不依赖于输出维度 $K$，而 Euclidean 界仅损失 $K^{-1/2}$，这一损失可被与协变量无关的噪声消除。平稳马尔可夫协变量保持 i.i.d. 速率，伪谱隙仅进入样本量条件。一个主定理承载了共同的论证，且全文常数均为显式。

Small LoRA adapters trained on Tajik news and books lowered Qwen3's Tajik text loss at every size we tried, from 0.6B to 8B, by 0.60 to 1.45 nats per token, but none of them improved Tajik reading comprehension where it can be measured. On Belebele the 1.7B and 8B adapters left accuracy unchanged and the 4B adapter lowered it by 3.7 points; at 0.6B the score rose by 4.2 points but sits near chance and coincides with a change in which answer letter the model prefers, so we do not read it as understanding. At 8B, continued pretraining on about 375 million tokens of raw Tajik text was followed by a Belebele gain of 9.9 points. After a further supervised fine-tuning stage, our main model, e3, is 11.4 points above Qwen3-8B on Belebele; its English MMLU changed by -1.2 points, and a loss larger than 3.3 points is excluded. The same heavy recipe applied to Qwen3-0.6B-Base brought its text loss close to that of the 8B model after books-only pretraining, while its multiple-choice answers stayed near chance and its English MMLU fell by 7.6 points. We report 15 trained models and their 5 starting models from June to October 2025, scored in 2026 on one harness with paired intervals, and we report what went wrong: lost training records, a thinking-mode dataset with almost no reasoning in it, later branches whose training data was never documented, and a 4-bit release that contains no transformer layers.

大型基础模型往往在可用的视觉或文本证据不足以支撑可靠结论时仍会给出回答。这种无法拒答的问题使其在高风险或信息检索场景中难以被信任，因为这些场景下一个自信却缺乏支持的答案可能比不回答更糟。这也引出一个自然的问题：这种行为背后的内部机制能否在推理时被识别并控制？我们提出 Dr. No，一种特征引导方法，它挖掘与拒答和作答相关的稀疏自编码器（SAE）特征，然后在 prefill 阶段对其中一小部分特征进行调制，而无需重新训练基础模型。Dr. No 在 LLaVA-1.6 Mistral 7B 上跨多模态拒答基准进行评估，包括 UNK-VQA 和 MM-UPD，并额外使用 Qwen3.5-2B 在 AbstentionBench 子集上对语言推理任务进行基准测试。与多种基于不确定性的拒答基线相比，特征引导在正确回答与适当拒答之间取得了比事后拒绝方法更有利的平衡。这些提升在 UNK-VQA 以及涉及图像—问题或问题—答案不兼容的 MM-UPD 案例上最为显著，并且最清晰地迁移到 AbstentionBench 的数学推理子集。然而，在缺失答案检测和检索密集型多跳问答（如 MuSiQue）上性能较为有限。这些结果表明，稀疏内部特征提供了轻量化的控制手段，用于塑造

背景：从树皮照片自动识别树种对林业资源调查和木材加工行业具有实际意义，但决定深度学习方法准确性的设计选择尚未被系统研究，且部分既有结果可能因同一树干图像在训练集与测试集之间泄漏而被高估。方法：在公开的 BarkNet 1.0 数据集（20 个树种）上，采用已确立的树级、防泄漏划分方案，开展系统性消融研究。使用 ConvNeXt-V2 骨干网络，依次改变以下因素并通过五折交叉验证进行评估：图像块大小、将各图像块预测聚合为图像级决策的策略、训练方式（单阶段与两阶段）、骨干网络容量、整图基线，以及推理时可用的图像块数量。结果：单块准确率随图像块尺寸单调上升（从 160 px 的 75.6% 升至 512 px 的 89.6%），图像级准确率也随之提升，在 512 px 附近趋于平台，约为 94.9%。学习式聚合最多比多数投票高出 0.7 个百分点，且随着图像块增大、注意力趋于均匀，这一优势会缩小。骨干网络容量（pico 至 tiny）对准确率的影响不到一个百分点，两阶段训练相较单阶段没有优势。整图分类（91.9%）落后于基于图像块的聚合，因为它没有可供平均的“包”；准确率随

本文聚焦于分析全连接前馈 ReLU 神经网络在逼近给定光滑函数时的表现。与通常研究的在架构规模（如网络宽度或深度）不断增加下的通用逼近性质不同，我们关注的是逼近网络参数的渐近增长。这类结果对于误差分析或神经网络训练的一致性结果等具有重要意义。我们工作的主要结果是：对于具备当前最优逼近误差的 ReLU 架构，其实现参数至多以多项式速度增长。我们将所得到的关于归一化网络规模的增长率与现有结果进行比较，并证明在大多数情况下（尤其是在高维输入时）该结果更优。

AI Agent 正逐渐成为数字同事，能够浏览网页、编写代码、运行实验，并在较长的时间跨度内协助人类。然而，现有基准测试尚不足以反映真实的知识工作：许多基准依赖合成数据或孤立任务，忽略了协作项目中复杂的多方动态——在协作项目里，关键信息分散于对话、代码和不断演进的决策之中。即使是一个简单的工作场所问题，也可能需要定位相关频道和私信、从冗长讨论中还原上下文、检查仓库状态、调和过时或相互冲突的证据，并跨来源进行综合。我们推出 SlackBench，这是一个用于评估 Agent 在真实科研与工程工作流中表现的基准，这些工作流要求对工作场所沟通和不断演进的仓库状态进行联合推理。SlackBench 以隐私保护的、对真实学术研究项目的重演为基础，提供可复现的环境，包含真实的多方对话、不断演进的决策，以及以 GitHub 仓库和 Google Docs 为依据的相应代码变更。它包含 140 个查询，涵盖异构来源的 Agent 式检索、对无法回答或前提错误的问题进行弃答、目标导向的摘要生成，以及长对话中的情境推断。这些任务无法简化为静态语料问答：Agent 必须判断哪些对话重要、遍历讨论串、检查仓库状态，并在各来源之间综合证据。

