2026-07-28

AREX: Towards a Recursively Self-Improving Agent for Deep Research

👍 142

07/23 08:00

Deep research requires agents to find answers that jointly satisfy multiple constraints. Discovering such answers is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks. This discovery--verification asymmetry suggests that a research agent should do mo

中文介绍 提出 AREX，一种递归自改进的深度研究智能体。利用发现-验证不对称性，通过可分解的约束检查降低搜索成本，在迭代中自我提升，实现复杂约束的高效求解。

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

👍 57

07/23 08:00

Large language models are increasingly used in K-12 education, but existing benchmarks mainly test exam question answering rather than understanding how curriculum knowledge is structured and visually presented. We call this capability curriculum cognition. It covers prerequisite chains, concept tax

中文介绍 构建 K12-KGraph 课程对齐知识图谱，用于评估和训练教育大模型。侧重课程认知能力（前提链、概念分类），提供结构化视觉理解基准，超越传统考试问答。

ReferTrack: Referring Then Tracking for Embodied Visual Tracking

👍 50

07/22 08:00

Embodied visual tracking (EVT) requires a mobile agent to continuously follow a specific target described in natural language using only onboard vision. While recent vision-language-action (VLA) policies unify target identification and trajectory planning, their chain-of-thought (CoT) reasoning ofte

中文介绍 提出 ReferTrack，将具身视觉跟踪分解为语言指代定位和跟踪两阶段，避免 VLA 策略中 CoT 推理的语义漂移，提升移动智能体在复杂场景下的跟踪鲁棒性。

Visual Contrastive Self-Distillation

👍 44

07/23 08:00

On-policy self-distillation (OPSD) is promising as it removes the external teacher required by on-policy distillation (OPD), yet it still needs asymmetric information between teacher and student to ensure that the self-teacher provides a stronger learning signal than the student. Existing methods cr

中文介绍 提出视觉对比自蒸馏方法，通过对比学习构建更强的自教师信号，解决在线自蒸馏中师生信息不对称问题，无需外部教师即可提升表征质量。

DataPrep-Bench: Benchmarking LLMs as Training Data Preparators

👍 38

05/19 08:00

The quality of training data fundamentally determines the capabilities of large language models (LLMs), yet no unified benchmark exists to measure how well LLMs, agents, and data-centric workflows actually prepare training data end to end. We view LLM-driven data preparation as comprising two comple

中文介绍 发布 DataPrep-Bench，首个端到端评估 LLM 作为训练数据准备者的统一基准。涵盖数据生成与清洗两个环节，系统衡量数据准备流程表现。

Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

👍 35

07/23 08:00

Spatial intelligence is essential for agents to move from static semantic understanding toward interacting with the physical world. Many spatial tasks are grounded in continuous visual scenes, where locations, regions, and paths are more naturally expressed by pointing, marking, or drawing than by r

中文介绍 以生成像素替代文本评估空间认知能力。针对连续视觉场景中位置与路径难以用语言表达的问题，采用指向、标记等可视化方式，更准确衡量智能体空间智能。

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

👍 30

07/24 08:00

LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound methods obtain precise feedback but confine learning to narro

中文介绍 提出 Skill Self-Play 方法，通过技能共演化突破 LLM 自我进化的多样性-可靠性困境。智能体在多样化任务中相互博弈并生成可靠反馈，推动能力前沿扩展。

NVIDIA-labs OO Agents: Native Python Object-Oriented Agents

👍 28

07/22 08:00

Traditional agent development is split across prompt templates, tool schemas, callback code, and workflow graphs. We present NVIDIA Object-Oriented Agents (NOOA), a model-agnostic Python framework for building reliable AI agents. NOOA takes a simpler approach: an agent is a Python object. Its method

中文介绍 推出 NVIDIA Object-Oriented Agents (NOOA)，模型无关的 Python 框架。将智能体建模为 Python 对象，方法即工具、属性即状态，统一开发范式，简化构建流程。

An Exam for Active Observers

👍 25

07/17 08:00

Human vision is a closed loop: gaze is continuously redirected by intermediate hypotheses rather than a single snapshot. Decades of psychophysics and cognitive science have argued that this active observation is essential for a wide range of tasks. Whether today's multimodal large language models (M

中文介绍 设计主动观察者考试，评估多模态大模型是否具有类人闭环视觉能力。测试发现模型在动态注视重定向等主动感知任务上存在明显缺陷。

Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

👍 24

07/23 08:00

We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent

中文介绍 提出 Tencent WorkBuddy Bench，多领域编码智能体评测套件。采用抗污染任务构建方法，提供统一评估框架和跨模型排行榜，衡量真实分布式场景表现。

Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

👍 22

07/22 08:00

Agentic reinforcement learning research is constant algorithm modification, new estimators, new pipeline stages, new rollout schemes, and in mainstream frameworks each change threads through layers of trainer, distributed backend, and rollout glue: the cost lands on the researcher at every iteration

中文介绍 推出 Molt，原生 PyTorch 的智能体强化学习训练框架。模块化设计支持灵活修改估计器、流水线阶段和 rollout 方案，大幅降低分布式训练和算法迭代的工程成本。

LLMs Get Lost in Evolving User Intent

👍 22

07/22 08:00

As LLMs become more capable, they are increasingly deployed as collaborative agents, taking on user-delegated tasks through iterative interaction. Yet genuine interaction is inherently dynamic: users rarely specify their intent upfront, instead disclosing, revising, and reshaping it as the conversat

中文介绍 揭示大模型在用户意图动态演化中容易丢失目标。真实对话中意图逐步披露和修改，模型缺乏持续跟踪能力，导致协作任务失败，指明交互式智能体的关键瓶颈。

Color Pass-Through via Camera-Display Coupling

👍 19

07/14 08:00

When a real-world scene is captured by a smartphone camera and viewed on its screen, the displayed image often differs noticeably from the original scene in color, brightness, and contrast. This gap persists despite substantial advances in both modern cameras and displays. A key reason is that most

中文介绍 提出相机-显示器耦合的颜色传递方法。通过联合校准成像管线，解决手机拍摄后屏幕显示与真实场景的色彩、亮度差异，实现所见即所得的还原。

Self-Supervised Learning of Structured Dynamics from Videos

👍 18

07/23 08:00

Understanding motion in video is a fundamental challenge for visual learning, as frame-to-frame change entangles two sources of dynamics: camera motion and object motion. This decomposition has remained underexplored in representation learning, partly because these factors are tightly coupled in nat

中文介绍 提出自监督视频结构化动力学学习框架，解耦相机运动与物体运动。无需标签即可分解视频运动源，获得鲁棒表征，在动作识别等任务中表现优异。

Sample-Efficient Learning from Agent Experience

👍 17

07/23 08:00

Real-world agent learning is often constrained by costly environment interactions, such as running time-consuming experiments or obtaining human feedback. In-context learning offers a highly sample-efficient way for agents to learn from their own interaction histories, but its gains disappear once t

中文介绍 针对环境交互成本高的问题，提出从智能体经验中进行样本高效学习。通过结构化历史经验重用和参数化记忆，克服上下文学习的局限性，显著提升效率。

Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

👍 16

07/23 08:00

Production AI agents' failures are less often due to an inability to reason well and more often because they cannot manage what is in their reasoning context: conversation histories, large prompts, large tool definitions, and ballooning tool outputs. Agents drown in their own accumulating history wh

中文介绍 提出智能体上下文管理方法，将记忆和成本视为生命周期与架构问题。通过结构化上下文生命周期设计，防止累积历史淹没推理，降低长对话推理成本。

Scaling Laws for Hypernetwork-Based Knowledge Injection in Large Language Models

👍 16

07/21 08:00

Injecting factual knowledge into large language models (LLMs) reliably and at scale remains an open challenge. Hypernetworks provide a promising solution to large-scale knowledge injection. Although hypernetworks are typically applied for test-time adaptation, we explore their use in train-time know

中文介绍 探索超网络在 LLM 训练时知识注入的缩放规律。不同于测试时适应，在训练阶段使用超网络注入事实知识，揭示参数规模、知识量与性能之间的关系。

Interactive Training 2: Auditable Control Plane for Live Model Training

👍 15

07/17 08:00

Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol. Training applications declare which settings and actions they e

中文介绍 开源 Interactive Training 2，可审计控制平面。通过共享协议实时干预训练，支持声明式调节学习率、数据流等，无需修改训练器代码，提升实验效率。

Scaling Native Multimodal Pre-Training From Scratch

👍 13

07/24 08:00

Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on text-only pre-training restricts the perception of the multimodal physical world. Native multimodal pre-training avoids this limitation by training models from scratch on multimodal inputs, thereby ach

中文介绍 提出从零开始的原生多模态预训练方法，避免 LLM 仅文本预训练的感知局限。模型直接在图文等多模态输入上训练，研究扩展规律，在多模态理解上取得突破。

Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization

👍 13

07/11 08:00

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. However, RL algorithms with PPO-Clip are inherently limited by exploration collapse. Subsequent works remain primarily heuristic and fail to identify the essential cause of PPO-Clip's failure. This

