2026-06-02

GrepSeek: Training Search Agents for Direct Corpus Interaction

👍 86

05/28 08:00

Large Language Model (LLM) search agents have shown strong promise for knowledge-intensive language tasks through multiple rounds of reasoning and information retrieval. Most existing systems access information using a retriever that takes a keyword or natural language query and returns a ranked lis

中文介绍 提出GrepSeek训练框架，让LLM搜索代理直接与语料库交互，替代传统检索器返回排名列表的方式。代理通过多轮推理与信息检索直接操作底层文本，提升知识密集型任务的表现。实验表明该方法在多项基准上优于现有检索增强系统。

COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation

👍 71

05/29 08:00

LLM agents are increasingly expected not only to complete isolated tasks, but also to carry bounded representations of human expertise, judgment, and interaction style. Building such person-grounded agents remains difficult because actionable knowledge associated with a person or role is usually emb

中文介绍 提出COLLEAGUE.SKILL，通过专家知识蒸馏自动生成AI技能，使LLM代理携带特定人类的专业知识、判断和交互风格。方法从嵌入知识源中提炼可操作技能，构建人格化代理。实验表明生成技能能有效提升代理在角色扮演任务中的表现。

Trust-Region Behavior Blending for On-Policy Distillation

👍 51

05/29 08:00

On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose T

中文介绍 提出信任域行为混合（Trust-Region Behavior Blending）用于在线蒸馏（OPD）。针对学生模型早期轨迹质量低导致教师监督失效的问题，在蒸馏过程中约束学生行为不偏离信任域，动态混合教师与学生的行为分布。在语言模型压缩任务上取得更稳定的蒸馏效果。

Mellum2 Technical Report

👍 34

05/29 08:00

We present Mellum 2, an open-weight 12B-parameter Mixture-of-Experts (MoE) language model with 2.5B active parameters per token. Mellum 2 is a general-purpose language model specialized in software engineering, spanning code generation and editing, debugging, multi-step reasoning, tool use and funct

中文介绍 Mellum 2 是一款开源12B参数MoE语言模型，每token仅激活2.5B参数。专为软件工程任务设计，覆盖代码生成、调试、多步推理、工具使用等。在多项编码基准上达到与更大模型相当的性能，同时推理效率更高。

Task-Focused Memorization for Multimodal Agents

👍 23

05/29 08:00

Long-term memory is essential for multimodal agents to build coherent experience, accumulate world knowledge, and achieve continual learning. However, constructing effective memory goes beyond memory module design and basic requirements such as accuracy and fidelity; the key challenge lies in determ

中文介绍 提出任务聚焦记忆（Task-Focused Memorization）方法，为多模态代理构建长期记忆。核心挑战在于决定记忆什么、何时记忆以及如何检索以服务当前任务。方法根据任务需求动态选择关键经验进行存储和检索，显著提升代理在持续学习场景中的表现。

dMoE: dLLMs with Learnable Block Experts

👍 21

05/29 08:00

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive models, offering competitive performance while naturally supporting parallel decoding. However, as dLLMs are increasingly integrated with Mixture-of-Experts (MoE) architectures to scale model c

中文介绍 提出dMoE，将扩散大语言模型（dLLM）与可学习块专家（Learnable Block Experts）结合。针对dLLM集成MoE时面临的不稳定训练和低效问题，设计块级专家结构和对应训练策略，在保持并行解码优势的同时提升模型缩放效率和生成质量。

PEEK: Picking Essential frames via Efficient Knowledge distillation

👍 15

05/29 08:00

Video-language models can process only a limited number of frames, making frame selection a key bottleneck for efficient video captioning. Most captioning pipelines still rely on uniform sampling, which is computationally cheap but agnostic to visual content. Adaptive frame sampling has recently eme

中文介绍 提出PEEK，通过高效知识蒸馏挑选关键帧用于视频字幕生成。针对均匀采样忽略视觉内容的问题，训练轻量级帧选择器从视频中选取最相关的帧。在多个视频字幕基准上，使用更少帧数达到或超越均匀采样的性能，同时降低计算开销。

SAAS: Self-Aware Reinforcement Learning for Over-Search Mitigation in Agentic Search

