2026-09-28

Training Object Permanence in World Models

👍 203

09/23 08:00

Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them ideal candidates for building human-like physical intelligence. Do video

中文介绍 针对视频生成类 world model 是否具备客体永久性与实体性这一人类认知先验，论文提出训练方法让 world model 掌握客体永久性。核心是在视频生成训练中引入遮挡与物体消失-重现场景，使模型学会追踪被遮挡后仍存在的物体。结果显示模型在遮挡与重现任务上的生成一致性显著提升，为构建类人物理智能提供路径。

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

👍 84

09/22 08:00

Long-term conversational memory in multi-party settings requires more than retrieving relevant content from long-term conversations: it must distinguish who said what, whom each statement concerns, how individuals perceive one another, what information is shared by the group, and how states change o

中文介绍 多轮多方对话的长期记忆不仅需检索相关内容，还要区分谁说了什么、涉及谁、个体间如何认知、群体共享信息及状态演变。论文提出 SpeakerMem-R1，以说话人为中心的双轨记忆机制，将内容记忆与说话人动态状态记忆分离并融合。该方法在多方对话记忆任务上提升了角色与状态追踪的准确性。

Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs

👍 75

09/24 08:00

While Large Language Models (LLMs) rely on highly non-linear components, in this work we demonstrate that they exhibit fundamental linearity: when inputs from distinct text streams are linearly combined, the model outputs a superposition of the individual next-token distributions. We term this the S

中文介绍 论文发现 LLM 尽管含高度非线性组件，却表现出基础线性性：将不同文本流的输入线性组合后，模型输出为各自下一 token 分布的叠加，作者称之为「叠加性」。该现象为理解 Transformer 的内部表征与多流推理提供线性视角，也提示可借此实现同时处理多条思路的能力。

Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World

👍 48

09/19 08:00

Spatial reasoning is essential for vision-language models (VLMs) to understand and act in the physical world. Reasoning in dynamic environments requires VLMs to perceive local state transitions caused by object motion and viewpoint changes and integrate them over long trajectories to maintain an upd

中文介绍 针对 VLM 在动态环境中空间推理需感知局部状态变化并跨长轨迹更新的问题，论文提出 Spatial-Interactor，让模型通过与可观测物理世界交互来学习空间推理。方法通过交互反馈整合物体运动与视角变化，维持更新后的空间状态。实验表明其在动态空间推理任务上优于静态感知基线。

HappyWorld-Bench

👍 43

09/21 08:00

Evaluating world models requires assessing both the quality of the worlds they generate and their consistency and responsiveness under exploration, interaction, and modification. We introduce HappyWorld-Bench, a comprehensive benchmark that evaluates whether generated worlds remain reliable as agent

中文介绍 评估 world model 不仅要看生成世界质量，还需考察其在探索、交互与修改下的可靠性与响应性。论文提出 HappyWorld-Bench 综合基准，把生成世界作为 agent 环境，检验其在被操作后是否保持一致。该基准为 world model 的交互式评测提供统一协议与可复现指标。

The Past Frames the Future: Memory for Autoregressive Video Generation

👍 39

09/23 08:00

Advances in generative models have improved video fidelity, enabling long-horizon generation, interactive world modeling, and evolving visual environments. Autoregressive (AR) video generation extends visual sequences through causal rollouts. However, a fundamental bottleneck emerges: as the generat

中文介绍 自回归视频生成可支持长时程生成与交互式世界建模，但随序列延长，因果 rollout 面临记忆瓶颈：早期帧信息被逐渐稀释。论文提出以历史帧构建记忆机制来约束未来生成，使模型在长视频生成中保持一致性与时序连贯，缓解长程漂移问题。

Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents

👍 35

09/23 08:00

Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time: once a task is completed, its trajectory is distilled into a fixed artifact, such as a reflection, workflow, skill, or reasoning strategy, that is later retrieved by sim

中文介绍 现有 agent 记忆系统多在写入时固化轨迹为反思、工作流或技能等静态产物，后续按相似度检索，难以适配任务。论文提出 Just-in-Time Memory，改为在任务执行过程中动态整理任务自适应记忆。该方法让 LLM agent 在需要时筛选与重组经验，提升跨任务复用效果。

OmniEcho: Spatial Audio Understanding for Embodied Agents

👍 22

09/20 08:00

Humans can effortlessly localize the direction of a sound source and integrate it with visual cues for reasoning, yet this remains challenging for embodied agents. In particular, it is still unclear how to effectively evaluate and model spatial audio understanding in embodied settings. To address th

中文介绍 人类能轻松定位声源方向并结合视觉推理，而具身 agent 尚缺空间音频理解能力，且缺乏相关评测。论文提出 OmniEcho，面向具身场景的空间音频理解基准与建模方法，将声音方位与视觉线索对齐用于推理。工作填补了具身 agent 空间听觉评测的空白。

