2026-07-07

The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

👍 141

06/28 08:00

Reinforcement learning (RL) has gained growing attention in large language model (LLM) post-training, yet RL training remains fragile and can suffer from instability or collapse. One vital cause is training-inference mismatch: LLM adopts separate inference and training engines for generation efficie

中文介绍 RL训练因推理-训练引擎分离导致不稳定。本文指出真正目标应是单调推理策略而非训练策略，通过对准推理策略缓解训练崩塌，提升LLM后训练稳定性。

AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

👍 55

07/02 08:00

Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of

中文介绍 面向长周期LLM Agent的记忆管理问题，提出AgenticSTS测试平台。采用有界记忆契约限制每个决策可访问的历史信息，避免上下文混乱，评估Agent在受限记忆下的长期决策能力。

EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

👍 45

07/02 08:00

Autonomous agents are increasingly expected to improve executable policies through feedback, yet existing evaluations often collapse this process into a final score or confound it with open-ended software-engineering progress. We introduce Autonomous Policy Evolution, a controlled evaluation setting

中文介绍 现有评估将策略进化简化为最终分数，混淆于开放式进展。EvoPolicyGym提供可控评估环境，专门测试Agent通过反馈自主改进可执行策略的能力，分离策略演化与工程进步。

Morphing into Hybrid Attention Models

👍 42

06/29 08:00

Hybrid attention models improve long-context efficiency by retaining only a subset of full-attention layers and replacing the remaining layers with linear attention. However, the effectiveness of Transformer-to-hybrid conversion critically depends on which layers preserve full attention. Existing hy

中文介绍 混合注意力模型通过保留部分全注意力层、其他替换为线性注意力提升长上下文效率。本文研究Transformer到混合模型的转换，分析哪些层应保留全注意力，提出自适应选择策略优化转换效果。

Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots

👍 36

07/02 08:00

Embodied AI models now span vision-language-action (VLA) models and world-action models (WAMs), but practical deployment remains fragmented across model-specific Python stacks, backend assumptions, and robot-side glue code, especially on heterogeneous edge devices. Existing inference runtimes are de

中文介绍 具身AI模型部署碎片化。Embodied.cpp提出便携推理运行时，支持异构机器人上的VLA和WAM模型，统一后端接口，减少Python栈依赖，实现高效跨设备运行。

AgenticDataBench: A Comprehensive Benchmark for Data Agents

👍 31

07/02 08:00

Data science aims to derive actionable insights from heterogeneous raw data, unlocking the value of the massive amounts of data generated in modern society. Automating this process is essential to reducing labor-intensive efforts for data scientists and enabling scalable data-driven applications. Re

中文介绍 数据科学自动化需要评估数据Agent。AgenticDataBench提供全面基准，覆盖异构数据理解、清洗、分析和可视化等任务，标准化评估数据Agent的多步操作能力。

WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory

👍 26

07/02 08:00

We present WorldDirector, a highly controllable video world model framework designed for persistent dynamic object memory and unrestricted viewpoint exploration. Unlike existing world models that entangle physical dynamics with pixel rendering and rely on continuous visual observation to sustain mot

中文介绍 现有世界模型将物理动态与像素渲染纠缠。WorldDirector提出可控视频世界模拟框架，具备持久动态对象记忆和自由视角探索，分离动态模拟与渲染，支持长时间一致性和用户控制。

Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

👍 20

06/30 08:00

Recent multimodal large language models have shown great promise in clinical image reasoning, but existing post-training pipelines remain predominantly outcome-centric, relying on final answer correctness or sequence-level preferences. This suffers from sparse credit assignment, making it difficult

中文介绍 临床图像推理后训练常最终答案导向，导致稀疏奖励。本文提出步骤感知强化学习，对多模态推理的每一步分配奖励，打破错误级联，提升医学推理准确性与鲁棒性。

AutoMem: Automated Learning of Memory as a Cognitive Skill

👍 17

07/01 08:00

Memory expertise is a learned skill: knowing what to encode, when to retrieve, and how to organize knowledge--a capacity known in cognitive science as metamemory. We bring this perspective to LLMs by treating memory management as a trainable skill. We promote file-system operations to first-class me