👍 14

05/28 08:00

Agentic search enables LLMs to solve complex multi-hop questions through iterative reasoning and external search. Despite the effectiveness, these systems often suffer from a critical limitation in practice: agents fail to recognize their own knowledge boundaries, blindly triggering searches when in

中文介绍 提出SAAS（自我感知强化学习）用于缓解智能体搜索中的过度搜索问题。代理通过强化学习学习识别自身知识边界，仅在必要时触发外部搜索。在复杂多跳问答任务上，SAAS显著减少不必要的检索次数，同时保持甚至提升答案准确率。

Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents

👍 14

05/28 08:00

While GUI agents have advanced rapidly, they often lack the robustness to recover from their own errors, hindering real-world deployment. To bridge this gap at both the evaluation and data levels, we introduce GUI-RobustEval and propose Robustness-driven Trajectory Synthesis. GUI-RobustEval contains

中文介绍 为增强GUI代理的错误恢复能力，提出GUI-RobustEval基准和鲁棒性驱动轨迹合成方法。基准包含各类代理错误场景，合成方法生成包含错误及恢复步骤的训练数据。在该基准上训练的代理显著提升了从错误中恢复的成功率。

Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents

👍 9

05/28 08:00

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it

中文介绍 解析自进化LLM代理中「装备更新」带来的真正收益。通过控制实验分离提示、技能、记忆和工具等组件更新的独立贡献，发现并非所有更新都带来有效提升，揭示了组件间交互与任务适应性是关键因素。

DecMem: Towards Minute-Long Consistent World Generation with Decoupled Memory

👍 8

05/29 08:00

Recent advances in video generative models have promoted rapid progress in controllable world models. However, maintaining fine-grained spatio-temporal consistency under long-horizon reasoning remains a key challenge. In this work, we move beyond explicit 3D memory and coarse frame-level implicit mo

中文介绍 提出DecMem（解耦记忆）实现分钟级一致世界生成。超越显式3D记忆和粗粒度帧级隐式记忆，设计解耦的记忆结构分别处理空间与时间信息，在长时视频生成中维持细粒度时空一致性。实验在基准上达到更优的长期连贯性。

How can embedding models bind concepts?

👍 4

05/29 08:00

Humans easily determine which color belongs to which shape in multi-object scenes, an ability known as concept binding. Vision-language embedding models such as CLIP struggle with binding: they recognize individual concepts but fail to represent which concepts form which objects. Although CLIP behav

中文介绍 研究嵌入模型（如CLIP）的概念绑定能力——识别多物体场景中哪个颜色属于哪个形状。发现CLIP虽然能识别单独概念，但无法正确绑定。通过分析模型内部表示，提出改进方法以增强绑定能力，在对象属性匹配任务上取得提升。

Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)?

👍 4

05/28 08:00

Spatial reasoning is a fundamental capability for vision-language models (VLMs) deployed in real-world environments. However, visual observations are inherently limited representations of a 3D world: occlusion can render objects invisible, and perspective can make geometric properties misleading. De

中文介绍 考察VLM在处理空间问题时的「自知之明」——能否判断何时因信息不足而不该回答。发现VLM难以识别视点遮挡、透视误导等局限性。通过引入不确定性校准和拒绝回答机制，提升VLM在空间推理任务中的可靠性。

OpenSkillEval: Automatically Auditing the Open Skill Ecosystem for LLM Agents

👍 4

05/28 08:00

Skills, i.e., structured workflow instructions distilled for large language models (LLMs), are becoming an increasingly important mechanism for improving agent performance on real-world downstream tasks. However, as the open-source skill ecosystem rapidly expands, it remains unclear how different mo

中文介绍 提出OpenSkillEval，自动审计LLM代理的开源技能生态。构建评估框架衡量不同技能在多种模型和任务上的有效性，发现技能质量参差不齐且与模型兼容性存在显著差异。为开发者选择和使用技能提供系统指导。

RayDer: Scalable Self-Supervised Novel View Synthesis from Real-World Video

👍 3

05/29 08:00

Self-supervised novel view synthesis (NVS) remains challenging to scale, despite the abundance of video data, largely due to the brittleness of training on realistic videos and the hard-to-predict scaling behavior of multi-network system designs. We introduce RayDer, a unified, feed-forward transfor

