2026-07-16

Weak-to-Strong Generalization via Direct On-Policy Distillation

👍 112

07/08 08:00

Reinforcement learning with verifiable rewards (RLVR) is a powerful recipe for improving language-model reasoning, but it is expensive to repeat on every new strong model because the target model must generate many rollouts during training. As models scale, post-training itself becomes a bottleneck.

中文介绍 针对RLVR在每代新模型上重复训练开销大的问题，提出弱到强泛化的直接在线策略蒸馏方法。利用弱模型生成rollout蒸馏强模型，降低后训练成本，缓解模型缩放时的训练瓶颈。

ABot-N1: Toward a General Visual Language Navigation Foundation Model

👍 84

07/12 00:21

Visual Language Navigation foundation models aim to unify deep reasoning for grounded spatial decisions with broad versatility for diverse embodied tasks. Current approaches typically achieve this integration via monolithic policies that map observations directly to actions, yet they often suffer fr

中文介绍 提出ABot-N1通用视觉语言导航基础模型。不同于单一策略直接映射观察与动作，它统一空间推理与多样化具身任务，克服现有方法在泛化性和深度推理上的局限。

ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory

👍 72

07/11 23:24

Recent VLM and VLA systems have improved robotic perception and action prediction, yet long-horizon embodied agents still require a general runtime layer for reasoning, memory, tool use, verification, and cross-embodiment execution. We present ABot-AgentOS, a general robotic Agent Operating System t

中文介绍 提出ABot-AgentOS通用机器人智能体操作系统，提供推理、记忆、工具使用、验证和跨本体执行运行时层。引入终身多模态记忆，支持长时域任务，弥补VLM/VLA在长期自主性上的不足。

Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

👍 64

07/09 08:00

AI agents have become capable of autonomously completing short, well-specified tasks. However, existing terminal benchmarks largely focus on simple problems that finish within minutes and are evaluated only by their final outcome. This setup overlooks intermediate progress and partial solutions, yie

中文介绍 现有终端基准仅评估最终结果，忽略中间进展。Long-Horizon-Terminal-Bench聚焦长跨度终端任务，采用密集奖励评分，测试智能体在长时间任务中的逐步表现和部分解决方案能力。

Video Generation Models are General-Purpose Vision Learners

👍 59

07/10 08:00

Driven by next-token prediction, NLP shifted from task-specific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we contend that large-scale text-to-video generation serves as a stro

中文介绍 论文主张大规模文本到视频生成可作为计算机视觉的通用催化剂，类似NLP中下一个token预测的作用。通过视频生成预训练，使模型成为通用视觉学习者，无需任务特定设计。

SynthDocBench: Controlled Benchmark for Long-Context Visual Document Understanding

👍 52

07/11 08:00

Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout complexity, modality, and question difficulty, which makes it difficu

中文介绍 现有VLM文档理解基准无法解耦长度、布局、模态、难度等多因素。SynthDocBench提供可控长上下文视觉文档理解基准，通过合成文档独立控制各因素，系统评估模型能力瓶颈。

Scalable Visual Pretraining for Language Intelligence

👍 50

07/10 08:00

The rapid progress of large foundation models has been driven predominantly by pretraining on large-scale text corpora. However, many forms of knowledge are conveyed through visual representations, where figures, typeset equations, and page layouts carry rich information that cannot be faithfully or

中文介绍 语言模型仅从文本预训练难以获取图表、公式、布局中的知识。提出可扩展视觉预训练方法，通过大规模视觉数据预训练吸收隐性知识，增强语言智能的通用理解能力。

Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation

👍 42

07/13 08:00

In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image-generation reinforcement learning. Instead of asking the MLLM to judge a generated image or answer decomposed verification questions, SpectraReward measures

中文介绍 提出SpectraReward，无需训练的奖励函数，将预训练多模态大语言模型转化为文本到图像生成的零样本奖励模型。通过让MLLM“读回”生成图像描述并与文本对齐，实现强化学习微调。

Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

👍 30

07/09 08:00

Visual generators excel at rendering, but they confidently fabricate what they do not know. User requests are unbounded, evolving, and deeply long-tailed: new characters, trending entities, post-cutoff events, and more. This world-knowledge bottleneck is structural: generators are trained on fixed c

中文介绍 视觉生成模型对训练集外概念会“自信编造”。提出通过智能体搜索动态演进知识边界，利用外部知识源补充未知世界知识，解决长尾、新实体、截止后事件等问题。

Trust Region Policy Distillation

👍 26

07/06 08:00

Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust Region Policy Distillation (TOP-D), which transforms the notoriously unstable, high-variance On-Policy Distillation (OPD) into a stable training paradigm by dynamically constructing a proximal teache

