AI 论文与研究

OmniSeek 是一个将 Omni-LLM 转变为原生工具调用多轮推理智能体的框架，让模型在推理中动态决定看或听、以及选取哪段时间窗口，把检索到的原始音视频片段回填上下文。

InterEvolve 通过测试时进化奖励程序，让控制器无需重新训练即可完成未训练过的人形机器人移动操作任务。该方法结合物体感知的前向-后向行为基础模型与 LLM 智能体，后者依据执行反馈和已验证程序技能库修改奖励程序结构，数值优化器负责调参。实验显示，进化出的程序释放了 FB 模型未被人工奖励触及的能力，并可在 Unitree G1 实体机器人上基于第一视角感知自主运行。

AI Theorist 是一套通过假设生成、第一性原理计算与证据驱动迭代来自主发现物理模型的 AI 智能体系统，被用于研究 Kitaev 量子自旋液体候选材料 α-RuCl_3 的光谱。它对该材料的光学与光电流观测给出新解释，识别出具有不同光学选择定则和实空间分布的激子态。据称这是首个 AI 系统自主建立物理模型、解释量子材料中此前未发表实验观测的演示。

DeskForge 是一个可组合真实桌面应用的可控环境，用于为 computer-use agent 生成大规模监督数据，并构建了包含 1.2M 条标注桌面观测、159.7M 个元素实例的 DeskForge-1M 语料。

Where-OPD 提出一种面向多模态大模型（MLLM）的在线自蒸馏方法，让教师模型获得文本形式的空间定位引导，从而定位并整合多个相关图像区域的证据，学生模型仅凭图像和问题学习复现该行为。

研究者提出 InFlowOp，用统一的"无标签成本"为多智能体工作流中的每个决策定价，在执行前双向决定任务分解粒度与智能体分配，执行中以最低成本修正故障。团队同时发布需要多智能体协作的基准 Braid，在多个领域和骨干模型上，InFlowOp 较单智能体基线最高提升 +11.97%，流内优化带来 +9.64%。

针对 DDPO 等扩散模型强化学习后训练方法在奖励优化中牺牲多样性与质量的问题，研究者提出 iADD，通过理论分析证明仅对扩散模型后段 timestep 更新可能损害多样性，并基于增量 Feynman-Kac 训练实现更优的对齐-多样性权衡。在三个任务上的实验与消融显示，该方法在对齐和多样性两方面均取得显著提升。

研究者提出基于选择的结构化推理（SSR），将多模态智能体的推理从开放式生成改为从预设可复用自然语言候选中选择，无需辅助任务头，并借助共享上下文 KV cache 实现并行打分。在 2B 和 4B 模型、七个多模态搜索基准上，SSR 平均成功率与同规模领先搜索智能体相当，单轮推理延迟降低超 90%，每题模型总推理延迟降低 28-54%。

Ego2Act 是一个面向目标导向的第一视角视频生成基准，包含来自 110 个真实世界任务的 2,640 段视频，覆盖不同物体杂乱程度与多步复杂度。该基准要求模型根据初始场景图像和高层目标，生成手部操作物体完成任务的逼真第一视角视频，并配套推出无参考评估流程 Ego2ActJudge。

Argo-Bench 是一个包含 210 个数据科学与分析任务的企业级数据智能体评测框架，模拟纽约市外卖平台 2024 年 8100 万订单，导出为 235 张表、75 亿行的 ERP 仓库。任务要求智能体在仓库中重建事实并执行封禁欺诈账号、分配骑手激励预算等操作，由模拟器按后果评分。14 个前沿与开源权重模型中最强者仅在 34.8% 的任务上得分 95 以上，平均 59.5 分。

PyRUA-Lean 框架让 GPT-6 Astra 驱动的机器人智能体在同等 LLM 调用预算下，将 LIBERO-PRO、RoboTwin 2.0 和 RoboCasa365 共 700 个仿真任务的整体成功率从 63.1% 提升至 71.7%。

VTR-Bench 是用于评估视频生成模型视觉文本渲染能力的系统性基准，包含覆盖广告、科学视频等五类场景的 300 条提示词，并配套自动化评估流程。在 11 个 SOTA 模型上的实验显示，场景文本渲染普遍存在困难，表现最好的模型整体词错误率（WER）为 0.250。

AutoGUIWorld 是一个数据生成框架，用图像生成器的视觉先验加规划器的任务知识合成 GUI 交互轨迹，无需部署或运行真实软件环境。它从操作系统上下文、视觉外观和界面状态的结构化规格中采样初始场景，经动作 grounding 与转移级质量过滤，在 Ubuntu、Windows、macOS 和 Chrome 上产出 79,266 条带空间标注的 step 级训练样本。

