2026-10-09

STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization

👍 100

09/29 08:00

Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors prop

中文介绍 针对线性注意力中循环状态量化导致严重精度下降的问题，论文提出 STEPQuant 量化方法。该方法深入分析 delta 规则下量化误差随时间步与特征维度的传播规律，识别出关键误差位置并实施针对性量化，有效缓解了并发推理时的显存瓶颈。

VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction

👍 53

10/05 08:00

Multimodal Large Language Models (MLLMs) have demonstrated remarkable potential in video understanding, yet their reliance on retrospective summarization and text-centric priors often limits their ability to bridge unobserved causal transitions when applied to Video Event Prediction (VEP). To addres

中文介绍 为解决多模态大模型在视频事件预测中难以建模未观测因果转移的缺陷，提出 VepAgent 框架。该框架将工具调用与强化学习结合，使智能体能主动获取辅助信息并跨越因果转移鸿沟，显著提升了事件预测的因果推理能力与准确度。

UniWAM: Unified World-Action Model

👍 48

10/01 08:00

Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet rema

中文介绍 论文提出统一世界-动作模型 UniWAM，旨在弥合视觉-语言-动作模型缺乏物理规律感知与世界-动作模型缺乏推理能力的缺陷。通过将视频生成的时空先验与视觉语言推理、动作控制统一建模，实现了具备环境动态感知的高性能具身智能控制。

RunningTab: Direct Workspace Interaction with Environment-Side Tabs

👍 34

10/07 08:00

Much knowledge work produces new deliverables from files a workspace already holds, and LLM agents are beginning to take such work over. Through direct corpus interaction, an agent can search and read any of those files from a terminal with no indexing, and producing a deliverable from many of them

中文介绍 论文提出 RunningTab，一种使 LLM 智能体通过环境端标签页直接与工作区交互的机制。智能体无需预先构建检索索引，即可通过终端直接搜索、读取工作区的大量现有文件，并基于多文件上下文高效生成交付成果，显著提升复杂知识任务处理能力。

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

👍 31

10/06 08:00

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may re

中文介绍 针对 RLVR 在所有轨迹奖励相同时难以提供有效相对反馈信号的问题，提出自我反思蒸馏方法。该方法通过在事后对已执行经验进行反思与提炼，转化为行动前的先验预见知识，使智能体在零散或同质奖励下仍能有效学习并改进策略。

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

👍 29

10/07 08:00

General-purpose agents increasingly write code, use tools, and complete complex digital tasks, raising the question of how far these capabilities carry into the physical world. To investigate this, we introduce RobotWorld, a challenging simulation testbed for robot use: turning instructions and obse

中文介绍 论文推出用于评估多模态智能体操作机器人能力的仿真基准 RobotWorld。该基准涵盖多种具身形态与复杂物理任务，测试智能体将指令和多模态观测转化为代码及工具调用的能力，系统评估了通用数字智能体向物理世界迁移的泛化边界。

PhysEvo: Astra Can Act, Let It

👍 26

10/06 08:00

Astra can act, yet reliable manipulation depends on the system through which it observes and controls the world. We introduce PhysEvo, a framework for physical recursive self-improvement (RSI) around a single frozen model. A task agent executes robot tasks; a meta-agent uses the resulting trajectori

中文介绍 论文提出围绕冻结模型构建的物理递归自我改进框架 PhysEvo。该系统包含执行机器人操作的任务智能体，以及基于执行轨迹进行分析与自我改进的元智能体，无需重新训练主干模型即可持续优化机器人的观测与控制系统，提升物理操作鲁棒性。

SWE-Game: Can Coding Agents Build the Games We Want?

