2026-10-08

Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

👍 153

10/06 08:00

On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels. In this paper, we examine whether expanding this alignment coverage improves le

中文介绍 跨 tokenizer 的 On-Policy Distillation 需在序列与词表两个层面做师生对齐。论文检验扩大对齐覆盖范围是否真能改善学习，指出收益有限，并将问题重新定位为监督可靠性，为跨词表蒸馏提供新的分析视角。

TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

👍 70

10/06 08:00

Reinforcement learning (RL) for post-training large language models (LLMs) incurs substantial computation and memory overhead during rollout generation, which motivates low-precision rollout for efficient RL training. However, existing FP4 RL methods suffer from a key limitation: they primarily opti

中文介绍 MoE 大模型强化学习后训练的 rollout 生成开销巨大，低精度 FP4 rollout 是提效方向，但现有 FP4 RL 方法存在关键局限。TRACE 提出由 rollout 引导的量化感知训练，将 FP4 支持扩展到 MoE 模型的强化学习训练全流程。

CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?

👍 52

10/06 08:00

Static-analysis checker synthesis requires agents to interpret a defect specification, inspect a repository, implement analyzer-specific logic, and refine the checker through repeated compilation and analysis feedback. Existing coding-agent benchmarks focus on tasks such as patch generation or vulne

中文介绍 现有 coding agent 基准多聚焦补丁生成或漏洞检测。CheckerBench 转而评测长程智能体能否合成静态分析 checker：需理解缺陷规范、检查代码仓库、实现分析器特定逻辑，并依据反复的编译与分析反馈迭代修正。

From Evidence to Action: How Tool-Using Agents Fail

👍 34

10/06 08:00

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and

中文介绍 工具调用智能体对外部状态做出实际改动，但结果正确并不意味着其动作有事先建立的证据支撑。论文系统研究从「证据」到「行动」这条链条在何处断裂，覆盖是否行动的决策、单步执行及后续环节的失败模式。

Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing

👍 24

09/29 08:00

Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes. We propose self-compensating VLA, a deployment-time adaptation meth

中文介绍 机器人机械磨损、负载变化等执行误差会使 VLA 策略的实际动作偏离指令并导致失败。论文提出自补偿 VLA，一种部署时的自适应方法，让策略在动作执行偏差下仍能完成任务，并配套压力测试评估鲁棒性。

World Action Learning via Interaction-Centric Spectral Latent Guidance

👍 24

10/02 08:00

Learning general-purpose robot policies requires large-scale real-world interaction data, yet collecting robot demonstrations remains expensive and difficult to scale. Egocentric videos offer abundant human interaction experience with task-relevant semantics for robotic manipulation, but direct tran

中文介绍 通用机器人策略学习依赖大规模真机交互数据，采集成本高且难以扩展；第一人称视频虽有丰富的人类操作经验，却难以直接迁移。论文提出以交互为中心的谱域隐空间引导方法，将视频中的交互语义用于世界与动作学习。

HuatuoGPT-3: RL-Only Domain Adaptation from Base Models

👍 21

10/05 08:00

Domain adaptation aims to turn a general-purpose large language model (LLM) into an expert for a target domain. While the dominant SFT+RL pipeline offers a convenient cold start, it may reduce exploration diversity and introduces additional complexity through multi-stage optimization. These limitati

中文介绍 领域适应通常采用 SFT+RL 两阶段流程，冷启动虽便利，却会压缩探索多样性并增加多阶段优化的复杂度。HuatuoGPT-3 提出直接从基座模型出发的纯强化学习领域适应方案，省去 SFT 阶段完成医疗领域专家化。

UNREAL: Unifying Retrieval and Long-Context with a Single Model

👍 19

10/06 08:00

Long-context inference and Retrieval-Augmented Generation (RAG) handle evidence selection at vastly different scales, from a single long prompt to an entire corpus. We ask whether a single model-internal mechanism can select evidence across this range. We introduce UNifying REtrieval And Long-Contex

中文介绍 长上下文推理在单个长提示内选证据，RAG 则要在整个语料库中检索，二者尺度差异巨大。UNREAL 提出统一的模型内部机制，用同一个模型同时覆盖两种尺度的证据选择，实现检索与长上下文的融合。

AGO AI Quality Gate: Evidence-First Release Decisions for Retrieval-Augmented Generation

👍 16

10/01 08:00

Enterprises adopting retrieval-augmented generation (RAG) face a recurring operational decision: promote, revise, or block a system version. The evidence is incomplete and the metrics come from fallible LLM judges. We report on AGO AI Quality Gate (AGO), an evidence-first quality-gate framework depl

