AI 论文与研究

微软研究院发布关于物理AI机器人推理基础设施的研究。通过系统评估移动操作任务，发现将推理从机载GPU卸载至边缘或云端，能解决机载算力限制模型规模、导致响应延迟（最高383%）及精度下降的问题。实验显示，卸载推理不仅提升了任务成功率，还通过替换高功耗机载GPU为轻量硬件（如树莓派），使机器人电池寿命延长高达160%。此外，微软推出了基于Kubernetes的Physical AI Toolchain工具集，支持自动化容器化与分布式推理部署。

Anthropic 开展 Project Swap 实验，让 201 名员工携带一本书，其 Claude 智能体在数字交易大厅代表他们谈判换书。5 分钟聊天后，Claude 构建的图书排序与本人自评排序在 61% 的书对上一致；市场效率为 0.55（最优 0.89），其中 85% 的差距来自智能体对参与者偏好的理解不足，而非谈判表现。

Microsoft Research 等机构的论文研究多智能体在任务执行中通过共享目录通信的收益。在 ARC-AGI-3 上，k 个通信智能体的成功率匹敌 4k 个独立智能体且差距随 k 扩大；在 polyomino packing 超越 best@k 和已知最优成绩，MNIST 压缩任务中四智能体团队写出 1,957 字节、99.4% 测试准确率的分类器。

Stanford 和 Together AI 的论文提出 Self-Organizing Agent Teams（SAT），让 o3-mini、Claude Sonnet 4 和 DeepSeek-V3 组成的团队学习协作策略。

Stanford 和 Together AI 等机构的论文提出 Self-Organizing Agent Teams（SAT），让 o3-mini、Claude Sonnet 4 和 DeepSeek-V3 组成的团队从过往协作中学习可复用的协作策略。

Data assimilation estimates a dynamical state from partial and noisy observations. Classical ensemble filters are efficient but restrict analysis updates through finite sample covariance and affine Gaussian distribution. We introduce the F…

Machine-learned interatomic potentials (MLIPs) for solid-liquid interfaces in advanced materials applications, e.g., electrochemistry, catalysis and corrosion, require training data that samples both liquid environments, the solid and the …

Most Turkish-capable large language models (LLMs) are evaluated using general-purpose benchmarks rather than long, structurally complex domain documents. This paper evaluates five open-weight 7B-8B models for Turkish document question answ…

Decoder-only Transformer language models cache keys and values (KV) to reuse past computation during generation. Sharing KV across layers saves storage but reduces the diversity of representations available across depth. We study what loca…

Effective summarization of interrogative dialogues is a critical task in forensic and investigative settings, requiring high factual accuracy, coherence, and attribute-specific relevance. In this work, we introduce CASPER, a novel Chain-of…

Small and medium-sized language models offer cost-effective executors for tool-using agents, making them attractive for local and large-scale deployment. However, in long-horizon and stateful environments, they often make structural errors…

Financial institutions are beginning to deploy agentic workflows in credit, fraud, collections, compliance, and operational control. Governance remains largely component-centric: each model or agent is specified, tested, authorized, and mo…

Methods operating on Vision Transformer (ViT) feature spaces typically rely on Euclidean distance or cosine similarity. This assumes that every direction is equally meaningful, but there is no reason to believe the true task geometry has t…

Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask questions that uncover relevant patient information. To train such dialogue policies, a…

Accurate remaining discharge time (RDT) prediction is challenging in real-world battery applications because future load profiles are unknown and highly dynamic. To address the uncertainty of continuous RDT regression, this paper introduce…

In this paper, we are concerned with matrices formed by block-diagonal factors interleaved with fixed permutations -- a flexible family of structured matrices. This class has recently drawn interest in deep learning architectures for its b…

KV-cache eviction is typically evaluated through average quality-memory trade-offs, yet a small average loss can hide requests whose utility degrades materially. We reformulate eviction as a deployment risk-control problem: a material degr…

Pruning large pre-trained transformer-based ASR models such as OpenAI's Whisper has seen great adoption, as pruning the decoder led to significant end-to-end transcription speedups. For instance, the {\tt whisper-large-v3-turbo} variant re…

The surge in AI-assisted generation of educational materials has outpaced our capacity to validate their pedagogical quality. Automated evaluation using Bloom Classifier models is a promising approach to assess educational materials at sca…

Vision-language-action models (VLA) promise human-steerable autonomous driving, but their training is bottlenecked by the scarcity of frames paired with natural-language instructions: while camera streams and expert trajectories are logged…

This paper advocates category theory as a practical framework for structuring and improving rein- forcement learning in high-dimensional, partially observable environments. We model symmetries between environmental states by partitioning t…

In value-based reinforcement learning, improving the accuracy of policy evaluation has been shown to improve downstream policy optimization performance. The widely adopted family of approximations relying on $n$-step truncation yields comp…

We propose the Memory-Efficient Neural Operator (MENO) as a high-performance PDE neural solver based on the Manifold Function Encoder (MFE). MENO features three primary advantages: (1) MENO has a significantly smaller memory footprint and …

Newton-Schulz (NS) iteration has recently been used in the Muon optimizer to transform update matrices during the training of large language models. Motivated by its spectral effect, we investigate applying NS directly to Transformer atten…