Meta 等机构发布 Context Language Models（CLMs）论文，让语言模型把上下文当作文件、原生自主管理上下文，无需外部 harness。

LongLive-Plug 视频生成的通用蒸馏 论文：https://huggingface.co/papers/2609.38154

Omni-IO Skills 让你的智能体原生全能 paper: https://huggingface.co/papers/2609.31847

LEGO-Anything 用于 3D 场景重建的编码智能体 论文：https://huggingface.co/papers/2609.36380

安全研究员 Johann Rehberger 在 BlueHat Asia 2026 上报告 Microsoft SQL Server Management Studio 中 SQL Copilot 的 SQL Server 提权漏洞 CVE-2026-65669，微软评级为 critical，建议用户保持安装更新。

Multimodal large language models (MLLMs) are rapidly advancing clinical diagnosis, yet their adaptation pipelines remain anchored to accuracy-based objectives. Clinical data are heavily class-imbalanced: a constant-majority predictor can s…

Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions remain sensitive to task-irrelevant prompt features. We investigate this sensitivity throu…

We find that major reported improvements in decoding words from non-invasive brain recordings are largely reproducible without any brain data. In the influential work of d'Ascoli et al. (2025), time series of brain activity from subjects p…

Recent video generation is increasingly realistic and controllable, yet video editing remains less developed, particularly for precise local edits that must preserve the original scene dynamics. Video scene text editing replaces text on sc…

We introduce VideoMSN, a Masked Siamese Network framework for efficient self-supervised spatio-temporal representation learning in videos. Instead of relying on heavy 3D architectures or reconstruction-based autoencoders for learning with …

Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions…

Differentially Private Stochastic Gradient Descent (DP-SGD) is a leading approach for privacy-preserving fine-tuning of large language models (LLMs). Many decoder-only LLMs employ weight tying between input and output embeddings, a design …

Automating the search for effective harnesses is an important step toward enabling agents to recursively self-improve. Existing harness optimizations typically produce a single global harness that is applied uniformly across task instances…

As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or ob…

We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for op…

Material generation should produce not only an appearance, but also the rules that construct it. We introduce MatLoom, a compact, layer-oriented language for text-to-material generation with pretrained language models. Each program compose…

Looped transformers and Mixture-of-Experts (MoE) offer complementary routes to efficient scaling: recurrence increases computational depth at fixed parameters, while MoE sparsity expands total capacity at fixed active compute. Yet existing…

Lossy compression is widely used in Federated Learning (FL) but is generally treated as an error source, while conventional poisoning defenses inspect update geometry. In this work, we instead treat the compressor's response as a security …

Pretrained robot policies provide useful action priors, but long-horizon manipulation still requires coordination between semantic reasoning and physical execution. Semantic reasoning operates at a coarser timescale than physical interacti…

Improving text-to-image models has traditionally relied on increasing model size or the number of denoising steps. In this work, we explore an alternative way to scale computation by repeatedly running shared Transformer blocks within each…

Recent autonomous machine learning engineering (MLE) agents have made significant progress on public leaderboards. Often motivated by progress stagnation over long-horizon cycles and limited Large Language Model (LLM) primitives, modern ML…

Web text makes up the majority of pretraining data and is increasingly AI-generated. After applying FineWeb quality filtering, we find that 27.5% of tokens from June 2026 web data are labeled as AI-generated by Pangram, rising to 31.1% by …

How is computation organized and reused across tasks and time in a trained recurrent network? Most analyses emphasize the geometry of neural activity, dynamical motifs, or local perturbation growth. We instead study the network's global fi…

In algorithmic statistics a string x is explained by a finite set containing it, and Kolmogorov's structure function records the smallest such model at each level of complexity. Vereshchagin's strong models, those computable from the data …

Physics-constrained generative models aim to generate physical fields that match a target distribution and satisfy prescribed constraints. However, enforcing these constraints often increases sampling costs through iterative corrections or…

LOCI 是一种混合空间记忆架构，在部分 Transformer 块中保留过去观测的 key-value 缓存，其余块仅处理当前 chunk 并辅以读写受投影相机几何条件约束的循环线性注意力记忆，使视角同时进入记忆寻址与存储内容。

发表于斯坦福 Journal of Online Trust and Safety 的研究识别了88个活跃托管未经同意 AI 生成亲密图像的网站，指出 Cloudflare、Google、Namecheap、WordPress 和 Proton 五家为主要基础设施服务商，其中 Cloudflare 提供了大部分托管、CDN 和域名解析服务，尽管各家服务条款均禁止非法内容。