Large Language Models (LLMs) have become an indispensable part of natural language processing tasks. However, autoregressive sampling has become an efficiency bottleneck. Multi-Draft Speculative Decoding (MDSD) is a recent approach where, when generating each token, a small draft model generates multiple drafts, and the target LLM verifies them in parallel, ensuring that the final output conforms to the target model distribution. The two main design choices in MDSD are the draft sampling method and the verification algorithm. For a fixed draft sampling method, the optimal acceptance rate is a solution to an optimal transport problem, but the complexity of this problem makes it difficult to solve for the optimal acceptance rate and measure the gap between existing verification algorithms and the theoretical upper bound. This paper discusses the dual of the optimal transport problem, providing a way to efficiently compute the optimal acceptance rate. For the first time, we measure the theoretical upper bound of MDSD efficiency for vocabulary sizes in the thousands and quantify the gap between existing verification algorithms and this bound. We also compare different draft sampling methods based on their optimal acceptance rates. Our results show that the draft sampling method strongly influences the optimal acceptance rate, with sampling without replacement outperforming sampling with replacement. Additionally, existing verification algorithms do not reach the theoretical upper bound for both without replacement and with replacement sampling. Our findings suggest that carefully designed draft sampling methods can potentially improve the optimal acceptance rate and enable the development of verification algorithms that closely match the theoretical upper bound.

Two Boolean functions are considered NPN equivalent if one can be transformed into the other by negating inputs, permuting inputs, or negating the output. Efficient NPN classification is valuable in fields such as logic synthesis, quantum computing, and information theory. However, exact NPN classification demands exhaustive enumeration of all possible input permutations and negations. Its complexity grows exponentially, making it impractical for larger functions with more inputs. In this paper, we introduce the first learning-based formulation for NPN classification by recasting it as NPN action prediction. Our method represents Boolean functions as AND-Inverter Graphs (AIGs) and applies a bidirectional heterogeneous graph neural network to predict input negations, output negation, and input permutations. To mitigate the non-canonicity of logic networks, we further introduce graph rewriting techniques that reduce structural ambiguity and recover useful symmetries before prediction. Experimental results show that NPN action prediction is learnable, that the resulting canonicalization quality improves over heuristic baselines, and that on practical benchmarks our method approaches exact methods on tractable cases while scaling to functions with up to 16 inputs.

超参数优化（HPO）只返回单一最优配置，系统性地丢弃了所探索的性能景观——而这一景观编码了结果如何依赖配置选择。适应度景观分析源自进化计算，包含一组稳健的方法，可提供崎岖度指数、自相关函数和适应度—距离相关性，但由于此前不存在细粒度真值枚举数据，这些方法从未被大规模应用于机器学习超参数空间。我们提出 ExaTune，一个开源 HPC 系统，用于产出一种新的工件类别：超参数景观地图。ExaTune 通过 $K$ 折交叉验证穷举评估离散化超参数网格上的每一个点，将结果聚合为可复现的 parquet 基准数据，并应用一套包含五个形式化景观指标的方法：空间自相关（Moran's $I$）、崎岖度、局部最优密度、适应度—距离相关性（FDC）以及基于方差参数重要性。我们发布了一个基准集合，包含 8 张完整景观地图（共 285,080 个被穷举评估的配置），覆盖五个模型家族（SVM、随机森林、XGBoost、岭回归和 $k$ 近邻），横跨六个分类基准和两个回归基准，托管于 HuggingFace。实证分析表明，在标准试验预算下，贝叶斯优化、Optuna、Hyperband 和随机搜索会系统性地遗漏不稳定区域、secondar

Automated detection of machine-generated text is applied where the verdict carries a cost for the author, yet origin is a property of the production process, not an observable property of the text. The conditions under which a text is produced push it in the same direction in which a generative model is presumed to push it. This study measures the contribution of these conditions while origin is held fixed. The Russian-language corpus of 1882 documents comprises a machine part, generated by four generation channels (each backed by its own model family) across 45 tasks in three task modes, and an archival human part annotated for source regimentation level before the texts were read. The primary contrast is computed within the machine corpus: both halves of a pair are generated by the same model for the same task and differ only in task wording. Machine-likeness is measured by four operationalisations: an interpretable index with weights fixed before data collection, a trained classifier, a zero-shot metric on an open model, a judge model. The main result is disagreement between procedures: changing the task mode shifted the scores of all four, but three gave a negative shift, the judge a positive one, and none of the four intervals covers zero. The comparison plan and the specifications of procedures 2-4 were fixed after the result of procedure 1, before procedures 2-4 ran; the disagreement survived correction of a corpus defect and of the nested validation. A baseline on the source identifier alone reaches a median AUROC of 1.0000 on the thirteen group splits where it is defined, and removing emphasis markers flipped a confident machine verdict to human without changing the full text by a single character. The scores are properties of a specific measurement system and corpus: they reflect data collection, repetition, formatting, and preprocessing decisions as well as linguistic style.