👍 25

10/07 08:00

We introduce SWE-Game, a benchmark of 247 tasks grounded in 41 executable reference Godot games spanning 13 gameplay categories in 2D and 3D. Five task types cover development from a brief, implementation from a game design document, skeleton completion, repair of 83 injected-fault cases, and Godot-

中文介绍 论文提出评估代码智能体游戏开发能力的基准 SWE-Game。该基准基于 41 个可运行的 Godot 参考游戏构建了 247 个任务，涵盖 13 种玩法类别与 5 类任务形式，系统衡量了智能体从设计文档实现、骨架补全到故障修复的全流程开发能力。

NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

👍 20

10/03 08:00

Sparse-view novel view synthesis is a central problem in 3D content creation, but diffusion-based approaches remain limited by iterative denoising, making multi-view generation expensive at inference time. We introduce NAMVIS, a diffusion-free framework that reformulates multi-view image synthesis a

中文介绍 针对扩散模型在新视角合成中推理迭代慢、计算成本高的问题，提出无扩散的自回归多视角图像生成框架 NAMVIS。该方法将稀疏视角生成重构为跨尺度自回归任务，在保证视角一致性的同时大幅提升推理效率，加速 3D 内容创作。

Q-Learning with Scalar Adjoint Matching

👍 17

10/07 08:00

Flow policies capture rich and diverse action distributions, and fine-tuning them with off-policy RL to improve beyond the demonstrations has drawn growing interest. However, fine-tuning a flow policy against a learned value function is not trivial, because the policy generates its action over many

中文介绍 论文提出结合标量伴随匹配的 Q-Learning 方法，用于离策略强化学习微调流匹配策略。该方法克服了流模型多步动作生成导致价值函数梯度反传困难的挑战，通过标量伴随系统高效计算策略梯度，成功使流策略超越演示数据的表现。

MIMESIS: Learning User Simulators as Training Environments for Interactive Agents

👍 16

10/07 08:00

Training and evaluating interactive language agents typically requires rich user interactions, yet collecting human feedback is expensive and difficult to scale. Simulated users offer a scalable alternative, but they must both resemble real user behavior and provide useful learning experiences for a

中文介绍 针对交互式语言智能体训练难以大规模收集真实人类反馈的问题，论文提出 MIMESIS 框架。该框架学习高保真用户模拟器作为智能体的训练环境，既能真实还原用户行为，又能为智能体提供富有挑战性的交互训练信号，显著降低微调成本。

From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery

👍 16

10/07 08:00

Test-time scaling (TTS) improves the reasoning capabilities of large language models by allocating additional inference computation. Existing approaches to improving TTS efficiency largely optimize accuracy against one resource dimension at a time, advancing either the accuracy--cost or accuracy--la

中文介绍 论文提出一种摊销智能体策略发现方法，实现个性化的测试时缩放。该方法打破了以往仅单向权衡准确率与成本或延迟的局限，能够自适应发现满足多维资源约束的最优智能体推理策略，在不同用户偏好下最大化大模型的推理效费比。

A self-learning scientific agent for X-ray diffraction

👍 12

10/06 08:00

A central challenge for scientific agents is to turn analytical experience into reusable expertise grounded in physical evidence. Here we introduce Gan Jiang, a self-learning agent for powder X-ray diffraction built on a diffraction-analysis ecosystem we developed: XMatcher, XQueryer, XDecomposer an

中文介绍 论文提出粉末 X 射线衍射自学习科学智能体「干将」。该系统依托配套的衍射分析工具生态，能将分析经验转化为基于物理证据的专业知识，自主进行谱图匹配、解析与相分离，展示了智能体在复杂材料结构分析中的自主科学发现能力。

QuadTok: Quadtree Visual Tokenizer for Autoregressive Image Generation

👍 12

10/07 08:00

We introduce QuadTok, a novel framework for visual tokenization and autoregressive image generation. Compared to traditional approaches using 2D grids or 1D token sequences, we propose a hierarchical quadtree structure, bridging the gap between 2D spatial binding and 1D sequence-level flexibility. T

中文介绍 论文提出用于自回归图像生成的四叉树视觉离散化框架 QuadTok。该框架采用分层四叉树结构替代传统 2D 网格或 1D 序列，既保留了图像的 2D 空间拓扑结构，又具备 1D 序列的生成灵活性，显著提升了自回归生成的建模效率与生成质量。

DLoop: Looped Speculative Decoding

👍 11

10/06 08:00

Speculative decoding accelerates autoregressive generation in large language models. In each drafting stage, a lightweight draft model proposes tokens that the target model subsequently verifies. With increasingly capable draft models, we find that the target model frequently accepts all tokens prod

中文介绍 论文提出循环推测解码框架 DLoop。针对高质量草稿模型生成 token 常被目标模型全部接收的现象，DLoop 允许草稿模型在特定条件下连续多轮起草，突破了原有单轮起草-验证的交替瓶颈，大幅提升了自回归大模型的推测解码加速比。

CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use

👍 10

10/06 08:00

Reconstructing an editable CAD model from a 3D shape remains a challenging engineering task. Existing methods can propose CAD operations, but no single source of proposals works equally well across different part geometries and stages of reconstruction. We introduce CADFather, an autonomous agentic

中文介绍 针对现有方法在复杂 3D 几何重构中单一策略难以泛化的问题，论文提出自主智能体 CADFather。该框架协调多种几何推理工具，在不同重构阶段动态自适应选择最佳 CAD 操作建议，实现了从 3D 点云/网格到可编辑 CAD 模型的高精度自动逆向工程。

Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

👍 10

10/05 08:00

Wrapping an image generation model in an agentic harness can effectively boost Text-to-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually construct and revise prompts, thereby eliciting better images.

中文介绍 论文提出同策略上下文蒸馏方法，将智能体套件中的提示词重写与迭代优化经验直接蒸馏内化至文生图扩散模型的权重中。该方法既摆脱了推理期对复杂智能体流调用的额外依赖，又显著保留了智能体协作带来的高质量图文生成优势。

SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision

👍 8

10/04 08:00

Transcribing music into a human-readable score requires a coherent understanding of rhythm, harmony, melody, and form. Two obstacles limit this goal: annotated recordings are scarce, and accurate local predictions can still produce inconsistent musical sequences. We present SheetSage2, a unified mus

中文介绍 论文提出统一音乐转录系统 SheetSage2，解决从音频生成可读铅纸乐谱时高质量标注匮乏及局部预测不连贯的问题。系统通过合成数据提供全流程监督，联合建模节奏、和弦与旋律，实现了结构连贯且符合音乐规范的高精度端到端自动记谱。

Learning Multimodal Embeddings with Evidence-Aligned Readout

👍 8

09/27 08:00

Multimodal large language models can expose task-relevant evidence through generation, but producing useful evidence does not by itself determine how it enters a retrieval embedding. We study whether the semantic organization of that evidence can also specify where representations are read. To addre

中文介绍 论文研究多模态大模型中检索表征的读取机制，提出证据对齐读出学习方法。通过利用模型生成过程中暴露出的语义组织与任务相关证据，指导检索嵌入向量的抽取位置，从而使生成证据与特征表示精准对齐，显著改善了多模态检索性能。

SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles

👍 7

10/07 08:00

Memory-augmented reinforcement learning strengthens LLM agents' ability to solve complex long-horizon tasks. Skills are one such form of memory, pairing instructions with an applicability condition over task types. However, retaining every skill indiscriminately as the policy improves lets obsolete

中文介绍 论文提出 SkillForge 框架，通过引入动态技能生命周期机制解决强化学习智能体记忆中技能陈旧冗余的问题。系统在策略提升的同时持续评估、演化并剔除失效技能，实现技能库与智能体能力的协同进化，显著增强了长程复杂任务的求解效率。

How to scale your app with an AI UGC Army (Full Workflow)

@jakecastilloooo · 18.7K 粉丝 · 141.1K 阅 · 507 赞 · 40 转

10/08 00:38

AI UGC (when done correctly) is one of the greatest hacks for finding and scaling winning videos for organic and paid social. (FYI I’m giving away 13 prompts throughout this article AND a AI UGC skill

中文介绍 应用推广与投放的 AI UGC 全套实操工作流，详解如何利用 AI 批量生成并测试高转化短视频。作者公开了 13 条核心 Prompt 模板与配套技能库，适合出海或 App 增长团队低成本搭建获客管线。

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 微软 AI 负责人 Suleyman 推介新刊首期访谈，诺贝尔经济学奖得主 Daron Acemoglu 预测未来 10 年 AI 仅能替代人类 5% 的工作，呼吁公众停止替代恐慌，从经济学视角为就业焦虑降温。

The Infinite SaaS Factory

@satyanadella · 9.6M 粉丝 · 82.6K 阅 · 587 赞 · 73 转

10/08 23:09

So much of what we do at work is still spent working around software. We switch between apps, rebuild context each time, and translate what we want to do into the steps each system expects. AI is

