2026-07-06

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

👍 51

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 AgenticSTS提出一个有界记忆测试床，用于评估长时域LLM代理的记忆合约设计。通过限制记忆容量，避免简单拼接导致的信息混杂，为代理长期决策提供更清晰的记忆管理基准。

EvoPolicyGym: Evaluating Autonomous Policy Evolution in Interactive Environments

👍 43

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 EvoPolicyGym提出自主策略演化评估框架，在受控交互环境中衡量代理通过反馈持续改进可执行策略的能力，避免与最终得分或软件工程进展混淆，提供更细粒度的演化分析。

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 Morphing提出将Transformer逐步转换为混合注意力模型的方法，通过选择性地保留部分全注意力层、其余替换为线性注意力，兼顾长上下文效率与模型性能。该方法自动确定最优层配置，优于现有启发式策略。

PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception

👍 39

06/26 08:00

We introduce PerceptionRubrics, a rubric-based evaluation framework that addresses the gap between saturated benchmark scores and real-world brittleness. Shifting evaluation from holistic semantic matching to rigorous atomic auditing, PerceptionRubrics pairs 1,038 information-dense images with over

中文介绍 PerceptionRubrics提出基于评分准则的多模态评估框架，从整体语义匹配转向原子化审核，使用1038张高信息密度图像和大量细粒度问题，校准模型与人类感知的一致性，揭示基准饱和下的实际脆弱性。

AgenticDataBench: A Comprehensive Benchmark for Data Agents

👍 30

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 AgenticDataBench提出数据智能体综合基准，涵盖数据科学全流程（探索、清洗、分析、可视化），评估LLM代理自动化处理异构原始数据的能力，旨在减少人工劳动并推动可扩展数据应用。

WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory

👍 24

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 WorldDirector提出可控视频世界模型框架，具备持久动态对象记忆和自由视角探索能力。解耦物理动力学与像素渲染，使模型能长期记忆对象状态并支持任意视点生成，提升视频世界模拟的可控性和连贯性。

ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

👍 24

07/01 08:00

In prefill-decode (PD) disaggregated LLM serving, each request is assigned to a decode worker after prefill. Existing decode routers balance only load; for mixture-of-experts (MoE) models this is incomplete: equally loaded workers can differ in latency, since each decode step loads the weights of ev

中文介绍 ELDR提出专家局部性感知的解码路由策略，用于预填充-解码分离的MoE模型服务。针对同等负载下因专家权重加载不同导致延迟差异的问题，通过路由考虑专家局部性，降低平均解码延迟，提升MoE服务效率。

Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning

👍 22

07/01 08:00

Multimodal Large Language Models (MLLMs) are often constrained by a language-space bottleneck, forcing complex visual reasoning into discrete tokens which can lose perceptual nuance. A promising alternative is continuous latent reasoning, where the goal is to discover implicit reasoning pathways tha

中文介绍 针对多模态大模型离散语言标记丢失视觉细节的问题，提出非对称互变分学习方法，在连续潜空间进行多模态推理。通过视觉与语言模态的非对称变分编码，发现隐式推理路径，保留感知细粒度，提升复杂视觉推理质量。

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 针对医学多模态推理中结果导向训练导致稀疏奖励难以定位错误步骤的问题，提出步骤感知强化学习方法。通过为推理过程每一步提供细粒度反馈，打破错误级联，显著提升临床图像推理的准确性与鲁棒性。

AutoMem: Automated Learning of Memory as a Cognitive Skill

👍 16

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 AutoMem将记忆视为可训练认知技能，在LLM中通过文件系统操作（编码、检索、组织）学习元记忆策略。使代理自主决定存储内容与时机，提升长期交互中的记忆管理效率和任务表现。

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

👍 16

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 SkillCoach提出自演化评分准则，用于评估和提升LLM代理的技能使用能力。通过生成细粒度、可更新的技能使用标准，解决仓库中技能重叠导致的选用困难，优于粗粒度最终验证器。

Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads

👍 15

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 提出Logit-Contribution Scoring（LCS）方法，识别LLM中执行非字面检索的注意力头——即基于语义合成而非直接复制。该方法通过分析标记贡献，发现现有检测器遗漏的合成行为，增强长上下文模型可解释性。

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

👍 10

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 DiscoBench提出澄清感知深度搜索基准，评估LLM搜索代理在用户查询不完整或模糊时主动提问澄清的能力。包含多步检索与推理任务，填补现有基准忽视查询歧义的空白。

InstanceControl: Controllable Complex Image Generation without Instance Labeling

👍 9

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 InstanceControl提出无需实例标注的可控图像生成方法，解决多实例场景中属性混淆问题。通过隐式实例分解和条件引导，在不依赖人工标注前提下实现精细的多对象布局与属性控制。

Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination

👍 9

07/01 08:00

Accelerating materials discovery requires AI systems that can generate scientifically valid hypotheses through multi-step, domain-grounded reasoning. Standard large language models often produce fluent but weakly traceable responses to open-ended materials design problems, making it difficult to det

中文介绍 提出图原生强化学习方法，通过概念重组生成可追踪的科学假设。在材料设计领域，将推理过程建模为图结构，结合RL使每一步假设可溯源，克服LLM生成结果无法验证的局限。

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

👍 8

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 指出LLM后训练中强化学习不稳定的根源在于训练-推理策略不匹配。提出优化单调推理策略比训练策略更重要，通过使推理引擎与训练目标对齐，避免策略崩溃，提升RL训练的稳定性和最终性能。

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

👍 8

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 DuoMem通过双空间蒸馏（语义空间与记忆空间），将大模型的内存增强能力迁移至小型设备。在资源受限环境中实现多轮任务的内存管理，兼顾性能与效率，推动端侧智能代理部署。

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 AnyGroundBench提出专门领域视频定位基准，评估VLM在特殊领域（如医疗、工业）的时空视频定位能力。填补现有基准仅覆盖日常场景的空白，推动VLM在专业场景的鲁棒性。

PixelEyes: Decoupling Perception and Reasoning for Pinpoint Visual Evidence Seeking

👍 8

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 PixelEyes通过解耦感知与推理解决多轮视觉定位中的冗余问题。将目标定位与逻辑推理分离，先精确感知再推理，消除MLLM在复杂场景中的无效搜索路径，提升证据定位效率。

PACE: A Proxy for Agentic Capability Evaluation

👍 7

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 PACE提出代理能力评估代理指标，利用非代理基准（推理、编码等）预测LLM代理在SWE-Bench等复杂任务上的表现。大幅降低评估成本与时间，提供高效、可扩展的代理能力预估方法。

The Fable Loop Library: 25 Workflows on Autopilot

@EXM7777 · 122.0K 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 EXM7777 分享 Fable Loop Library，包含 25 个自动化工作流，每个配提示词和工具接口，方法沿袭 Karpathy 思路，让 Fable 5 自动运行。

Your AI, your growth

@arthurmensch · 74.0K 粉丝 · 162.7K 阅 · 539 赞 · 76 转

07/04 05:31

Of course you need to use open-source models if you’re an enterprise leader. Close model providers, that are now forcing data retention, are gaining immense leverage on your business if you don’t. As

中文介绍 arthurmensch 主张企业应使用开源模型，闭源模型强制数据留存将侵蚀企业自主权。选择开源是保护业务增长的关键。

How to build a second brain with Fable 5

@EXM7777 · 122.0K 粉丝 · 83.6K 阅 · 518 赞 · 48 转

07/03 22:06

I'm going to show you, step by step, how to turn Fable 5 into a machine that knows your business inside out... and ships outputs that look nothing like what everyone else is getting the tool is a