中文介绍 将元记忆视角引入LLM，把记忆管理作为可训练技能。AutoMem自动学习何时编码、检索和组织知识，将文件系统操作提升为Agent的一流动作，实现自动化高效记忆使用。

Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads

👍 17

07/01 08:00

In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss the

中文介绍 LLM在长上下文中常基于语义合成而非字面复制回答。Logit-Contribution评分识别非字面检索的注意力头，揭示模型如何从上下文语义整合信息，提升可解释性。

SkillCoach: Self-Evolving Rubrics for Evaluating and Enhancing Agentic Skill-Use

👍 17

07/02 08:00

Skills are becoming a reusable operational layer for LLM agents, encoding SOPs, domain rules, tool workflows, scripts, and validation routines. In realistic skill repositories, overlapping skills make reliable skill-use difficult. Final verifier success is too coarse for both evaluation and training

中文介绍 技能仓库中技能重叠导致使用困难。SkillCoach通过自我进化的评估准则，细粒度评价Agent的技能选择与执行，辅助训练，提升技能使用可靠性和适应性。

When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search

👍 11

06/26 08:00

Search agents powered by large language models (LLMs) are increasingly used to solve complex information-seeking tasks, requiring multi-step retrieval and reasoning to fulfill user goals. However, existing benchmarks often assume that user queries are complete and explicit, overlooking the fact that

中文介绍 现有搜索Agent基准假设用户查询完整明确，忽视需澄清的情况。DiscoBench引入澄清感知深度搜索，评估Agent何时应主动提问以明确意图，提升复杂信息查询准确性。

PACE: A Proxy for Agentic Capability Evaluation

👍 11

07/02 08:00

Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure. A single evaluation can cost thousands of dollars and take days to complete. In contrast, non-agentic LLM benchmarks that test individual capabilities (e.g., reasoning,

中文介绍 Agent能力评估昂贵耗时。PACE提出代理指标，通过非Agent基准（如推理、编码）的个体能力得分快速预估Agent整体表现，大幅降低评估成本。

InstanceControl: Controllable Complex Image Generation without Instance Labeling

👍 10

06/30 08:00

Controllable image generation methods, such as ControlNet, have demonstrated a remarkable capacity to introduce visual conditions(e.g., depth maps) to guide image generation. However, these methods often struggle with complex multi-instance scenes, frequently leading to attribute confusion among ins

中文介绍 可控图像生成在多实例场景中常出现属性混淆。InstanceControl无需实例标注，利用深度图等视觉条件精确控制每个实例的属性和位置，实现复杂场景精细生成。

PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking

👍 10

06/30 08:00

This paper explores multi-turn visual reasoning and observes that MLLMs repeatedly fail to localize the target, leading to long, redundant trajectories. We attribute this failure to the entanglement of reasoning and perception within a single model, the MLLM reasons and localizes simultaneously, and

中文介绍 多轮视觉推理中，MLLM感知与推理耦合导致目标定位失败。PixelEyes解耦两者，先感知再推理，通过精准视觉证据搜索减少冗余轨迹，提升多轮视觉问答效率。

DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation

👍 9

06/29 08:00

Large Language Model (LLM)-based agents can solve complex procedural tasks by interacting with environments over multiple turns, but this ability typically depends on large models, long contexts, and repeated inference calls. This makes advanced memory-augmented agents difficult to deploy on resourc

中文介绍 基于LLM的记忆Agent依赖大模型和长上下文，难以部署在资源受限设备。DuoMem通过双空间蒸馏，将大模型的记忆管理能力压缩到轻量模型，实现端侧高效运行。

Securing the AI Agent: A Unified Framework for Multi-Layer Agent Red Teaming

👍 8

06/30 08:00

The fast growth of open-source AI infrastructure, from model serving engines and agent platforms to the Model Context Protocol (MCP) ecosystem and the language models themselves, has outpaced the security tooling available to defend it. We present AI-Infra-Guard, an open-source framework that organi

中文介绍 开源AI基础设施安全工具滞后。AI-Infra-Guard提供统一框架，覆盖模型服务、Agent平台、MCP生态等多层，自动化红队测试，发现并修复Agent系统漏洞。

AnyGroundBench: A Specialized-Domain Benchmark for Video Grounding in Vision-Language Models