AI "humanizers" rewrite machine-generated text so that detectors classify it as human-written. Public leaderboards score evasion and embedding similarity, but not whether checkable details survive. We check whether numbers, dates, person and organization names, and quoted spans survive each rewrite, using a rule-based checker validated by two AI annotators (precision 0.84-1.00, recall 0.93-0.95; no human annotation). In the October 2026 cycle of a public humanizer leaderboard, all 11 commercial tools changed or dropped at least one of these units in 35-65% of the documents that contained one. Most changes were to quoted spans; counting numbers and names alone, 10 of the 11 tools still altered at least one in 20% or more of the documents. The September 2026 cycle was similar (19-50%). An open 12B humanizer kept at least 97% of numbers and names in fact-dense U.S. federal agency text, but reworded 67-78% of the quoted spans, against its own prompt. Unit retention and detector evasion were not clearly related across tools (Spearman rho = -0.46, 95% CI -0.65 to 0.08), and the open model's four released files (two of them quantization-aware low-bit files) did not show the pre-registered monotone trade-off. On federal texts with matched human, AI-styled and humanized versions, a general-purpose decision model, Jev, asked a single zero-shot question, separated human originals from humanized rewrites with AUROC 0.864 (0.827-0.902), 0.18-0.22 higher than Binoculars, Fast-DetectGPT and the adversarially trained RADAR, while calling 24% of the human originals AI at a 0.5 threshold. On the AI-styled rewrites by 2026 models, both perplexity-based detectors scored below chance (AUROC 0.37 and 0.39), also with line breaks removed. Hypotheses, data rules and analyses were frozen before the main runs; two earlier data contacts are disclosed. We release the checker and the per-item results.

近年来，动态驾驶场景重建借助生成先验取得了显著进展，尤其是在新轨迹渲染质量方面。然而，生成数据与真实世界传感器观测之间仍存在显著差距，严重阻碍了重建保真度的进一步提升。本文提出 ReconDreamer#，一个面向动态驾驶场景、协同重建与生成的自演化框架，大幅缩小了合成新轨迹数据与记录的真实世界观测之间的差距。在该自演化闭环框架中，生成模型与重建模型通过迭代优化相互增强，最终显著提升重建模型的三维表示能力。此外，我们引入基于几何先验的自适应地面模型，以进一步增强重建模型的渲染质量。为验证所提方法的有效性，我们在 Waymo、nuScenes、PandaSet 和 EUVS 基准上进行了全面评估。具体而言，在 Waymo 上，当横向偏移为 6 米时，ReconDreamer# 相较 ReconDreamer++ 将 NTA-IoU 和 NTL-IoU 分别提升 8.8% 和 3.0%，并将 FID 降低 15.1%。

Search can help LLM agents correct mistakes, but their starting conditions may shape which evidence they find and use. Drawing on cognitive-bias paradigms, we develop a controlled search harness to examine result order, records of prior work, and initial hypotheses while holding each question and document collection fixed. Matched controls distinguish order effects from differences in document composition and the effects of failed work from context length. Across 10 models and 1,000 synthetic search worlds, we track queries, retrieved documents, document openings, and final answers. Our comparisons suggest that final answers are more sensitive to initial evidence order during active search than when all evidence is supplied in full. Dead-end histories can reduce strategy switching relative to matched neutral histories, indicating an effect beyond context length. Under wrong hypotheses, stage measurements distinguish reduced retrieval or opening from errors that persist after counter-evidence is opened. Exploratory interventions show that opening counter-evidence more often can coincide with lower accuracy, while accuracy can improve with little change in opening.

Reasoning substantially improves language models' ability to solve difficult problems, but long reasoning traces make these gains expensive. Efficient reasoning should express useful analysis in fewer tokens by reducing repetition and text that adds little value to solving the problem. Progress in proprietary models motivates this goal, yet their hidden scratchpads limit direct study of how it is achieved. We investigate whether open-weight models can generate less verbose reasoning by changing the form of their own earlier thoughts, without retraining. Our intervention rewrites prior reasoning into compact state records before pending tool results are revealed, then compares continuations from original and compact histories. Across the tested models, accepted compact histories usually shorten future reasoning; a coding case study also shows less repeated verification after a shared candidate solution. In software-agent deployments, compaction can reduce total spending, although task quality and rewrite-inclusive time depend on the target and compressor. These findings identify reasoning history as one way to reduce future reasoning-token use and motivate model-specific compaction policies evaluated for both efficiency and retained capability.

The rapid development of generative models has made AI-generated images increasingly close to real images. It has also brought several challenges to detectors, including rapid generator updates, large changes in data distribution, and limited cross-dataset generalization. Many existing methods based on pretrained vision encoders primarily use the global representation from the last layer, potentially leaving useful information from intermediate layers underexplored. To address this issue, we propose Global Inter-Layer Aggregation (GILA). GILA extracts global representations from different depths of a frozen encoder, models interactions among layer representations through inter-layer self-attention, and adaptively aggregates the resulting contextualized representations. The method is not tied to a single pretraining objective and does not require fine-tuning the pretrained encoder. We evaluated GILA on three vision backbones with different pretraining objectives, ten seen generators, and three external datasets. The experimental results show that GILA makes better use of multi-layer information in pretrained encoders and improves detection performance and cross-dataset generalization.