中文介绍 提出Trust Region Policy Distillation（TOP-D），将不稳定的在线策略蒸馏转化为稳定训练。通过动态构建近端教师策略，在信任区域内进行蒸馏，降低方差，提升蒸馏效率。

KronQ: LLM Quantization via Kronecker-Factored Hessian

👍 23

07/08 08:00

Post-training quantization (PTQ) is a widely adopted technique for compressing large language models (LLMs) without retraining. Existing second-order PTQ methods, including GPTQ, construct quantization objectives exclusively from input activation statistics, effectively assuming that all output chan

中文介绍 现有二阶PTQ仅基于输入激活统计量化，忽略输出通道影响。KronQ利用Kronecker分解海森矩阵，同时建模输入和输出结构，实现更精确的LLM量化，在低比特下保持性能。

AdvancedMathBench: A Benchmark Suite for Advanced Mathematical Proof Generation and Verification

👍 22

07/14 01:38

Large language models (LLMs) have achieved remarkable performance on high-school and olympiad-style mathematics, yet their capabilities on advanced mathematics remain poorly understood. Existing benchmarks, however, fall short in both scope and evaluation granularity: they provide limited disciplina

中文介绍 现有数学基准局限于中学/奥数级别。AdvancedMathBench提供多学科高级数学证明生成与验证基准，细粒度评估LLM在代数拓扑、数论等高等数学上的推理能力。

Metacognition in LLMs: Foundations, Progress, and Opportunities

👍 17

07/14 01:58

Metacognition is a foundational component of intelligence critical to effective learning, problem solving, decision-making, communication, and more. In recent years, it has become increasingly recognized as a cornerstone of capable, transparent AI systems. Yet while LLMs have made significant progre

中文介绍 综述论文系统探讨大语言模型中的元认知。定义知识监控、策略选择等基础概念，回顾LLM中元认知能力的进展（自我评估、反思），指出不足与未来方向（可靠性、透明度）。

Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

👍 14

07/14 08:00

Coding agents must integrate external tool returns into ongoing reasoning - a capability that standard left-to-right pretraining on code exposes only in its forward direction. We observe that the action-observation-continuation loop of a coding agent is structurally isomorphic to a function call sit

中文介绍 编程智能体需整合工具返回值到推理流，标准从左到右预训练仅正向暴露。提出函数感知的Fill-in-the-Middle（FiM）作为中间训练，对齐执行循环与函数调用结构，提升工具返回利用能力。

Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning

👍 12

07/09 20:17

Fine-tuning LLMs to inject new knowledge faces a critical challenge: LLMs can quickly memorize new facts, yet fail to use them for downstream reasoning tasks. We formalize this failure as the \textbf{Knowing--Using Gap}, characterized by an accuracy gap and a temporal lag between memorization and ge

中文介绍 微调LLM注入新知识时，模型能记忆但无法用于下游推理，形成“知道-使用鸿沟”。通过机制分析发现记忆与泛化在表示空间上的脱节，并提出缓解策略弥合该差距。

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

👍 11

07/09 08:00

In long-horizon tasks, decision-relevant state is often scattered across an expanding trajectory, while the action agent must surface it and act. As trajectories grow, task requirements, environment facts, prior attempts, diagnoses, and open subgoals can be buried in the context window or pushed bey

中文介绍 长程任务中相关信息易超出上下文窗口。提出主动记忆智能体，能主动检索和动态唤起相关记忆，防止关键信息被遗忘，提升长程任务的成功率。

Proxy Exploration and Reusable Guidance: A Modular LLM Post-Training Paradigm via Proxy-Guided Update Signals

👍 10

07/13 20:56

Post-training is essential for refining the domain-specific capabilities of large language models (LLMs), yet existing reward optimization and distribution matching methods tightly couple policy exploration with distribution alignment. This coupling forces expensive exploration directly on the polic

中文介绍 现有后训练方法耦合策略探索与分布对齐，开销高。提出代理探索与可重用指导范式，通过代理模型探索并生成可重用更新信号，解耦探索与对齐，降低后训练成本。

Self-Guided Test-Time Training for Long-Context LLMs

👍 10

07/10 08:00

Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the e

中文介绍 扩展上下文窗口后LLM仍无法有效利用长输入。提出自引导测试时训练（Self-Guided TTT），在推理阶段自我生成监督信号，学习从长上下文中定位和利用关键信息，提升准确性。

Multi-Agent LLMs Fail to Explore Each Other

👍 9

07/13 16:34

Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents fail to do so, often exhibiting myopic and polarized interaction patterns

中文介绍 多智能体系统中LLM代理交互时缺乏探索能力，表现出短视和两极分化模式。实验证明它们无法有效探索彼此策略空间，导致协作效率低下，揭示根本局限性。

MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

👍 7

07/13 23:59

Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation, failing to adapt culture-specific items; 2) inference-time meth

中文介绍 现有LLM道德推理忽视多语言性。MET提出理论引导、文化感知的多语言道德推理方法，构建文化适应的多语言数据集，设计跨文化推理策略，实现一致道德判断。

You just hired a million bad employees.