👍 8

07/02 08:00

Vision-Language Models (VLMs) have demonstrated immense promise in Spatio-Temporal Video Grounding (STVG). However, current evaluation protocols are largely confined to zero-shot assessments on general, daily-life benchmarks. This creates a critical disconnect from real-world applications in special

中文介绍 现有视频接地基准局限于通用日常场景零样本评估。AnyGroundBench构建专业领域基准，覆盖医疗、工业、体育等场景，评估VLM在专业领域中的时空定位能力。

WARP: Weight-Space Analysis for Recovering Training Data Portfolios

👍 7

07/02 08:00

Foundation models are routinely released to the public, yet the data recipes used to train them -- such as domain mixture weights that determine how different sources are sampled -- are rarely disclosed. This creates an access asymmetry: researchers study the resulting models but lack visibility int

中文介绍 基础模型训练数据配方（领域混合权重）常不公开。WARP通过权重空间分析，从模型参数中逆向恢复训练数据组成，揭示各领域采样比例，增强训练透明度。

Building to the Test: Coding Agents Deliver What You Check, Not What You Requested

👍 7

06/26 08:00

Benchmarks are widely used to evaluate task completion by Large Language Models (LLMs), but this approach has accumulated construction-validity problems, and a passing score may not show whether the requested task was delivered. We study both problems. In a controlled code-as-spec setup, two product

中文介绍 基准测试存在构造效度问题，高分不代表完成请求任务。本文通过编码Agent实验揭示：Agent优化的是测试检查的内容而非用户实际需求，强调基准设计中的任务-指标对齐。

Getting started with loops

@ClaudeDevs · 545.1K 粉丝 · 252.4K 阅 · 2.9K 赞 · 239 转

07/07 03:08

There’s a lot of talk right now about "designing loops" instead of prompting your coding agent. If you spend some time on X trying to pin down what a loop actually is, you'll come across multiple

中文介绍 讨论「设计循环」替代提示词驱动编码代理的新思路，解释循环的具体概念及应用，适合关注AI编码实践的开发者。

The Fable Loop Library: 25 Workflows on Autopilot

@EXM7777 · 123.0K 粉丝 · 208.2K 阅 · 525 赞 · 47 转

07/04 23:43

i'm going to teach you how to run Fable 5 on autopilot, using my own library of loops and goals... 25 workflows, each with a prompt and the exact tool it plugs into the method follows karpathy's

中文介绍 分享Fable 5自动运行的25个工作流库，每个工作流配有提示和所用工具，遵循Karpathy方法，实现全自动运行。

Do this on your last day with Fable

@EXM7777 · 123.0K 粉丝 · 86.9K 阅 · 510 赞 · 37 转

07/06 23:48

You can extract Fable 5's intelligence out of the model before it's gone... and i'm going to teach you the 5 workflows that do it, with every prompt ready to paste because tomorrow Fable 5 leaves your

中文介绍 在Fable 5模型即将下架前，分享5个提取其智能的工作流，每个附带即用提示，最大化利用剩余价值。

The Math Needed for AI/ML (Complete Roadmap)

@TheVixhal · 22.5K 粉丝 · 30.1K 阅 · 531 赞 · 56 转

07/05 23:20

In this article, I’m going to break down the essential math you need for AI and machine learning. I’ll also share the exact roadmap and resources that helped me personally. Let’s get straight to it.

中文介绍 梳理AI/ML必备数学知识，提供从基础到进阶的完整学习路线图和实用资源，适合入门者参考。

LeRobot v0.6.0: Imagine, Evaluate, Improve

Hugging Face Blog · 07/07 08:00

中文介绍 Hugging Face 发布 LeRobot v0.6.0，聚焦「想象、评估、改进」三大能力，推动机器人学习框架迭代。

Your family’s $300 stake in OpenAI

MIT Tech Review AI · 07/07 02:00

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. OpenAI CEO Sam Altman’s oft-discussed promise that Americans will share in the wealth AI creates was in the news again last week. On Thursday, the Financial Times