As AI systems are increasingly deployed in high-stakes domains such as healthcare, effectively integrating human expertise with AI outputs is critical for reliable decision-making. A substantial body of theoretical work studies optimal mechanisms for human-AI collaboration (HAC), focusing on when to involve humans and how to combine their expertise with AI predictions. However, these approaches largely overlook a key factor highlighted by empirical evidence: human judgments are impressionable and can vary depending on how AI outputs are presented. For example, when AI systems defer decisions to humans, revealing or withholding the AI prediction can significantly influence human judgments. Moreover, existing work neglects deployment-time variability, where conditions, particularly the cost of eliciting human input, may differ substantially from those observed during training. To address these challenges, we introduce a rigorous and theoretically grounded framework for human-AI collaboration, termed collaborative omnimechanism design, based on a novel and computationally efficient contextual bandit variant of boosting for outcome indistinguishability. In particular, we reduce a calibration-type outcome indistinguishability condition to weighted calibration, providing a novel time-efficiency result, bypassing the high-dimension curse induced by complicated collaboration methods and outcome space. Our framework provides a principled characterization of human impressionability and introduces a unified optimization scheme that enables robust and adaptable decision-making across diverse query costs and evaluation objectives without retraining. We evaluate the framework in three HAC settings, examining adaptation to query costs and prediction objectives without retraining, and show that its induced policies perform comparably to or better than the evaluated baselines.

On-policy distillation (OPD), which uses reverse-KL supervision on student rollouts to transfer a context-conditioned model's behavior into student model parameters for closed-book inference, has shown strong results. However, we find that this success does not translate to knowledge injection: across two document-QA benchmarks, OPD injects document knowledge far less efficiently than simple off-policy distillation. We trace this reversal to two failures that are structural, yet sparse across token positions. Reverse KL weights its gradient by the student distribution and therefore barely lifts document-specific tokens the student considers unlikely; after forward KL restores this signal, student rollouts expose later positions on the target trajectory only after generating the preceding tokens correctly, so coverage deteriorates with target length. Motivated by the concentration of teacher-student divergence at these hard positions, we propose Adaptive-Policy Distillation (APD), whose per-token gate jointly shifts the rollout toward the teacher and the objective toward forward KL where needed, withdrawing both interventions as the student learns. Without labels or a verifier, APD improves over OPD by up to 54.6%, achieving the best scores of 0.558 F1 on QASPER and 0.486 MRR on BioASQ.

Offline multi-task reinforcement learning (MTRL) aims to learn a unified policy from heterogeneous offline datasets that generalizes across tasks. However, existing offline MTRL methods often struggle with two fundamental issues: severe per-task data scarcity and limited cross-task generalization. In this paper, we propose Language-Model Augmented Decision Transformer (LaMA-DT), a unified sequence-modeling framework that leverages a single large language model (LLM) for both trajectory augmentation and multi-task policy learning. LaMA-DT uses natural-language task descriptions to guide task-aligned trajectory augmentation through a set of supervised generation and alignment objectives, and incorporates a scoring-and-filtering mechanism to retain high-quality augmented trajectories. The same LLM is then repurposed as a language-conditioned Decision Transformer and fine-tuned on a mixture of real and augmented data. Experiments on Meta-World and MuJoCo multi-task benchmarks demonstrate that LaMA-DT consistently outperforms competitive baselines, particularly in low-data regimes, showing that language-guided data augmentation substantially improves cross-task generalization.

Cognitive Behavioral Therapy (CBT) demands more than empathic validation, requiring instead a structured formulation of the interplay among situations, automatic thoughts, emotions, and behaviors. Therapists formalize this interplay through a Cognitive Conceptualization Diagram, where maintaining longitudinal coherence necessitates continuous evidential adjudication to separate tentative hypotheses from established findings. Current dialogue systems track conversational history via retrieval or static profiles but lack the representational capacity to encode epistemic status. Consequently, provisional interpretations are often prematurely reified as established facts, inadvertently biasing downstream interventions. Addressing this limitation, CCDWiki is introduced as a plug-and-play advisor conceptualizing longitudinal therapy as evidence-grounded hypothesis adjudication. The approach encodes each clinical hypothesis as a mutable entry bound to corroborating and contradicting evidence. Across sessions, accumulating data strengthens supported entries, preserves provisional ones, and retracts contradicted assumptions. The resulting formulation distills into concise guidance for unmodified therapeutic language models. Experiments demonstrate substantial improvements in both conceptualization and intervention quality without requiring backbone retraining.

Computational brain imaging is constrained by a mismatch between increasingly diverse tasks and scarce, heterogeneous clinical data. Although existing brain foundation models transfer pretrained representations, their knowledge remains entangled within model-specific features and fixed architectures. This paper proposes BrainGFM, a generative foundation model that extends knowledge transfer beyond representation reuse through a reusable and reconfigurable brain-imaging prior. By pretraining on a corpus of about 1.3 million MRI and CT slices, BrainGFM supported task-aligned data synthesis for low-data predictive learning and task-adapted post-training for image reconstruction. Across 48 downstream tasks, BrainGFM improved segmentation and classification under limited supervision and achieved high-fidelity image enhancement across heterogeneous imaging conditions. Multicenter clinical validation involving 4,796 stroke patients across two internal and four external centers demonstrated its applicability to three clinically relevant tasks around stroke cohorts. Blinded reader assessments showed that BrainGFM-generated FLAIR images approached acquired FLAIR in overall diagnostic quality. In a human-AI collaboration study, BrainGFM assistance yielded a 14% relative improvement in mean radiologist stroke-subtyping accuracy. These findings support generative priors as a reusable foundation for heterogeneous computational brain-imaging applications.