Agent-Editing World Model: Rethinking World Modeling for LLM Agents

👍 19

09/23 08:00

Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool res

中文介绍 现有语言 world model 常预测环境观测，但重建高熵、依赖执行细节的工具返回代价高且噪声大。论文提出 Agent-Editing World Model，将世界建模转为编辑式抽象，预测状态变化的高层编辑而非原始观测。该范式降低预测熵并提升 LLM agent 长程任务表现。

Rufus-Air: An Open LLM Post-Training Recipe

👍 17

09/24 08:00

Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure

中文介绍 论文公开 Rufus-Air，一个在 GLM-4.5-Air-Base（106B-A12B）上可复现的后训练配方，由 SFT、Reasoning RL、Coding RL、指令跟随 RL、通用 Agent、Coding Agent、Search Agent 与 RLHF 八阶段串行组成。报告详述数据、奖励设计、基础设施与消融，为开源 LLM 后训练提供完整参考。

PACT: From Credit Assignment to Critic Alignment

👍 16

09/22 08:00

Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclear. We formulate three regularity conditions, namely Completen

中文介绍 RL 已成为 LLM 后训练核心，但 token 级信用分配缺乏公认数学定义，与常用训练信号关系不清。论文提出完整性、一致性等三项正则条件，将信用分配问题形式化，并据此提出 PACT 从信用分配走向 critic 对齐，为理解与设计 RL 训练信号提供理论框架。

Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models

👍 15

09/22 08:00

The rapid capability gains of frontier language models are widely attributed to improved reasoning abilities, yet this cannot be verified as raw CoT traces in closed-source systems are hidden. By registering a simple custom tool through a standard API feature, we induce frontier models to externaliz

中文介绍 前沿模型能力提升常归因于推理增强，但闭源系统隐藏原始 CoT 难以验证。论文通过标准 API 注册简单自定义工具，诱导前沿模型将隐藏 CoT 外化。借此提取并刻画其隐藏推理特征，发现模型能力强劲却推理精简，为闭源模型的推理机制研究提供新途径。

Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms

👍 12

09/23 08:00

Language-model agents increasingly face long-horizon tasks with evolving state, interdependent decisions, and delayed outcomes. Scaling their training requires diverse agentic environments, dependable outcome signals, and low extension cost. Existing generation pipelines commonly construct an enviro

中文介绍 训练长时程 agent 需多样、可靠且低扩展成本的环境。论文提出可验证隐藏动态玩法，从已解机制自动生成 agentic RL 环境，让环境状态演化可验证且结果信号可靠。该方法以低扩展成本批量产出高质量 RL 训练环境，提升 agent 长程决策学习效率。

Learning to Discover Interesting Mathematics

👍 11

09/23 08:00

Recently, Large Language Models (LLMs) have been increasingly able to solve advanced mathematical problems, including many that have been open for decades. This opens the door to expansion of mathematical knowledge at unprecedented scale. Yet, while LLMs may be able to conjecture and prove more and

中文介绍 LLM 已能解决诸多长期未解的数学问题，但关键在于如何发现「有趣」的数学而非仅验证或证明。论文提出让 LLM 学习发现有趣数学的方法，通过训练或搜索机制筛选有价值猜想与问题。工作为以 LLM 规模化扩展数学知识提供方向。

Coding Agents for Generalized Task and Motion Planning Problems

👍 11

09/24 08:00

Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem insta

中文介绍 任务与运动规划（TAMP）即使全观测、对象化状态下仍因离散决策与几何、运动学、动力学约束耦合而困难。论文提出用编码 agent 求解泛化 TAMP，利用问题实例间规律生成可执行规划代码。方法在泛化场景中减少对手工几何启发式的依赖，提升求解通用性。

Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

👍 11

09/24 08:00

The rapid progression of large language models is extending AI from passive content generation into the active workflows of engineering and scientific discovery. This shift raises a compelling question: can AI be both the object of development and an active participant in building next-generation AI

中文介绍 论文提出 Qwen-Planner-Agent，一个闭环 AI-for-AI 框架，用于真实移动端 planner agent 的构建与迭代。框架让 AI 同时作为开发对象与参与者，通过闭环反馈自动优化规划策略。工作探索了 LLM 在端侧规划任务中自我改进与落地的可行性。

PackLab: A Comprehensive Framework for Developing, Training, and Evaluating MLLMs in Robotic Bin Packing

👍 11

09/20 08:00

Robotic bin packing requires long-horizon sequential decision-making, as each object placement affects the available space for subsequent packing. Existing methods primarily rely on hand-crafted geometric heuristics that optimize predefined objectives or reinforcement learning policies learned throu