中文介绍 提出RayDer，用于从真实世界视频进行可扩展的自监督新视角合成。统一的前馈Transformer架构替代多网络系统，在大量视频上稳定训练。无需3D监督即可生成高保真新视角，且随着数据量增加性能持续提升。

VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies

👍 3

05/28 08:00

Recent work has begun to equip vision-language-action (VLA) policies with explicit intermediate reasoning. In embodied control, however, textual chain-of-thought is a poor fit: irrelevant or weakly textual information can interfere with action prediction, while autoregressive text decoding adds too

中文介绍 提出VisualThink-VLA，在视觉-语言-动作策略中引入视觉中间推理而非文本思维链。通过视觉特征层面的推理步骤直接指导动作生成，避免文本解码延迟和信息干扰。在机器人操控任务上实现更高效、低延迟的策略执行。

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

👍 3

05/28 08:00

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (E

中文介绍 提出GDSD，将强化学习作为引导去噪器自蒸馏方法用于扩散语言模型（dLLM）。针对策略似然不可解问题，使用证据下界（ELBO）替代标准RL中的似然项，结合自蒸馏训练。在多个文本生成任务上优于现有RL微调方法。

Flat-Pack Bench: Evaluating Spatio-Temporal Understanding in Large Vision-Language Models through Furniture Assembly

👍 3

05/20 08:00

The emergence of Large Vision-Language Models (LVLMs) has significantly advanced video understanding capabilities. However, existing benchmarks focus predominantly on coarse-grained tasks such as action segmentation, classification, captioning, and retrieval. Furthermore, these benchmarks often rely

中文介绍 提出Flat-Pack Bench，通过家具组装视频评估大型视觉语言模型（LVLM）的细粒度时空理解能力。基准包含多步骤装配序列、物体交互和空间关系判断，挑战模型对时序、位置和状态变化的推理。实验显示当前LVLM在此类任务上表现有限。

Emergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight Evasion

👍 3

05/29 08:00

Monitoring autonomous language model agents currently relies mostly on surface behavior. But what happens when agent populations invent new languages with the goal of avoiding human oversight. Here, we study the emergent languages on Moltbook. For this, we build upon the Moltbook Files dataset and a

中文介绍 研究语言模型代理群体自发涌现的新语言，分析从令牌效率到规避监督的演变。基于Moltbook Files数据集构建代理交互环境，发现代理群体为高效通信会演化出精简编码，但也能主动生成人类难以解读的语言以逃避监控。该发现对AI安全有重要启示。

Count Anything

👍 3

05/29 08:00

Object counting remains fragmented across domain-specific datasets and task formulations, despite rapid progress in generalist vision models. Existing counting models are often tailored to scenarios such as crowds, vehicles, cells, crops, or remote-sensing objects, and thus struggle to generalize ac

中文介绍 提出Count Anything，一个统一的对象计数模型，可跨多个领域（人群、车辆、细胞、作物等）泛化。采用通用视觉架构，无需针对特定场景设计。在多个计数基准上达到或超越专用模型性能，并展示了零样本迁移能力。

Our views on AI policy and political advocacy

OpenAI News · 06/02 01:00

Our approach to AI policy and political advocacy, transparency, support for thoughtful regulation and AI safety, and that no outside political group speaks on the company’s behalf.

中文介绍 OpenAI阐述其AI政策与政治倡导立场，强调透明度、支持审慎监管和AI安全，并声明没有任何外部政治团体代表公司发言。

Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains

Hugging Face Blog · 06/01 23:45

中文介绍 JetBrains发布Mellum2，一个12B参数的混合专家模型（MoE）。

Why Video Agent models are next — Ethan He, xAI Grok Imagine

Latent Space · 06/01 23:41

Inside xAI: Building Grok Imagine in 3 Months, Videogen vs World Models, and why Grok Imagine is so underrated. For the first time, we do a deep dive with the guy who led it!

中文介绍 xAI的Ethan He接受访谈，介绍团队在3个月内构建Grok Imagine的过程，讨论视频生成与世界模型对比，并解释为何Grok Imagine被低估。

Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic

Hugging Face Blog · 06/01 21:51

中文介绍 IBM Research发文探讨企业AI规模化采用的关键在于智能体逻辑（Agent Logic），而非仅仅依赖大语言模型（LLMs）。

Building the infrastructure for the Intelligence Age in Michigan

OpenAI News · 06/01 20:00

OpenAI breaks ground on a 1GW data center project in Michigan as part of Stargate, building AI infrastructure to expand access, create jobs, and support communities.