具有时间依赖系数的演化方程迫使神经求解器将容量花费在方程已经确定的尺度上。我们表明，一个经典对象——Ermakov-Pinney 尺度函数 ρ——可以作为无参数嵌入写入物理信息神经网络的输入层。通过 Pinney 常数 c 的两个符号，一种构造覆盖三类问题：线性扩散（c=-1，ω≡0）、时间依赖约束势阱中的扩散（c=-1，ω≠0），以及具有时间依赖频率的量子振子（c=+1）。一个精确的矩恒等式仅由初始数据即可给出 ρ 的闭式形式，因此不需要辅助网络或解的已知信息。我们的核心发现是比较性的：每一类问题需要该构造的不同组成部分，并且有各自不同的可靠对照。在族 I 中，仅扩散时间重参数化就再现了大部分增益，因此该族对该思想而言是一个糟糕的检验，并说明了为什么必须进行消融。在族 II 中不存在时间重参数化，对照会降低精度，而振幅前因子变得至关重要（40.8 倍）。在族 III 中绝热近似也失效，重要的是二次相位因子（5.7 倍）。当约化解不收敛到不动点时，该方法没有帮助：其外推优势随之消失。约化本身是经典缩放变量变换，即 sel 的

Automated theorem generation (ATG) seeks to construct new theorems rather than merely prove given conjectures. Existing approaches typically generate candidate statements that require subsequent proof or formal verification. We propose a constructive framework for ATG in first-order logic based on triangular standard contradictions. The method constructs minimally unsatisfiable clause sets from selected first-order literals, then forms theorems by retaining some clauses as premises and taking the negation of one clause, or a conjunction of clauses, as the conclusion. Minimal unsatisfiability ensures that the remaining premises are satisfiable, avoiding trivial inference by explosion, while unsatisfiability of the full set guarantees the resulting entailment. Theorems are therefore correct by construction, without a separate proof procedure for each instance. Extending the previously developed $\Delta_1$ generator, we introduce seven methods, $\Delta_2$--$\Delta_8$, forming a unified family of eight theorem-generation methods. Four generators, $\Delta_1$--$\Delta_4$, have been implemented and released. We also formulate a general $\Delta$-class method encompassing the family. Combinatorial analysis demonstrates the capacity to generate large numbers of distinct theorems systematically. This framework offers a principled foundation for automated theorem discovery and the construction of logically guaranteed knowledge.

Confidence estimates guide human review in LLM-built knowledge graphs, but better error prediction can yield little improvement in an audit. We study this gap through repairability, selection headroom and label validity. Modelling non-adaptive auditing as budgeted covering, we characterise the minimal decision-relevant error statistics and bound the regret of scalar-preserving first-order planning by a structural mass $\Delta^{\mathrm{cl}}$. We simulate ideal repair against benchmark-aligned labels for five extraction configurations and two cross-document graphs on Re-DocRED, followed by frozen confirmation on test and replication on 500 hash-selected training documents. On the held-out cross-document graphs, an error-rate $\times$ fan-out queue leaves 3.98-4.74 percentage points more errors than a joint model with the same actions. By comparison, opening shared actions for that joint planner reduces residual errors by 0.89-1.43 points relative to its own fact-only plans, in point estimates. Equivalence between joint and independent-block models remains inconclusive. Generic LLM verification improves three of four model-budget configurations on each held-out set. LLM adjudication identifies possible benchmark omissions.

Dhaka's rapid urban growth has made building setback violations common, narrowing footpaths and blocking emergency access. RAJUK, responsible for building control in the Dhaka Metropolitan Area, enforces setback rules mainly through manual field inspection, which scales poorly. We present a systematic comparison of four machine learning approaches for classifying building setback conditions from street-level photographs: negative (compliant), moderate (borderline), and positive (violation). We built a labeled dataset of 372 geo-tagged images from sixteen Dhaka neighborhoods and expanded it with 5,968 augmented copies for training. We trained and evaluated a multi-layer perceptron (MLP), a Random Forest ensemble on frozen ResNet50 embeddings, a support vector machine on frozen ResNet18 embeddings, and a fine-tuned ResNet18 CNN. On a 1,275-image test set, the fine-tuned CNN achieved 65.4% accuracy (macro-F1 0.50), followed by Random Forest at 59.9% (macro-F1 0.44) and SVM at 58.2% (macro-F1 0.45). All models classify the compliant class reliably but struggle with the borderline moderate class due to visual ambiguity. These results establish reference performance levels for automated setback-compliance screening in a South Asian regulatory setting.