中文介绍 机器人装箱需长时程序列决策，现有方法多依赖手工几何启发式或强化学习策略。论文提出 PackLab，一个用于开发、训练与评估 MLLM 完成机器人装箱的综合框架，覆盖数据、训练与评测。该框架为多模态模型在装箱任务中的长程规划能力提供统一基准。

MemBodied: Recurrent Associative Memory for Vision-Language-Action Models

👍 11

09/23 08:00

Vision-Language-Action models provide a strong foundation for general-purpose robot control, yet a vast majority of policies do not preserve and leverage episode-level information beyond the current observation. This limitation is consequential in history-dependent manipulation tasks that depend on

中文介绍 多数 Vision-Language-Action 策略仅依赖当前观测，不保留 episode 级历史信息，在依赖历史的操作任务中受限。论文提出 MemBodied，为 VLA 模型引入循环联想记忆，持续保存并复用既往观测与动作。该方法提升历史依赖操作任务的成功率与长程一致性。

Hunyuan-A13B Technical Report

👍 10

09/23 08:00

We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rig

中文介绍 论文发布 Hunyuan-A13B，一个基于 MoE 的开源大语言模型，总参数 80B、推理仅激活 13B，兼顾能力、计算效率与部署成本。模型经大规模预训练与后训练优化。报告给出架构、训练数据与评测结果，展示其在相近激活规模模型中的竞争力。

IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

👍 9

09/24 08:00

Deep search requires LLM agents to decompose complex queries, search for evidence, and synthesize grounded answers, yet existing ReAct-style agents suffer from two limitations: role coupling, where one policy must handle planning, evidence use, and synthesis; and context accumulation, where growing

中文介绍 深度搜索要求 LLM agent 分解查询、检索证据并合成有据回答，但 ReAct 式 agent 存在角色耦合与上下文累积两大局限。论文提出 IterSynth，以角色解耦的迭代合成框架将规划、证据使用与合成分离。方法缓解上下文膨胀，提升深度搜索的证据利用与答案质量。

The New Copilot

@jacobandreou · 7.8K 粉丝 · 521.0K 阅 · 500 赞 · 77 转

09/25 21:02

The bottleneck is no longer intelligence. Over the last six months, intelligence has accelerated dramatically. But smarter models are not enough for everyone to benefit from novel intelligence. New

中文介绍 作者提出「瓶颈已不再是智能」——过去半年模型智力飙升，但光有更强的模型并不能让所有人受益。帖子聚焦「新 Copilot」这一话题，探讨如何把前沿模型能力真正转化为人人可用的工具，属于产品思路层面的观察，而非具体教程。

We're open sourcing the company brain, Here's how we designed the multi-player harness

@DhravyaShah · 63.2K 粉丝 · 112.4K 阅 · 608 赞 · 45 转

09/26 10:08

It's been just 2 weeks since we discontinued the company brain harness. We had to offboard tons of people to other products, but many of our customers, and others suggested us to open source our work

中文介绍 Dhravya 宣布开源自家「company brain」项目。距团队下线该 harness 仅两周，期间不得不把大量人员调往其他产品线，但客户和社区都建议开源。推文重点讲多人协作式 harness 的设计思路，含金量在于一个真实生产系统的架构复盘。

Why AI is booming, but productivity isn't

@chamath · 2.4M 粉丝 · 97.8K 阅 · 551 赞 · 44 转

09/26 02:06

In June, I wrote that vibe coding was dead and that ROI-driven analysis of AI was about to go from a nice-to-have to a necessity. Here's what AI ROI is and why it's tricky to see in GDP numbers so

中文介绍 Chamath 延续其 6 月的判断——vibe coding 已死、AI 的 ROI 分析将从可选项变成刚需。他解释什么叫 AI ROI，以及为什么它难以在 GDP 数据里体现出来。属于宏观视角的泼冷水式分析，提醒别把模型能力增长直接等同于生产率提升。

Prescriptions for Prosperity in the Digital Economy

@saylor · 5.2M 粉丝 · 96.2K 阅 · 661 赞 · 98 转

09/26 21:11

Artificial intelligence will make it possible for individuals and companies to produce far more than they can today. That makes the freedom to create, finance, own, and exchange things more important.