中文介绍 企业部署 RAG 时常需决定版本是发布、修订还是拦截，但可用证据不完整，指标又来自可能出错的 LLM 评判。AGO AI Quality Gate 提出「证据优先」的质量门框架，为 RAG 版本放行提供可复现、可审计的判断依据。

MiniCorp: The Last Mile of the AI Agent Firm

👍 15

10/05 08:00

The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. Historical archives are also frequently incomplete and record only what actu

中文介绍 企业级通用智能体需要长期纵向的企业数据来训练与适配，但这类数据稀缺、获取昂贵且受隐私约束，历史档案也往往不完整，只记录实际发生的事。MiniCorp 围绕「自运行公司」这一目标探讨数据与智能体构建路径。

DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation

👍 12

10/03 08:00

On-policy distillation (OPD) has emerged as a widely used paradigm for post-training large language models, reducing the train--test mismatch of conventional distillation by supervising the student on its own generated trajectories. However, existing OPD objectives remain largely token-local and out

中文介绍 现有 On-Policy Distillation 目标大多停留在 token 局部层面，未考虑轨迹难度差异。DiffGate 提出难度门控的教师指导机制，按样本难度自适应调节教师监督强度，从而更有效地利用学生自生成轨迹完成后训练。

Rationale-Guided Policy Optimization: Learning to Reason with Adaptive Rationale Scaffolding

👍 11

10/05 08:00

On-policy reinforcement learning has become a central paradigm for improving the reasoning abilities of large language models. However, its effectiveness is often limited by reward sparsity: when a model fails to discover correct trajectories for difficult problems, the optimization process receives

中文介绍 on-policy 强化学习是提升 LLM 推理能力的主流范式，但奖励稀疏使其效果受限：模型难以发现正确轨迹时，优化几乎得不到信号。论文提出 Rationale-Guided Policy Optimization，用自适应理由脚手架引导推理。

SlimWise: Decoupling Expert Pruning Across Prefill and Decode for Efficient MoE Serving

👍 10

09/28 08:00

Mixture-of-experts (MoE) models activate few experts per token, yet batched decoding can access nearly the entire expert pool, making expert-weight traffic a major bottleneck. Expert pruning reduces this traffic, but conventional approaches also prune compute-bound prefill, sacrificing model quality

中文介绍 MoE 模型批量解码时几乎会访问整个专家池，专家权重传输成为主要瓶颈。传统专家剪枝会连带裁剪计算密集的 prefill 阶段而损害质量，SlimWise 将 prefill 与 decode 的专家剪枝解耦，仅针对解码路径优化。

Adaptive Latent Capacity for World Models

👍 10

09/26 08:00

We introduce Adaptive LeWorldModel (ALeWM), a world model based on a joint-embedding predictive architecture (JEPA) that learns to concentrate predictive information in compact prefixes of a wide latent representation. To encourage this ordering, ALeWM learns a sequence-conditioned distribution over

中文介绍 提出 ALeWM（Adaptive LeWorldModel），一种基于联合嵌入预测架构 JEPA 的世界模型。它学习序列条件下的分布，促使预测信息集中到宽隐表示的前缀中，从而实现隐容量的自适应分配，提升世界模型的预测效率。

NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale

👍 10

10/06 08:00

Agentic reinforcement learning (RL) disaggregates training from rollout, so each policy update must reach the rollout clusters before the next batch. Transferring a full 1T checkpoint for such weight synchronization (refit) takes 87.5 min between two AWS regions. Measurements of BF16 training show t

中文介绍 万亿参数规模的 agentic RL 将训练与 rollout 分离，每次策略更新都须把完整 checkpoint 同步至 rollout 集群，跨 AWS 区域传输一次需 87.5 分钟。NeMo-DCR 采用位精确的增量压缩 refit，显著降低同步开销。

DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks

👍 9

10/06 08:00

LLM agents often lack the operational knowledge to act reliably in new environments, as they must discover specific tool behaviors or environment conventions on their own. Without memory of past attempts, they repeat the same mistakes across tasks, leading to more task failures and longer trajectori

中文介绍 LLM 智能体进入新环境时常缺乏工具行为与环境约定等操作知识，没有历史记忆便会在不同任务上重复犯错，导致失败增多、轨迹变长。DAEDALUS 让智能体自主生成任务并从中自举构建记忆，提升后续任务表现。

Sherpa: Teaching LLMs to Teach Adaptively

👍 6

10/06 08:00

Large language models (LLMs) have become increasingly capable problem solvers, but being able to solve a problem is not the same as being able to teach it. Existing approaches to training LLMs as teachers rely on demonstrations, preference data, or predefined pedagogical criteria that specify what g