Vision-Language-Action (VLA) policies often rely strongly on the camera viewpoints seen during training, causing substantial performance degradation when deployed from unseen perspectives. Collecting demonstrations from sufficiently divers…

百度智能云灵医大模型在医疗大模型评测平台 MedBench 更新榜单中以综合得分 76.1 位列第一，并在医学知识问答、医学语言生成（并列第一）、复杂医学推理、医学语言理解等细分维度领先。灵医以智能体 Harness 架构为基础，依托持久会话记忆、技能管理和多智能体协作等机制面向真实临床场景持续优化。相关产品已累计落地医院客户 800+ 家、基层医疗机构 4000+ 家。

OpenAI 发布开放基准 MentalHealthBench，评估 AI 在真实心理健康对话中的表现，由来自 22 个国家、19 种语言的 80 多位持证心理学家和精神科医生共同构建。

JevBench v1.4.1 发布，这是 Benchmark Heaven 面向 Jev 类决策模型的可重复基准，输入状态与有界评分标准、输出类型化答案，基于 534 个公开加 308 个密封决策、协议 jevbench::v1.4 评分。榜单共 77 个系统，Jev 1.13.0 以 77 分居首，JevK5 v0.2.0 与 Hopper 分列二、三。

论文《Look Light, Think Heavy》指出多模态模型存在"看轻想重"问题：看图越来越浅、语言推理越来越长，靠语言自迭代和解释脱离视觉本身。GPT-6 上这一现象越来越明显，在 image 2.5 的使用中造成明显干扰与脱节，作者认为用解释代替观看也是人类常犯的错误。

数学领域公开披露的AI使用率在不到6个月内从1.39%升至14.09%。 作者扫描了2026年3月1日至8月20日期间发布的32,944篇数学相关arXiv论文。 证明构建是最大的类别，出现在1,225篇论文中。 AI在数学各领域的渗透并不均衡：组合数学的AI辅助论文总数最多，而度量几何中自身论文使用AI的比例最高。

DeepSeek 发布 DSec（DeepSeek Elastic Compute）论文并署名梁文锋，公开用于 Agent 训练的沙盒基础设施技术细节。

斯坦福与牛津论文 MedRSI 显示，医疗智能体可从自身错误中自我改进，但新能力须先在后续新患者上验证才能固化。若每轮都立即注册新工具，第 30 轮准确率降至 76.9%（57 个工具）；放慢注册、仅保留 18 个工具时准确率达 94.4%。该智能体还会优先处理可能造成更大临床危害的错误，而非仅关注最常见错误。

感谢 @ValsAI 的高水准评测！"flash"这个词现在有点"误导性"了。Ling-3.0-flash-fin 总参数 124B、激活 5.1B，是一款高智能密度的"flash lite"。趁免费 API 还在，尽情享用吧。我们还有 fp4 量化版本可用于本地 AI 😛

Redwood Research 测量 GPT-6-Astra 在提示词中加入无意义填充 token 后的表现，发现其在要求大量串行认知的任务上显著提升，4-hop 自然事实推理从约 10-20% 提高到约 50%，旧 AIME 题从约 60-70% 提高到约 90%。

论文提出 Just-in-Time Memory（JitMem），保留原始轨迹并把记忆整理推迟到读取时，由整理器根据当前任务合成任务自适应的紧凑载荷。在 ALFWorld、WebShop 和 τ^2-bench 上，JitMem 分别超过最强基线 16.2、16.3 和 3.9 个绝对成功率点，未训练的整理器也已与基线相当或更优。

Apple 提出 probe guidance，利用现有扩散模型冻结的内部状态构建引导信号，无需在推理时增加额外前向计算，即可让弱模型与强模型保持相近动态。该方法在连续扩散语言模型的无条件生成上取得新的 SOTA，应用于 1.7B 扩散语言模型时在多项选择题基准上持续提升。研究还发现，传统 autoguidance 中的弱模型必须来自训练的低熵区域。

研究者提出 Agent-Editing World Model（AEWM），不再模拟工具响应，而是建模推理与动作如何塑造任务进展，通过 Action Judge 区分 Critical、Exploratory、Noisy 决策，并用 State Revision 编辑噪声推理-动作续写。

DeltaWAM 用稠密锚点、稀疏 delta 与动作三路流联合预测视觉变化量和动作，替代传统世界动作模型对完整未来帧的稠密预测。在 RoboTwin 上配合 Streaming Delta Memory（SDM）后，平均成功率从 Fast-WAM 的 81.3% 提升至 85.4%，视觉随机化下从 75.8% 提升至 83.9%。

研究者在将 Qwen3-4B 和 Qwen3-8B 蒸馏到 Qwen3-1.7B-Base 时发现，潜在监督会让 MATH-500 准确率在 10 步内从 25 升到 46，随后却崩到 11 且无法恢复，而对齐指标仍在持续改善。

Epoch AI 更新并扩展了 OECD 的半导体拆分投入产出表，将半导体增强版 ICIO 数据更新至 2016-2022 年，并按处理器、存储和其他芯片进一步细分，其中韩国出口的堆叠存储（含高带宽内存）单独可见。该表可追踪各国半导体的最终买家，并衡量各国支出对外国芯片生产商的暴露程度，用于校准贸易模型、估算贸易战或自然灾害等供应链冲击的影响。