中文介绍 EXM7777 分步讲解如何将 Fable 5 打造成第二大脑，使其深度理解业务并产出独特成果。教程侧重个性化知识系统构建。

A Field Guide to Fable: Finding Your Unknowns

@trq212 · 299.3K 粉丝 · 40.1K 阅 · 534 赞 · 43 转

07/04 01:43

Working with Claude Fable 5 keeps re-teaching me an old lesson: the map is not the territory. The map, a representation of the work to be done, is my prompts and skills and context, it’s what I give

中文介绍 trq212 分享使用 Claude Fable 5 的感悟：提示词和上下文只是工作地图，不等于实际成果。强调区分表征与实质，帮助找到未知领域。

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 TheVixhal 解析 AI/ML 必备数学知识，提供完整学习路线图和自用资源。从线性代数到概率论，手把手规划学习路径。

🤗 Kernels: Major Updates

Hugging Face Blog · 07/06 08:00

中文介绍 Hugging Face博客宣布Kernels重大更新。

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

DeepMind Blog · 07/03 22:25

中文介绍 Google DeepMind与电影制作公司A24宣布一项独特的研究合作。

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）以关于循环的辩论、AI工程状况报告以及关于下一步构建内容的闭幕主题演讲结束。

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软件主管Andrew Qu解释了其代理框架eve的创建过程，以及为何技能、沙箱和代理可读网站现在至关重要。

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

TLDR AI · 07/03 08:00

中文介绍 Meta推出Watermelon项目，Anthropic与三星在芯片上合作，以及自主研究（autoresearch）的实际应用。

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“代理型网站”，根据用户意图生成页面。在AIEWF上，Carlos Sanchez讨论了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 Lean Six Sigma和业务流程管理（BPM）等框架通过统计严谨和质量控制实现运营秩序，AI正被用于进一步提升运营卓越性。

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讨论了Impeccable、在“循环最大化”时代中的人类判断，以及为何代理仍需要人类引导。

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在聊天机器人和图像生成之外，正在物理基础设施、运营连续性和安全等领域发挥重要作用，例如用于涡轮机的工业场景。

[AINews] not much happened today

Latent Space · 07/02 15:10

another quiet day.

中文介绍 AI新闻简报报道当天没有重大事件发生。

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每日快讯讨论自主研究与人类能动性之间的紧张关系，演讲者捍卫人类理解与控制的软件工厂愿景遇到阻力。

not much happened today

Smol AI News · 07/02 13:44

**Fullstack Code Arena** extends coding agent evaluation to include **databases, API keys, deployments, and structured tool use**, marking a shift to end-to-end app shipping. **LangChain** released **LangSmith** with unified tracing and **OpenWiki** for auto-generated docs, while **LlamaIndex** demo

中文介绍 Fullstack Code Arena扩展了编码代理评估范围，涵盖数据库、API密钥、部署和结构化工具使用。LangChain发布了LangSmith（统一追踪）和OpenWiki（自动文档生成）。

Gemini Flash upgrade ⚡️, Meta AI cloud 🌐, ZCode 👨‍💻

TLDR AI · 07/02 08:00

中文介绍 谷歌Gemini Flash升级，Meta推出AI云服务，以及ZCode项目。

Autoresearch: The feedback loop behind self-improving agents

Latent Space · 07/02 07:52

Introspection co-founder Roland Gavrilescu explains autoresearch, agent “recipes,” self-improving loops, and why humans remain central to the software factory.

中文介绍 Introspection联合创始人Roland Gavrilescu解释了自主研究、代理“配方”、自我改进循环，以及为何人类仍是软件工厂的核心。

How Cursor deploys AI inside the enterprise

Latent Space · 07/02 03:03

Cursor's Pauline Brunet explains how her team of Forward Deployed Engineers help organizations implement agents — essentially setting up software factories.

中文介绍 Cursor的Pauline Brunet解释了其前部署工程师团队如何帮助企业实施代理，即建立软件工厂。