中文介绍 微软 CEO 纳德拉探讨企业软件范式转移：员工大量精力耗费在不同 SaaS 工具切换与上下文重建中，AI 将重塑人机交互界面，推动从「人迁就系统」迈向无缝打通所有业务流的「无限 SaaS」新时代。

We reverse engineered ChatGPT Intelligent UI. Here's how it actually works

@rabi_guha · 3.1K 粉丝 · 52.8K 阅 · 622 赞 · 31 转

10/09 00:49

ChatGPT launched Generative UI (finally) and renamed the field over night to Intelligent UI. We had to see how they implemented it. This post is an in depth study of how OpenAI implemented their

中文介绍 针对 OpenAI 刚上线的 ChatGPT 生成式界面（官方定名为 Intelligent UI），作者团队进行了逆向工程拆解，深度解析其底层实现架构与动态组件渲染机制，为前端 AI 原生界面开发提供一手参考。

How to make a cartoon explainer video with Claude

@free_ai_guides · 60.4K 粉丝 · 28.6K 阅 · 501 赞 · 44 转

10/08 01:31

I made a 78-second cartoon video with one prompt and four approvals. Claude wrote the script, recorded the voice, drew three cartoon characters into 28 shots and cut it all together with captions. My

中文介绍 基于 Claude 的低门槛动画制作工作流：仅凭 1 条 Prompt 与 4 次人工确认，端到端生成了一支 78 秒动画短片，由 Claude 统筹完成剧本、配音、28 个分镜与字幕剪辑，极大压缩视频制作周期。

how to use AI agents to run your marketing team (the playbook)

@aashatwt · 16.5K 粉丝 · 24.8K 阅 · 513 赞 · 43 转

10/07 20:05

i haven’t built a million-dollar app yet. but i’ve worked with many of the top AI companies over the last six months, and one thing is obvious: they move so fucking fast. here’s what that speed can

中文介绍 总结过去半年与顶尖 AI 公司合作经验的营销落地手册，拆解头部团队如何配置多 AI Agent 跑通自动化营销流水线，将小团队的内容产出与获客速度拉升数倍，提供了一套可复用的 AI 驱动团队运作方案。

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 针对多目标、易内耗的 ADHD 学习者设计的个人 AI 系统搭建教程。作者利用 Agent 与大模型架构，教你在 1 小时内搭建一套专注力与目标推进系统，帮助兴趣广泛的开发者破除拖延、推进多线任务。

Roundtables: A Conversation With the Creator of AI-Designed Viruses

MIT Tech Review AI · 10/09 08:08

Friday, October 16, 2026 Can AI design new life forms? In 2025, Stanford University PhD student Samuel King came up with a preliminary answer when he used a generative AI model to propose genetic blueprints for microscopic viruses. It isn’t yet an example of AI-generated life, but that could be next

中文介绍 斯坦福大学博士生塞缪尔·金（Samuel King）利用生成式人工智能模型提出了微型病毒的基因蓝图。《麻省理工科技评论》举办圆桌对话，与其探讨人工智能设计新生命形式的可能性及相关进展。

[AINews] not much happened today

Latent Space · 10/09 07:29

a quiet day.

中文介绍 Latent Space 发布今日人工智能行业简报，称本日行业动态较为平淡，未出现重大技术突破或关键产业事件更新。

Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

Latent Space · 10/09 00:27

A special Science pod and Engineering pod crossover.. with Forward Deployed Engineering kicker!

中文介绍 Periodic Labs 的利亚姆·费杜斯（Liam Fedus）与埃金·多古斯·库布克（Ekin Dogus Cubuk）做客播客节目，探讨从半导体到超导体的超级智能合成，以及前向部署工程的实际应用。

How Oracle turns days of work into minutes with ChatGPT and Codex

OpenAI News · 10/09 00:00

Across recruiting, engineering, and operations, Oracle turns specialist knowledge into fast, repeatable workflows with ChatGPT Work and Codex.

中文介绍 甲骨文公司（Oracle）宣布通过引入 ChatGPT Work 与 Codex 工具，在招聘、工程和运营等多个领域将专业知识转化为快速可重复的工作流，成功将数天的工作时长压缩至数分钟。

Pollo AI turns creative ideas into campaigns with OpenAI

OpenAI News · 10/08 20:00

With GPT-5.6, GPT-6 Astra, and GPT‑Image‑2.5, Pollo AI helps creators turn bold ideas into detailed images and cinematic video ads.