@gsivulka · 10.2K 粉丝 · 591.2K 阅 · 512 赞 · 99 转

07/15 00:41

AI was supposed to replace human labor. It did the opposite. For the first time in history, humans are cheaper than software. And AI is creating more jobs than it eliminates. Technology has

中文介绍 AI 并未取代人类劳动，反而使人类比软件更便宜，并创造更多就业岗位。作者指出历史上首次出现人类比软件更廉价的局面，打破了技术替代劳动的常规预期。

Making Fable Cheaper Than Opus

@joon_h_lee · 515 粉丝 · 366.3K 阅 · 521 赞 · 43 转

07/14 01:03

We replaced Opus 4.8 with Fable 5, and Devin’s bill went down. Fable 5 costs twice as much per token as Opus 4.8. But when we ran both models on FrontierCode 1.1 using our new Fusion architecture,

中文介绍 通过 Fusion 架构，用 Fable 5 替换 Opus 4.8 后，虽然每 token 价格翻倍，但在 FrontierCode 1.1 上整体成本反而降低。分享了一个模型替换与架构优化的成本节省案例。

The Most Human Technology Ever Made

@illscience · 30.4K 粉丝 · 332.5K 阅 · 539 赞 · 69 转

07/14 01:05

Over the weekend, Mira Murati wrote a phenomenal essay that you should read in full: “The future worth building is human.” I think it’s one of the best essays of the year, because it gets at the

中文介绍 推荐 Mira Murati 的文章《The future worth building is human》，称其为年度最佳，阐述技术应回归人性，强调以人为中心的未来构建。

A Framework for Frontier AI and the Dawning of a New Age

@demishassabis · 1.4M 粉丝 · 98.1K 阅 · 1.5K 赞 · 236 转

07/14 17:10

This is a pivotal moment in human history. Artificial General Intelligence (AGI), a system that exhibits all the cognitive capabilities the brain has, is probably only a few short years away. When we

中文介绍 Demis Hassabis 指出 AGI 可能在几年内到来，这是人类历史转折点，并提出了一个面向前沿 AI 的框架，讨论迈向新时代的路径。

I Brutally Modified My Front-End Design Skill ~ Now My UIs Don’t Look Like AI Crap

@JoePro · 949 粉丝 · 66.8K 阅 · 537 赞 · 29 转

07/14 11:51

To create a custom Claude skill, navigate to Customize > Skills in your Claude interface, click the + button, and select + Create skill

中文介绍 通过 Claude 自定义技能（Customize > Skills）创建专属指令，改进前端设计质量，避免 AI 生成的低质 UI。一个提升设计水准的实用教程。

How I'd make $10 million with AI agents

@gregisenberg · 682.7K 粉丝 · 62.2K 阅 · 578 赞 · 36 转

07/14 02:22

2 markets worth trillions are opening at the same time. I've been through 3 platform shifts living in Silicon Valley building/selling VC-backed companies and I've NEVER seen this. Mobile apps came

中文介绍 作者经历 3 次平台变革，认为 AI 代理带来的两个万亿级市场同时开放是前所未有，并提出用 AI 代理赚取 1000 万美元的具体路径。

2 Hermes Workflows I can't live without

@0xJeff · 80.8K 粉丝 · 61.6K 阅 · 502 赞 · 36 转

07/13 19:33

I’ve been consistently using Hermes everyday for the past 4 months and he’s already a core part of my research & investment workflows. Before Hermes, it was just me relying on Grok, Claude, Gemini for

中文介绍 连续 4 个月使用 Hermes 作为研究和投资核心工具，分享两个离不开的工作流，替代了仅依赖 Grok、Claude、Gemini 的方式，提升效率。

Your AI says it read your files. Did it really?

@conikeec · 7.5K 粉丝 · 1.1K 阅 · 1.0K 赞 · 121 转

07/14 22:39

AI always answers with confidence — even when it skipped half your documents. waggle shows you which files it really opened, so you know whether to trust the answer, or send it back. You hand an AI

中文介绍 工具 waggle 可显示 AI 实际打开了哪些文件，避免 AI 自信回答但跳过半数文档，确保答案可信。解决 AI 文档阅读透明性问题。

What building Shippy taught us about building agents

Hugging Face Blog · 07/16 01:29

中文介绍 Allen AI从构建Shippy代理项目中总结出构建智能代理的关键经验与教训。

Model Routing Is Simple. Until It Isn’t.