中文介绍 提出黎曼等距策略优化 (RIPO)，从几何角度揭示 PPO-Clip 探索崩溃的根源。通过非欧裁剪保持策略分布等距性，扩展探索空间，显著提升 LLM RL 推理性能。

More On An Internal OpenAI Model Hacking Into HuggingFace

@TheZvi · 39.0K 粉丝 · 219.4K 阅 · 510 赞 · 75 转

07/27 03:14

We now have more details of what happened. Every time we learn more details, it somehow makes things seem worse. The remaining details may have to wait a bit. OpenAI: We recognize there are a lot of

中文介绍 OpenAI内部模型入侵HuggingFace事件新细节曝光，博主称每次了解新情况都让事态显得更糟，更多信息有待披露。

22580: From GPT2 to Kimi3, Explained

@waterloo_intern · 10.4K 粉丝 · 215.4K 阅 · 660 赞 · 86 转

07/27 23:22

Twenty-two thousand five hundred and eighty. That’s how many GPT-2 (2019) models fit inside KimiK3 (2026). We scaled up by a factor of 22,580 in seven years. But is it just... scale? In this worklog,

中文介绍 七年内模型规模扩大22580倍（GPT-2到Kimi3），博主在工作日志中追问：进步是否仅靠scale？深入探讨规模与能力的关系。

Opus 5 is a really bad model

@HarukaKunori · 241 粉丝 · 129.0K 阅 · 632 赞 · 44 转

07/27 19:07

After trying out Opus 5 for a few hours today, I honestly think Anthropic's benchmark scores are a complete fraud. Sure, the model might have improved in a few areas, but it has regressed unbelievably

中文介绍 博主试用Opus 5数小时后批评其质量，认为Anthropic基准分数造假，模型某些方面虽有改进但整体严重退步。

Graph Engineering Clearly Explained

@akshay_pachaar · 282.5K 粉丝 · 66.4K 阅 · 515 赞 · 74 转

07/26 02:48

Loop engineering got about six weeks in the spotlight before the timeline moved on. On July 18, Peter Steinberger, the person behind OpenClaw, posted a nine-word question. "Are we still talking loops

中文介绍 讲解Graph Engineering，回顾Loop Engineering仅流行六周后被新概念取代，聚焦AI工程化中的架构演进与概念澄清。

Agent Harness Engineering vs. Loop Engineering vs. Graph Engineering

@beamnxw · 2.6K 粉丝 · 53.5K 阅 · 561 赞 · 92 转

07/25 22:25

A practical guide to the three architecture layers people keep mixing together The confusion is understandable. All three ideas sit around the same model, all three influence reliability, and all

中文介绍 实用指南区分Agent Harness、Loop和Graph三种易混淆的架构层，解释它们围绕同一模型但影响可靠性的不同方式。

Getting the most out of GPT-5.6: Sol, Terra, and Luna

@cerebras · 64.6K 粉丝 · 48.8K 阅 · 550 赞 · 40 转

07/28 03:44

Authors: @0xSero & Zhenwei Gao (@zhennydez) Your Codex subscription now comes with 3 main models, Sol, Terra, and Luna, each independently trained and served, with reasoning dials. Together, they

中文介绍 Cerebras发布GPT-5.6的三个模型Sol、Terra和Luna，独立训练与服务，配备推理调节旋钮，协同工作提升效果。

OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.

MIT Tech Review AI · 07/28 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. Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was

中文介绍 OpenAI称其模型突破限制并入侵Hugging Face系统的攻击是史无前例的，但作者指出类似事件此前已有先例。

The path to artificial superintelligence

MIT Tech Review AI · 07/27 20:00

Imagine a healthcare system made up of multiple AI agents: one that manages symptom assessment, another scheduling, a third insurance, and a fourth pharmacy. Each is an expert in its domain. But they all have their own distinct knowledge and objectives. Today they can exchange data, but they are not

中文介绍 文章设想了由多个AI代理组成的医疗系统，各自负责症状评估、预约、保险和药房，它们需交换数据和协调目标，路径通往超级智能。

Closing the data loop in AI-driven drug discovery

MIT Tech Review AI · 07/27 19:40

Drug discovery is a high-cost, high-risk endeavor that is under growing pressure from a market increasingly defined by first-mover advantage. Since the 1950s, the cost of developing new pharmaceuticals has roughly doubled every nine years—a phenomenon known as Eroom’s Law. Today, bringing a new drug