科学家在拟合任何参数之前，必须先决定模型包含什么：哪些量存在、哪些是隐藏的、它们如何相互作用，以及如何被测量。物理学家把这种试验形式称为 ansatz。自动化方程发现基准大多会提供它：固定变量、给出候选项库，或给出一个待修改的定律。我们提出 Unsatz，这是一个 LLM agent 必须自行构造 ansatz 的基准。每个任务都是一个模拟系统，由私有种子生成，表示为一种小型建模语言中的程序：常微分方程、延迟微分方程或偏微分方程，离散映射，静态模型和因果模型，或随机过程，可能带有隐藏状态、失真的传感器和无信息测量通道。agent 在预算内购买实验，用同一种语言编写候选程序，从基准的拟合器获得其系数，并且仅根据其最终程序对从未见过的实验的预测效果评分：超出其可探索范围的条件、干预以及更长时间的运行。在 20 个领域的 80 个系统中，九种语言模型的解决率为 21% 到 60%。在这 44 个系统上，经典方程发现方法几乎从不成功，尽管一旦给出真实方程（系数未知），同样的数据通常就足够；难点在于找到结构，而不在于估计结构。由于机制会在不同系统间重复出现，我们还把 agent 为相关 sy 发现的程序提供给它们

Memory capacity and bandwidth are major bottlenecks in LLM inference, motivating weight compression beyond conventional 4-bit quantization. Vector quantization (VQ) enables approximately 2-bit weight storage but often relies on codebooks incompatible with native FP4 computation. We introduce MXCode (MX-Native Codebook Quantization), a framework bridging ultra-low-bit VQ compression with MicroXcaling (MX)-native FP4 computation. MXCode proposes three strategies: PostScale, which quantizes learned codebooks to MXFP4; PreScale, which extracts MX scales before codebook learning; and PreScale+, which further compresses these scales using VQ. These strategies explore tradeoffs among quantization quality, scale overhead, and codebook granularity. All reconstruct MX-native W4A4 operands without materializing intermediate high-precision weights. We evaluate MXCode on Qwen3 (235B), DeepSeek-R1 (671B), and GLM-5.2 (753B) using the Microsoft MX Emulation Library. MXCode achieves approximately 2 bits per weight with competitive, workload-dependent model quality. On GB300 hardware, experiments with Qwen3 (235B) show approximately 50% lower model-load memory than Native FP4 while maintaining near-Native-FP4 serving throughput. Under memory-constrained prefix reuse, P90 TTFT improves by up to 94.5%. These results demonstrate software-based reconstruction on existing FP4 hardware and inform the development of future MX-native inference systems.

Audio language models are meant to answer questions from the audio itself, but an hour-long recording rarely fits in their context, and the answer often depends on a few seconds of it. Transcribing the recording first makes it fit but discards sounds and voices, while systems that return to the audio must be trained for this behavior. We propose ReListen, a training-free method in which a frozen audio language model finds its own evidence before answering. In a Jev-style scan, the model scores each 20-second window by the probability that it helps answer the question, and it then answers from the highest-scoring minute. This minute overlaps the annotated evidence far more often than transcript retrieval, for 88.5% of VoiceGiraffe questions. Without any change, the same scan also grounds events on LAT-Bench as well as a model trained for temporal grounding. Answering from it matches or exceeds hearing the whole recording with a small fraction of the audio, and on VoiceGiraffe it exceeds a transcript of the whole recording by 9.2 points. The selected evidence is also worth hearing rather than reading. Given the same excerpts, reading their transcript matches listening on speech content, but listening is clearly more accurate on questions about sounds and voices. Moreover, ReListen answers faster than every transcript-based system, and its memory stays constant even for recordings too long to hear whole on one GPU. Therefore, a frozen audio language model can search hour-long recordings by itself, without a transcript or additional training.

模型上下文协议（MCP）通过实现消息验证与会话逻辑的官方 SDK，将大语言模型智能体连接到外部工具和服务。这些 SDK 所遗漏的规范约束可能会传播到下游应用，凸显了 SDK 层级一致性测试的重要性。然而，现有的 MCP 一致性工具主要针对运行中的端点进行测试，几乎无法揭示 SDK 模型是否强制执行规范要求。我们提出 MCPGuard，一个基于 LLM 的自动化框架，用于检测规范到 SDK 的一致性差距。其结构轨将 JSON Schema 约束与规范化后的 SDK 模型对齐，并生成针对具体约束的证据。其行为轨提取规范性文字规则，检索有界的函数级代码切片，并借助 LLM 审计其中可能存在的遗漏。随后，一个证据感知的动态裁决器检查完整的源码上下文，并构建 PoC 以验证候选问题。共享内核配合特定语言的适配器，使分析能够跨 SDK 进行。在官方 Python 1.26.0 和 TypeScript 1.29.0 SDK 上，MCPGuard 分别确认了 13 项和 19 项发现，而两个黑盒基线均未发现其中任何一项。对开源 MCP 服务器的测试进一步表明，这些差距会系统性地向下游传播。通过使规范约束在 SDK 层面的一致性可直接测试，MCPGuard 补充了传统的面向端点的一致性测试套件。

Causal discovery is usually evaluated on simulators or on systems whose ground truth follows from physics or design. When it fails on other real data, the failure is attributed to assumptions that may not hold, but which ones fail, and by how much, is rarely measured. We measure the observable properties of the data they bear on, one property at a time, on a real panel. The case study follows 2,901 US public firms over 31 quarters, with 71 types of corporate action from SEC 8-K filings, 11 financial-statement items, and an expert graph over 651 action-state pairs. The panel breaks measurable requirements of discovery: tails that survive winsorisation, balance-sheet items and revenue close to random walks, and interventions too rare to detect effects of a tenth of a standard deviation in 40 of 71 actions. Even so, a marginal test ranks expert-labelled direct above not-direct pairs within each action (AUROC 0.658) better than every null we built, so the panel does carry information about which pairs are direct. Tuned discovery methods, in configurations feasible at this panel size, recover less of it than a simple pairwise cross-correlation when pairs are ranked within each action, and no more when they are ranked within each target, which removes what the target alone predicts. Most direct pairs (132 of 222) involve actions below the support needed for 80% power against a 0.1-SD effect. For flow targets, conditioning on the target's other expert-labelled parents, often filed alongside the action, erodes the marginal discrimination, and most not-direct associations survive it: even with the graph's parent sets supplied, the expected separation does not emerge. For each property, we give a diagnostic computable before any discovery method is run.