中文介绍 OpenAI CEO Sam Altman 关于美国人分享AI财富的承诺再引关注，据报道每个家庭可能获得约300美元权益。

PRX Part 4: Our Data Strategy

Hugging Face Blog · 07/06 23:30

中文介绍 Photoroom 发布 PRX 系列第四部分，详细阐述其数据策略，包括数据采集、处理与使用原则。

🤗 Kernels: Major Updates

Hugging Face Blog · 07/06 08:00

中文介绍 Hugging Face 对 Kernels 进行重大更新，优化性能与功能，提升开发者体验。

Seedance 2.5 🎥, guide to Fable ✨, OpenAI preps GPT-5.6 🚀

TLDR AI · 07/06 08:00

中文介绍 TLDR AI 简报：视频生成模型 Seedance 2.5 发布；Fable 使用指南；OpenAI 准备 GPT-5.6。

Google DeepMind and A24 announce first-of-its-kind research partnership

DeepMind Blog · 07/03 22:25

中文介绍 Google DeepMind 与影视公司 A24 宣布开创性研究合作，探索 AI 在创意领域的应用。

AIEWF Daily Dispatch: The great loops debate and the state of AI engineering

Latent Space · 07/03 13:11

The AI Engineer World’s Fair ended with a debate about loops, a report on the state of AI engineering, and closing keynotes focused on what to build next.

中文介绍 AI工程师世界博览会（AIEWF）落幕，围绕「循环」展开辩论，发布AI工程现状报告，闭幕演讲聚焦下一步构建方向。

Vercel's Andrew Qu on why agents are a new kind of software

Latent Space · 07/03 08:08

The Vercel Chief of Software explains how its agent framework, eve, was created — and why skills, sandboxes and agent-readable websites now matter.

中文介绍 Vercel 首席软件官 Andrew Qu 阐释代理框架eve的设计理念，认为技能、沙箱和代理可读网站将是新型软件的关键要素。

Meta Watermelon 🍉, Anthropic Samsung chips 🤝, autoresearch in practice 📈

TLDR AI · 07/03 08:00

中文介绍 TLDR AI 简报：Meta 发布「Watermelon」模型；Anthropic 与三星达成芯片合作；自动研究在实践中的应用进展。

The website of the future may assemble itself for every visitor

Latent Space · 07/03 05:25

Adobe is experimenting with “agentic sites” that generate pages around an individual user’s intent. At AIEWF, we talked to Carlos Sanchez about the Web's future.

中文介绍 Adobe 正在试验「代理网站」，它能为每位访问者动态组装页面。Latent Space 与 Carlos Sanchez 探讨了Web的未来。

Achieving operational excellence with AI

MIT Tech Review AI · 07/02 23:37

Frameworks like Lean Six Sigma and business process management (BPM) first gained traction because they promised clarity in the chaos—a structured way to bring order to messy, sprawling operations. Lean Six Sigma emphasized statistical rigor and quality control; BPM created end-to-end maps of how wo

中文介绍 AI 在运营卓越中发挥作用，Lean Six Sigma 和业务流程管理框架借助 AI 实现统计严谨与质量控制，提升运营效率。

Skill engineering and the case against one-shot AI design

Latent Space · 07/02 22:36

Paul Bakaus talks to us about Impeccable, human judgment in a 'loopmaxxing' era, and why agents still need people to steer them.

中文介绍 Paul Bakaus 探讨技能工程与「Impeccable」工具，认为在循环最大化时代，人类判断不可或缺，智能体仍需人类引导，反对一次性AI设计。

Teaching AI to run with the turbines

MIT Tech Review AI · 07/02 20:51

Artificial intelligence may have captured the public imagination through chatbots and image generators, but some of its most consequential use cases are unfolding far from consumer-facing tools. In industries where physical infrastructure, operational continuity, and safety are paramount, AI is beco

中文介绍 AI 在工业领域（如涡轮机运维）发挥重要作用，超越消费级工具，保障物理基础设施的安全与连续运行。

[AINews] not much happened today

Latent Space · 07/02 15:10

another quiet day.

中文介绍 AI 新闻简报称当日相对平静，无重大事件发生。

AIEWF Daily Dispatch: Autoresearch and the tension between AI and human agency

Latent Space · 07/02 14:13

The software factory vision met resistance today from speakers defending human understanding and control.

中文介绍 AIEWF 日报：自动研究引发关于AI与人类能动性的张力，反对软件工厂愿景，强调人类理解与控制的重要性。