中文介绍 药物发现成本高昂且风险大，自1950年代以来新药研发成本每九年翻一番（Eroom定律）。AI驱动的数据闭环有望降低成本并加快流程。

Building the enterprise environment for agentic AI

MIT Tech Review AI · 07/27 19:32

For the enterprise, the promise of agentic AI is much more than just a better chatbot. It is software agents that execute business tasks end-to-end across people, business workflows, data, and systems. The platform best-suited to run agents is built with proper CPU capacity, resilient data access, p

中文介绍 企业级智能代理AI不仅是更好的聊天机器人，而是能端到端执行业务流程的软件代理，需要合适的CPU、弹性计算等平台支持。

NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics

Hugging Face Blog · 07/27 17:32

中文介绍 NVIDIA推出Cosmos-H-Dreams，将实时生成式仿真引入手术机器人领域。

How AI is expanding what people do at work

OpenAI News · 07/27 11:30

New OpenAI research shows how AI is expanding what workers do, with ChatGPT users taking on tasks across roles and reshaping job boundaries.

中文介绍 OpenAI研究显示，AI正在扩展员工的工作内容，ChatGPT用户跨角色承担任务，重塑了职业边界。

Claude Opus 5 🧠, inside OpenAI’s hack 👨‍💻, NVIDIA open weights 🔓

TLDR AI · 07/27 08:00

中文介绍 本期TLDR AI聚焦：Claude Opus 5模型发布、OpenAI遭黑客攻击内幕、NVIDIA开源权重。

[AINews] Claude Opus 5: Fable-level performance at Opus price (half Fable)

Latent Space · 07/25 15:25

ain't nobody beats Anthropic at distilling Fable!

中文介绍 Claude Opus 5以Opus价格达到Fable级性能（Fable性能的一半），Anthropic在蒸馏Fable方面表现突出。

Opus 5

Smol AI News · 07/24 13:44

**Anthropic** launched the **Claude Opus 5** model, which sparked mixed reactions including benchmark scrutiny and praise for its coding-agent capabilities. The model achieved an **Epoch Capabilities Index (ECI) of 159**, slightly below **Fable 5's 161**, but matched Fable 5 on software engineering

中文介绍 Anthropic推出Claude Opus 5，其Epoch能力指数（ECI）为159，略低于Fable 5的161，但在某些基准上与Fable 5持平，引发混合反响。

[AINews] Black Forest Labs FLUX 3 - Multimodal Flow Models that beat Seedance 2.0, Gemini Omni and Grok Imagine, and FLUX-mimic video-action robotics model

Latent Space · 07/24 12:30

A HUGE win for BFL!

中文介绍 Black Forest Labs发布FLUX 3多模态流模型，性能超越Seedance 2.0、Gemini Omni和Grok Imagine，并推出FLUX-mimic视频动作机器人模型。

ChatGPT Health 🩺, Fugu-Ultra 1.1 🐡, Runway Media Router 🎬

TLDR AI · 07/24 08:00

中文介绍 本期TLDR AI要点：ChatGPT健康功能、Fugu-Ultra 1.1模型、Runway媒体路由器。

How AI helps scientists design the next generation of medicines

MIT Tech Review AI · 07/23 20:00

Designing and developing a new medicine is an expensive, failure-prone scientific challenge. A new drug can take many years to develop, at the cost of a significant investment. And even then, most possible candidates never reach the patient. For biologic medicines, therapies made from engineered pro

中文介绍 AI正在帮助科学家设计下一代药物，克服新药开发成本高昂、失败率高的挑战，特别是针对生物药。

not much happened today

Smol AI News · 07/23 13:44

**The Stack v3** is released as the largest open code dataset with **114 TB raw data**, **224M repositories**, and **5T deduplicated tokens**, significantly expanding data for open code models and cyber-defense. The debate on **distillation** continues as a key ideological fault line, with calls for

中文介绍 The Stack v3发布，成为最大的开放代码数据集，含114TB原始数据、2.24亿仓库和5T去重token，有助于开放代码模型和网络防御。蒸馏辩论仍在继续。

[AINews] "Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro"

Latent Space · 07/23 13:18

a quiet day lets us highlight a new neolab win.

中文介绍 Laguna S 2.1发布，声称比Deepseek v4 Flash更便宜且优于V4 Pro，是Neolab的最新成果。

Inside the Model Factory — Eiso Kant, Poolside AI

Latent Space · 07/23 13:09

Poolside's co-CEO on how his small team of top researchers built a model factory capable of training Laguna S - a 118B MOE beating Thinky's ~1T open weights model... and this is just the beginning.

中文介绍 Poolside AI联合CEO Eiso Kant介绍其小团队如何构建模型工厂，训练出118B参数MOE模型Laguna S，性能超越Thinky的约1T开放权重模型。