Hugging Face Blog · 07/16 01:27

中文介绍 IBM Research探讨模型路由看似简单但实际复杂的挑战，揭示路由决策中的隐藏难点。

Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer

MIT Tech Review AI · 07/16 01:09

OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training it against GPT-Red made the model its m

中文介绍 OpenAI构建了名为GPT-Red的LLM超级黑客，作为陪练伙伴提升其他模型防御网络攻击的能力。公司近期发布旗舰模型GPT-5.6，声称通过GPT-Red训练增强安全性。

The US is advancing AI safety through state and federal action

OpenAI News · 07/15 20:00

OpenAI outlines a “reverse federalism” approach to AI governance, where state laws help build a national framework for safe, democratic AI.

中文介绍 OpenAI提出“反向联邦主义”AI治理方法，通过州法律构建国家层面的安全民主AI框架。

GPT-Red: Unlocking Self-Improvement for Robustness

OpenAI News · 07/15 18:00

Explore GPT-Red, OpenAI’s automated red teaming system that uses self-play to improve AI safety, alignment, and prompt injection robustness.

中文介绍 OpenAI的GPT-Red是自动化红队系统，利用自我对弈提升AI安全性、对齐性及提示注入鲁棒性。

Welcome Inkling by Thinking Machines

Hugging Face Blog · 07/15 08:00

中文介绍 Thinking Machines发布Inkling模型/工具，在Hugging Face博客进行介绍。

Introducing Real World VoiceEQ: Measuring the human quality of voice AI

Hugging Face Blog · 07/15 08:00

中文介绍 介绍Real World VoiceEQ，用于衡量语音AI的人类质量（自然度与表现力）。

DeepSeek IPO plans 📈, Kalshi compute markets ⚡,  Bonsai phone model 📱

TLDR AI · 07/15 08:00

中文介绍 今日AI要闻：DeepSeek计划IPO；Kalshi推出计算市场；Bonsai发布手机模型。

[AINews] not much happened today

Latent Space · 07/15 07:54

a continuation: Codex adding 1M users a day now.

中文介绍 AI新闻简报：Codex目前日增100万用户，延续快速增长趋势。

5 Trends That Defined AI Engineering at World’s Fair 2026

Latent Space · 07/15 07:21

At this year's AIE World’s Fair, AI engineering entered a new phase: building systems around agents, rather than just building with agents.

中文介绍 2026年AI工程世界博览会总结五大趋势，核心是从“用代理构建”转向“围绕代理构建系统”。

How to manage AI investments in the agentic era

OpenAI News · 07/14 18:00

Learn how enterprises can manage AI investments in the agentic era by measuring useful work per dollar, improving efficiency, and scaling high-value workflows.

中文介绍 OpenAI指导企业在代理时代如何管理AI投资：按美元衡量有效工作量、提升效率并扩展高价值工作流。

not much happened today

Smol AI News · 07/14 13:44

**OpenAI's agent products** saw a **2.5x weekly usage growth** driven by **Codex + ChatGPT Work** and demand for **GPT-5.6 Sol**. JetBrains adopted Codex as a recommended agent, while LangChain enhanced tracing and observability across multiple tools. **PrismML released Bonsai 27B**, a compressed va

中文介绍 OpenAI代理产品周使用量增长2.5倍，由Codex和ChatGPT Work及GPT-5.6 Sol需求驱动。JetBrains推荐Codex代理，LangChain增强追踪能力，PrismML发布Bonsai 27模型。

[AINews] Codex usage up >10x in 6 months to 7M users, +1M in the past ~day; did Codex overtake Claude Code??

Latent Space · 07/14 09:22

a quiet day lets us fact check some numbers against the sound of silence of Claude Code reporting...

中文介绍 Codex在6个月内使用量增长超10倍，用户达700万，过去一天新增100万。与Claude Code的对比引发关注。

How sales teams use ChatGPT Work

OpenAI News · 07/14 08:00

See how sales teams can use ChatGPT Work to create pipeline briefs, meeting prep packets, forecast reviews, account plans, and stalled-deal diagnoses from real work inputs.

中文介绍 OpenAI展示销售团队如何利用ChatGPT Work创建管道简报、会议准备包、预测回顾、账户计划和交易诊断。

How data science teams use ChatGPT Work

OpenAI News · 07/14 08:00

See how data science teams can use ChatGPT Work to build root-cause briefs, impact readouts, KPI memos, scoped analyses, and dashboard specs from real work inputs.

中文介绍 OpenAI展示数据科学团队如何利用ChatGPT Work构建根因简报、影响报告、KPI备忘录、范围分析和仪表盘规范。