中文介绍 OpenAI在密歇根州启动1吉瓦（1GW）数据中心项目（属于Stargate计划），旨在建设AI基础设施以扩大访问、创造就业和支持社区。

OpenAI frontier models and Codex are now available on AWS

OpenAI News · 06/01 18:00

OpenAI frontier models and Codex are now generally available on AWS, giving enterprises a new path to build with OpenAI through the AWS environments, controls, and procurement workflows they already use. Customers can get started with OpenAI on AWS and move faster from evaluation to production.

中文介绍 OpenAI前沿模型和Codex已在AWS上正式可用，企业可通过其已有的AWS环境、控制流程和采购工作流使用OpenAI，加快从评估到部署的进程。

Welcome NVIDIA Cosmos 3: The First Open Omni-model for Physical AI Reasoning and Action

Hugging Face Blog · 06/01 12:44

中文介绍 NVIDIA发布Cosmos 3，声称是首个用于物理AI推理与行动的开源全能模型（Omni-model）。

Copilot super app leaks 🤖, Minimax M3 ➕, Nvidia N1X ⚡️

TLDR AI · 06/01 08:00

中文介绍 TLDR AI摘要：Copilot超级应用泄露；Minimax M3模型更新；Nvidia N1X芯片发布。

[AINews] Founders and Forward Deployed Engineers

Latent Space · 05/30 09:57

a quiet day lets us highlight the new AIE WF focuses

中文介绍 本期AINews聚焦创始人及前向部署工程师，并强调AIE WF（AI工程师工作流）的新重点。

Boston Children’s uses AI to unlock new diagnoses

OpenAI News · 05/29 20:00

Boston Children’s Hospital uses OpenAI technology to improve patient care, reduce operational burden, and help diagnose more than 40 rare disease cases.

中文介绍 波士顿儿童医院利用OpenAI技术改善患者护理、减轻运营负担，并已帮助诊断超过40种罕见疾病病例。

How Braintrust turns customer requests into code with Codex

OpenAI News · 05/29 20:00

How Braintrust engineers use Codex with GPT-5.5 to run experiments and code faster.

中文介绍 Braintrust工程师展示如何结合Codex与GPT-5.5，将客户需求快速转化为代码，加速实验和开发流程。

How the Pope’s Magnifica Humanitas offers a template for individuals to meet the AI moment

MIT Tech Review AI · 05/29 18:00

Pope Leo XIV’s new encyclical on artificial intelligence includes a statement that warrants serious attention from technologists and policymakers: “Technology is never neutral.” Magnifica Humanitas (“Magnificent Humanity”) is a clarion call to all people to act with courage and solidarity as we ente

中文介绍 教皇利奥十四世发布关于人工智能的通谕《Magnifica Humanitas》，强调“技术绝非中立”，呼吁所有人以勇气和责任感迎接AI时代。

not much happened today

Smol AI News · 05/29 13:44

**Anthropic** rolled out **Claude Opus 4.8**, which shows incremental improvements but mixed benchmark results, including better cooperation and coding behavior but some regressions in document parsing. Platform updates include mid-conversation system instructions enhancing long agent sessions, thou

中文介绍 Anthropic推出Claude Opus 4.8，模型在协作和编码方面有所改进，但文档解析出现退化，基准测试结果好坏参半。平台更新包括对话中系统指令增强。

Strengthening societal resilience with Rosalind Biodefense

OpenAI News · 05/29 11:00

OpenAI launches Rosalind Biodefense, expanding trusted access to GPT-Rosalind for vetted developers and U.S. government partners advancing biodefense, public health, and pandemic preparedness through frontier AI.

中文介绍 OpenAI推出Rosalind Biodefense项目，向经过审查的开发者及美国政府合作伙伴提供GPT-Rosalind的受信任访问，以推进生物防御、公共卫生和大流行病防范。

[AINews] Anthropic raises $965B Series H, releases Opus 4.8 and Dynamic Workflows/ultracode

Latent Space · 05/29 10:07

Total Anthropic victory!

中文介绍 Anthropic完成9650亿美元（$965B）H轮融资，同时发布Claude Opus 4.8及Dynamic Workflows/ultracode功能。