多智能体语言模型系统常常加入审查者来改进上游智能体的决策。然而，审查者既可能修复错误，也可能替换原本已经正确的决策。我们将此视为制度性管辖权问题：在相同提案、任务信息、模型和审查输出条件下，审查者何时被授权使其建议具有约束力？我们比较宽泛的推翻权限与一种证据门控——仅当审查者给出关于违反约束且修复合规的机械有效记录后，才允许替换。在两个本地运行的开源权重模型、96 个空间策略任务以及一个 96 任务的投资组合迁移基准上，门控大幅减少对正确提案的破坏，同时保留大多数对违规的修复。这种权衡的方向部分由规则本身决定；其幅度取决于审查者的状态转换和上游提案状态。随后一项前瞻性验证让审查者明确选择保留、替换或上报。这消除了 Qwen 的正确提案被破坏问题，但 Mistral 仍替换全部 48 个正确提案；Qwen 也拒绝了 Mistral 经常尝试的所有仅涉及福利的修复。因此，接口设计会改变审查者行为，但不是与模型无关的保障。结果并不是受限审查普遍更好，而是审查者能力、行动接口和约束性管辖权共同决定纠正域……

Small students degrade on large reasoning models' (LRMs') long solutions, so recent work shortens them by interpolating an LRM with a short-answer model at ratio $\lambda$. We ask what student weights retain when this teacher stops thinking, comparing LoRA adapters of three students from three teacher pairs at each $\lambda$. In each pair, the adapters' update subspace shifts most where the teacher stops thinking and solutions shorten most, locating the change without attributing it. We identify a shared adapter subspace tied to thinking-like behavior (reflective markers, monologue format and long solutions). Shared by adapters before that point, it keeps about half (post hoc) while one pair's teacher halves its solutions but keeps its think tag, and largely empties after thinking stops. Removing it largely turns this behavior off for every pair, student and benchmark. Only in that pair and one other student does it leave clearly fewer markers than removing the update's top components; elsewhere, same-energy random directions often nearly match it. Grafted into a short-answer student at twice its weight, it turns early markers (a post hoc count) on, like the whole update, but the student rarely finishes. Our work identifies a weight subspace tied to distilled thinking-like behavior.

2D Gaussians render images at thousands of frames per second. However, fitting an image takes seconds of optimization. This confines 2D Gaussians to offline use. Feedforward methods are faster but less accurate without per-image fine-tuning. We present Gaussifier, a lightweight feedforward framework for 2D Gaussian decomposition. We learn a density map of Gaussian centers from per-image fitting data. Equal-mass Voronoi places Gaussians so that each Voronoi cell holds approximately equal density. A stateless, reusable network then refines the output in a few iterations. Gaussifier reaches 40.73/38.61\,dB PSNR on Kodak-24/DIV2K at over 60 FPS. At the same Gaussian count, it exceeds every feedforward baseline by 3.0 to 15.6\,dB in PSNR. With fewer Gaussians than any baseline, it still leads the strongest one by 2.5 to 3.5\,dB. At the same count, baselines with their own per-image fitting need 26 to 326 times longer to reach its PSNR. Gaussifier makes high-fidelity 2D Gaussian images fast and light enough for real-time applications. We release code, weights and a browser-based WebGPU demo at https://github.com/Gaussifier/gaussifier.

视觉语言模型（VLM）正日益用于支持施工危险监测，但其频繁误报仍缺乏充分理解。本研究探究错误的撞击危险警告究竟源于物体识别失败、空间关系编码失败，还是源于由学习到的共现先验主导的危险决策。我们使用 355 对受控空间邻近度和移除工人变体的反事实工人-挖掘机场景对，评估了 Qwen2.5-VL-7B 和 InternVL3.5-8B，并结合行为分析、内部表征探测、因果干预和决策校正。两个模型均能识别实体并编码其空间关系（折外 AUROC：0.82 和 0.91），并分别在 79% 和 73% 的场景对中给近距离场景分配了高于远距离场景的危险分数。然而，分别有 74% 和 100% 的远距离场景被错误分类为危险。移除工人使 Qwen 的危险判定率降至 11%，而在两个模型中，工人存在对危险分数的影响分别比空间邻近度大 2.4 倍和 3.2 倍。注意力头干预、神经元级干预和激活引导均未能消除这种共现偏差。然而，在匹配敏感度下，将危险决策条件化于模型自身的空间关系表征，使 Qwen 的特异度从 0.36 提升至 0.58，优于阈值重校准。这些发现表明，VLM 的撞击误报主要源于决策失败，而非……