中文介绍 创意平台 Pollo AI 接入 OpenAI 的 GPT-5.6、GPT-6 Astra 及 GPT-Image-2.5 模型，协助创作者将创意构想高效转化为高精度图像及电影级视频广告。

LegalOn halves Codex costs while maintaining development speed

OpenAI News · 10/08 20:00

LegalOn cut estimated daily Codex costs by 65% while maintaining development speed. It matched Astra, Sol, and Luna to tasks and managed budgets strategically.

中文介绍 法律科技公司 LegalOn 通过将特定任务与 Astra、Sol 及 Luna 模型精准匹配，并在战略层面优化预算管理，在保持开发速度的同时，将 Codex 每日预估使用成本降低了 65%。

AI breakthroughs in robotics won’t change your life any time soon

MIT Tech Review AI · 10/08 17:00

The story is a collaboration between MIT Technology Review and Aventine, a non-profit research foundation that creates and supports content about how technology and science are changing the way we live. A robot shaped like a human—white with a black head and torso—has been popping up on video feeds.

中文介绍 《麻省理工科技评论》与非营利研究机构 Aventine 联合报道指出，尽管人形机器人近期频繁亮相，但机器人领域的人工智能突破在短期内仍难以真正改变大众的日常生活。

Building a safer path to autonomous industrial AI

MIT Tech Review AI · 10/08 16:17

Industrial AI is entering a new phase. After decades of predictive analytics and other specialized applications, advances in foundation models, physical AI, and agentic AI are making it possible to automate more complex tasks across industrial environments. But unlike AI that operates purely in the

中文介绍 工业人工智能正步入全新发展阶段。随着基础模型、物理人工智能及智能体技术的持续演进，工业环境中的复杂任务正逐步实现自动化，行业同时也在探索更安全的自主化落地路径。

[AINews] Claude Haiku 5.5 — better than GPT-6 Luna at the same pricing

Latent Space · 10/08 15:27

yay small models

中文介绍 Latent Space 发布动态称，Anthropic 旗下 Claude Haiku 5.5 模型在保持同等价格水平的前提下，性能表现已超越 GPT-6 Luna，小型模型竞争力进一步提升。

Disrupting AI-enabled “false front” operations

OpenAI News · 10/08 08:00

OpenAI disrupted two AI-enabled influence operations that used false-front journalists and a think tank to spread geopolitical messaging.

中文介绍 OpenAI 宣布阻断了两起利用人工智能技术展开的虚假阵线影响力行动。涉事行动通过虚构记者身份及虚假智库机构，试图在网络环境中散播特定地缘政治立场信息。

The model that didn't exist, so you made it yourself

Hugging Face Blog · 10/08 08:00

中文介绍 Hugging Face 官方博客发布开发者经验分享，探讨当市场上缺乏契合特定应用场景的现成机器学习模型时，开发者如何从零自主构建与训练专属模型。

ChatGPT Intelligent UI 🧠, Claude Haiku 5.5 ⚡, Grok Bot 3rd party models 🤖

TLDR AI · 10/08 08:00

中文介绍 TLDR AI 科技资讯汇编报道最新行业动态，重点涵盖 ChatGPT 智能用户界面上线、Claude Haiku 5.5 模型发布，以及 Grok Bot 开始支持第三方模型等内容。

Multimodal open d1 decision models for the edge

Hugging Face Blog · 10/08 00:54

中文介绍 LiquidAI 在 Hugging Face 平台推出 open d1 多模态决策模型，该系列模型专门面向边缘计算设备优化，旨在提升终端侧复杂环境下的智能决策处理能力。

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 联合创始人克雷格·麦克拉克（Craig McLuckie）与乔·贝达（Joe Beda）探索将智能体运行框架全面转向云原生架构，以提升智能体在桌面端之外的可靠性。

Introducing Falcon ASR

Hugging Face Blog · 10/07 21:21

中文介绍 阿联酋技术创新研究所（TII）在 Hugging Face 博客上正式发布 Falcon ASR 自动语音识别模型，进一步丰富其 Falcon 开源模型家族在语音处理领域的布局。