中文介绍 能解题并不等于能教学。现有将 LLM 训练为教师的方法依赖示范、偏好数据或预设教学标准，规定「怎样教」。Sherpa 转而让模型学习自适应教学，依据学生状态动态调整教学策略，而非套用固定的教学准则。

Learning to Read the Contextual Tokens in Diffusion Transformers

👍 6

10/05 08:00

Multimodal Diffusion Transformers (MM-DiTs) jointly process visual and textual representations throughout generation. These models repeatedly update the text tokens through multimodal attention, forming dynamic contextual tokens whose function is not well understood. In this work, we introduce a fra

中文介绍 多模态 Diffusion Transformer 在生成全程联合处理视觉与文本表示，并通过多模态注意力反复更新文本 token，形成功能尚不明确的动态上下文 token。论文提出分析框架刻画其作用，加深对这类模型内部机制的理解。

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model

👍 6

10/07 01:56

Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current def

中文介绍 网页智能体需读取第三方撰写的页面，页面中植入的指令可能将其带离用户目标，但页面又承载任务所需的值与控件，无法简单忽略。AdvSim2Real 在 Web 世界模型中以自适应提示注入训练智能体，提升其抗注入攻击能力。

Understanding and Enhancing Backdoor Persistency in LLM Agent Post-Training

👍 6

10/05 08:00

Developers can build LLM agents by adapting third-party models through benign post-training. We study a supply-chain threat in which an attacker supplies a model with a backdoor: hidden behavior that produces malicious outputs when a particular input pattern appears. Focusing on software-engineering

中文介绍 研究 LLM 智能体后训练中的供应链后门威胁：攻击者提供的模型在特定输入模式下触发隐藏恶意行为。论文聚焦软件工程场景，分析后门在良性后训练过程中的持续机制，并探索增强其后门持久性的方法，为防御提供依据。

He Who Controls the Compute Controls the Universe

@stevenmarkryan · 458.8K 粉丝 · 124.9K 阅 · 541 赞 · 84 转

10/05 13:25

Incredible things are happening in AI. What’s more incredible is that it’s still EXTREMELY early. As in, Day 0.000. Things are only going to get crazier from here. In case you’ve been living under a

中文介绍 博主谈 AI 时代的算力权力格局，认为掌控算力者将掌控一切，并强调眼下仍处于「Day 0.000」的极早期阶段，接下来的变化只会更加剧烈。偏宏观叙事与个人判断，未涉及具体工具或数据。

Every language learner gets a tutor now

@TryLiveAvatar · 1.1K 粉丝 · 123.4K 阅 · 692 赞 · 188 转

10/06 02:45

The best way to learn a language was always a patient tutor. For the first time, everyone can have one. Here's what that changes for learners, teachers and schools. The hardest part is speaking Ask

中文介绍 面向语言学习者的 AI 导师产品愿景：人人都能拥有一个有耐心的私人外教，最难也最关键的一环是「开口说」，并讨论这对学习者、教师和学校分别意味着什么。属于 AI 口语陪练方向的发布式内容。

AI won't take your job anytime soon. In 10 yrs, only 5% of what humans do will be replaced by AI

@mustafasuleyman · 1.2M 粉丝 · 99.2K 阅 · 587 赞 · 104 转

10/06 22:16

This is the prediction Nobel laureate Daron Acemoglu makes in the first issue of The Humanist Review, our new magazine exploring the future of AI, published by MAI. He argues we need to stop

中文介绍 援引诺贝尔经济学奖得主 Daron Acemoglu 在 MAI 新杂志 The Humanist Review 首期的预测：未来十年 AI 只会接管人类工作内容的约 5%，不会很快抢走饭碗，并主张停止过度恐慌、理性看待就业冲击。

If you have too many interests, build this AI system in 1 hour

@Hesamation · 90.6K 粉丝 · 21.4K 阅 · 618 赞 · 51 转

10/06 17:46

if you have an ADHD brain, you probably have six different goals, you get overwhelmed, and you miss days without working on any one of them. that's me. I love agents, harnesses, RL, LLM architecture,

中文介绍 针对兴趣过多、目标分散、常常几天不动手的 ADHD 型人群，用 1 小时搭一套 AI 系统来同时管理六个目标。博主自述关注 agents、harness、RL、LLM 架构，属于个人工作流搭建分享。

Multimodal open d1 decision models for the edge

Hugging Face Blog · 10/08 00:54

中文介绍 Liquid AI 在 Hugging Face 博客发布 open d1 系列多模态决策模型，面向边缘设备场景，官方未提供更多细节说明。

Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?

Latent Space · 10/07 22:10

Kubernetes co-creators Craig McLuckie and Joe Beda aim to bring agent harnesses fully into the cloud.