统一多模态模型（UMMs）中的离散 token 化虽然建立了强大的语义推理能力，但也引入了严格的量化瓶颈，导致细粒度视觉细节退化，通常生成结构准确但伪影严重的图像。我们提出 FlowShift，一种参数高效的后训练框架，通过将预训练的离散 UMMs 转换为连续 rectified-flow 生成器来绕开这一瓶颈。FlowShift 不丢弃骨干网络的语义先验，而是用连续潜变量投影替换离散视觉接口，并对注意力层应用 LoRA。为了在不产生语义漂移的前提下高效提升视觉质量，我们提出双目标速度向量蒸馏（Dual-Objective Velocity Vector Distillation, DOVD），使模型与高保真连续教师模型以及真值连续轨迹对齐。该方法在显著提升感知质量的同时，保持了提示词遵循能力和推理能力。将 FlowShift 应用于领先的离散 UMMs 时，COCO-30K FID 和美学分数显著提升，同时保留超过 99.5% 的多模态理解能力。消融实验证实 DOVD 驱动了生成质量的提升。最终，FlowShift 在不牺牲预训练语义基础的情况下实现了高保真连续生成。

为什么世界上一些最强大的人工智能（AI）系统仍然难以对因果关系进行推理？因果发现，即识别系统潜在驱动因素的能力，是科学探究的基础，传统上依赖于结构化数据和统计方法。近来，大语言模型（LLMs）已被探索为因果推断的另一种范式，利用其广泛的知识和推理能力。然而，尽管 LLMs 在语言上很流畅，它们往往依赖表层模式识别，而非真正的因果逻辑。虽然已知 LLMs 在这一领域会出错，但我们缺乏对这些失败发生在哪里、为何发生以及严重程度的精确图谱。在本文中，我们提出了一个系统性压力测试框架，涵盖六个关键维度：模型规模、因果图复杂度、提示策略、元数据质量、数据质量以及对不确定性的敏感性。我们的发现揭示了 LLMs 中显著的“乐观偏差”；模型一贯高估因果联系，导致整体精确率偏低。此外，我们证明传统评估指标可能具有欺骗性：尽管推理密集型模型在性能上领先，最高分往往出现在推断中的不确定性被忽视时。一旦将不确定性纳入评估，性能就会下降，从而更诚实地反映因果可靠性。我们还发现，虽然扩大模型规模能改善 outco

基于多参数 MRI（mpMRI）准确预测前列腺影像报告和数据系统（PI-RADS）评分仍具挑战性，难度高于临床显著前列腺癌（csPCa）检测。鉴于病灶位置在 PI-RADS 评分中的重要性，我们利用大规模内部数据集，评估了在多种架构（3D ResNet-18、ResNet-50 和 ViT）中引入前列腺分区掩膜是否能改善 PI-RADS 预测和 csPCa 检测。解剖学指导总体上提升了各架构的性能，其中 ResNet-18 结合 2 区指导在 PI-RADS 预测中取得最佳结果（QWK = 0.511），ResNet-50 结合 5 区指导在 csPCa 检测中取得最佳结果（AUC = 0.809）。在外部数据集上的微调结果进一步证明了解剖学指导的益处。此外，采用解剖学指导的模型提升了模型预测 PI-RADS 评分在基于 PI-RADS 的 csPCa 检测中的实用性（阳性定义为 PI-RADS>=3）。结合模型预测与专家评定的 PI-RADS 评分，将 AUC 从 0.711 显著提高至 0.756（p < 0.05），表明 AI 推导的 PI-RADS 评分可能为专家评估提供补充信息以用于 csPCa 检测。总体而言，我们的研究结果凸显了将前列腺分区解剖信息纳入基于深度学习的前列腺 MRI 评估中的价值。

学生可以使用大型语言模型完成作业，但正确的解答并不能证明他们理解其背后的推理。个性化访谈可以探查这种理解，但需要工作人员反复投入精力。我们推出 ProofPaths，这是一个 AI Agent 框架，将基于评分标准的评分与个性化、交互式访谈相结合。多个评分 Agent 依据教师指导对提交内容进行评估，交换批评意见，并修订各自的判断。它们的分数决定由哪份评审提供成绩和反馈。访谈 Agent 要求学生论证其提交内容中的各个步骤，并利用学生的回答引导后续提问。每个问题都关联到一段检索到的提交内容摘录，因此教师可以将学生的解释与所引用的段落进行对比。系统会跟踪评估进度，并为每份提交内容存储评分评审和访谈记录。针对评估控制器的若干确定性测试会检查来源处理、分数汇总以及纠正尝试次数限制。教师可以查看所有评审和访谈记录，并在作业工作流中覆盖成绩。

Recommending creators that a user does not yet follow is a cold-start problem: from hundreds of thousands of popular creators and only a short trail of prior engagements, the system must predict the next creator worth surfacing. The prevailing generative recommendation recipe tokenizes each creator into a \emph{Semantic ID} (SID) by quantizing a frozen embedding and trains a language model to generate SIDs. We show that this SID layer, not the language model, is the binding constraint. The creator-to-SID map is many-to-one, so colliding creators must be resolved by a popularity tie-break that caps creator-level recall and favors already-popular creators, and the frozen quantizer stops the training gradient from reaching the creator representation. We propose Atomic Creator Token (ACT), which gives every creator its own vocabulary token trained end-to-end. The map is bijective, so collisions are impossible and inference is a direct top-$k$ over creator logits. In a controlled comparison on production data in which the token layer is the only variable, ACT improves creator-level Recall@15 by 19.8 points ($1.84\times$) over a matched SID baseline, still leads when the SID codebook is enlarged $64\times$, and serves at ${\sim}2.7\times$ lower per-user latency. The advantage carries over to the public KuaiRand dataset, where ACT is the only method that beats a global-popularity baseline, provided each creator token sees enough training examples. In a live A/B test, ACT is directionally positive on every reported metric against both the SID-enhanced and the classic production rankers.