中文介绍 Kubernetes 联合创始人 Craig McLuckie 与 Joe Beda 提出新项目，目标是把智能体运行框架完整搬进云端，探讨云原生架构能否让智能体在桌面端之外仍保持可靠。

One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

Hugging Face Blog · 10/07 20:45

中文介绍 英伟达在 Hugging Face 博客介绍，通过微调 Nemotron 模型家族，在国际信息学奥林匹克（IOI）与国际数学奥林匹克（IMO）两项赛事中均取得金牌级成绩。

Helping teens learn, plan, and shape the future of AI

OpenAI News · 10/07 20:00

College Planner is coming to ChatGPT for Teens to help students manage college applications, alongside new flashcards, quizzes, and a teen AI council.

中文介绍 OpenAI 宣布 ChatGPT for Teens 将上线 College Planner，帮助学生管理大学申请，同时推出闪卡与测验功能，并设立青少年 AI 委员会参与相关议题。

Radisson Hotel Group brings hotel discovery into ChatGPT

OpenAI News · 10/07 15:00

Radisson partnered with Accenture to build a ChatGPT plugin using OpenAI technology, helping travelers find, compare, and book hotels while planning their trips.

中文介绍 丽笙酒店集团与埃森哲合作，基于 OpenAI 技术开发 ChatGPT 插件，帮助旅客在规划行程时查找、比较并预订酒店。

[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

Latent Space · 10/07 12:55

Our head hurts.

中文介绍 Latent Space 报道称，OpenAI 发布 722 篇数学论文，解决了 500 个顶级未解数学问题中的 90 个，并称这是数学百余年来「最重要的时刻」。

GPT-6 and Intelligent UI for everyone

OpenAI News · 10/07 08:00

GPT‑6 is rolling out globally in ChatGPT with Intelligent UI, delivering faster responses with visuals and interactive experiences you can explore and use directly.

中文介绍 OpenAI 宣布 GPT-6 在全球范围内向 ChatGPT 用户推出，并引入 Intelligent UI，提供更快的响应速度以及可直接探索和使用的可视化交互体验。

Mistral Large 4 🧠, OpenAI Decisions API ❓, Nano Banana 2.1 🍌

TLDR AI · 10/07 08:00

中文介绍 TLDR AI 汇总当日动态：Mistral 发布 Large 4 模型，OpenAI 推出 Decisions API，另提及 Nano Banana 2.1 相关更新。

EmbeddingGemma 2: an open, lightweight multimodal embedding model

DeepMind Blog · 10/07 03:57

中文介绍 DeepMind 发布 EmbeddingGemma 2，定位为开源、轻量级的多模态嵌入模型，具体性能与参数细节未在摘要中披露。

Atlassian and OpenAI expand partnership to turn enterprise knowledge into action

OpenAI News · 10/07 00:00

Atlassian and OpenAI are expanding their partnership to connect frontier models with enterprise knowledge and help teams plan, build, and deliver work.

中文介绍 Atlassian 与 OpenAI 扩大合作，将前沿模型与企业知识库相连接，帮助团队完成规划、构建与交付等协作工作。

How Jump Trading is scaling quant research with ChatGPT

OpenAI News · 10/06 20:00

Jump Trading uses OpenAI to expand quantitative research. See how longer-running AI workflows combine multiple data sources with human review.

中文介绍 量化交易公司 Jump Trading 使用 OpenAI 技术扩展量化研究，做法是让长时间运行的 AI 工作流整合多种数据源，并结合人工审核。

Sharing AI progress in mathematics

OpenAI News · 10/06 20:00

OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub.

中文介绍 OpenAI 公布其内部前沿模型在若干公开数学问题上的新结果，并在 GitHub 上分享 Lean 证明形式化文件及相关研究细节。

Advancing computer use with Ironclad

OpenAI News · 10/06 18:00

Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.

中文介绍 OpenAI 与 Ironclad 合作，围绕复杂的合同工作流训练和评估 AI 智能体，以推进面向专业工作的计算机操作能力。

[AINews] Reflection Beam - 501B-A23B American Open Model

Latent Space · 10/06 14:28

A small win for US open source

中文介绍 Latent Space 报道，Reflection 发布 Beam 模型，参数规模为 501B、激活 23B，属美国开源模型，被评价为美国开源生态的一次小胜利。

Instinct group chats 💬, Reflection’s 501B model 🧠, OpenAI text watermarks 🏷️

TLDR AI · 10/06 08:00

中文介绍 TLDR AI 汇总当日动态：Instinct 推出群聊功能，Reflection 发布 501B 模型，OpenAI 则涉及文本水印相关进展。