中文介绍 MicroStrategy 的 Saylor 谈数字经济下的繁荣处方：AI 会让个人和企业产出远超今天，因此「自由地创造、融资、拥有和交换」变得更重要。核心是 AI 生产力红利必须搭配产权与资本自由才能兑现，属于观点型长文，不含具体工具或教程。

How to build motion design studio with Opus 5.5 ( Full-course )

@0xMovez · 36.4K 粉丝 · 78.9K 阅 · 737 赞 · 46 转

09/27 22:29

Most people who try motion design with Opus 5.5 end up with the same video: centered text on a gradient, everything fading in, a logo at the end. They don't give it a reference, don't give it a

中文介绍 一份用 Opus 5.5 做动态设计工作室的完整课程。作者点出多数人产出雷同的根源：居中文字加渐变、全部淡入、结尾放 logo，因为不给模型参考素材、也不给具体约束。课程教的是如何用参考和提示把作品做出差异，属于实操向教程。

Using Claude Code: Spending your effort

@trq212 · 354.7K 粉丝 · 62.7K 阅 · 965 赞 · 52 转

09/26 04:04

One of the best parts of our newest Claude models is how they respond to effort without breaking the prompt cache in Claude Code, but I’ve received a lot of questions on this from users. What is

中文介绍 Anthropic 的 trq212 解答用户高频疑问：新版 Claude 模型能响应「effort」设定，且在 Claude Code 里不会因此破坏 prompt cache。推文解释这套机制是什么、该如何合理分配你的 effort，属于 Claude Code 的进阶使用技巧分享。

We built a 20x faster Grok bot

@cerebras · 76.6K 粉丝 · 37.6K 阅 · 554 赞 · 28 转

09/25 23:28

Written by @milksandmatcha and @0xSero A personal assistant should save you time and effort. Over the past few weeks, we’ve been obsessively testing AI personal assistants on everyday tasks, from

中文介绍 Cerebras 团队复盘如何做出快 20 倍的 Grok 机器人。他们用数周时间在日程安排等日常任务上反复压测 AI 个人助理，目标是让它真正省时间。内容偏工程优化与推理加速实践，展示自研硬件在低延迟场景下的优势。

OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha

Latent Space · 09/26 07:14

In 2023 most people doubted that there could be more than 1 or 2 frontier model labs. Now there are dozens.... and Stripe just bought the best known one for $7B.

中文介绍 Latent Space 播客对话 OpenRouter 的 Alex Atallah 与 AMP 的 Anjney Midha。节目中提及 2023 年多数人质疑前沿模型实验室能否超过一两家，如今已有数十家，而 Stripe 刚以 70 亿美元收购其中知名度最高的一家。

Proaction boosts sales 60% and saves 75+ hours with Codex

OpenAI News · 09/26 03:00

With Codex, GPT-Live-1, and GPT-6 Astra, Proaction builds, operates, and sells modern fleet management faster.

中文介绍 OpenAI 称 Proaction 借助 Codex、GPT-Live-1 和 GPT-6 Astra，更快地构建、运营和销售现代车队管理系统，实现销售额增长 60%、节省 75 小时以上工时。

The Pentagon wants $30 million to build an AI-powered lie detector

MIT Tech Review AI · 09/25 17:16

The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according to a Department of Defense budget request. The program, called Polygraph+ or Polygraph Next, will focus on scoring algorithms that use artificial intelligence and machine learning a

中文介绍 美国国防部预算申请显示，美国政府计划未来五年投入 3030 万美元研发改进型测谎设备。该项目名为 Polygraph+ 或 Polygraph Next，重点在于使用人工智能的评分算法。

[AINews] The Future of Latent Space

Latent Space · 09/25 13:37

A quiet day lets us discuss the work behind the scenes - now open for business!

中文介绍 Latent Space 在一条消息较少的日期发布 AINews，介绍其幕后工作，并宣布现正式对外开放业务。

Runway’s WorldPrompt and the Engineering of Real-Time Worlds

Latent Space · 09/25 09:30

GWM Worlds 2 uses persistent context and timed actions to steer a world model generating video and audio in real time.

中文介绍 Runway 推出 WorldPrompt，其 GWM Worlds 2 采用持久上下文与定时动作，实时引导世界模型生成视频和音频，Latent Space 讨论了实时世界生成背后的工程实现。

ChatGPT Pro Max 🤖, Muse realtime avatar 🎭, DeepSeek $1B ARR 💰

TLDR AI · 09/25 08:00

中文介绍 TLDR AI 当日快讯汇总：ChatGPT Pro Max、Muse 实时虚拟形象，以及 DeepSeek 达到 10 亿美元年度经常性收入（ARR）。

Introducing Gemini 3.8 Live with Live Avatar

DeepMind Blog · 09/25 00:20

中文介绍 DeepMind 发布 Gemini 3.8 Live，并引入 Live Avatar 实时虚拟形象功能。

Foundries vs Navigators: Lowering the Cost of Science

Latent Space · 09/24 23:03

Guest Post: In science, thinking has gotten cheap but doing has not. This asymmetry is reshaping how research companies operate, largely inconspicuously.

中文介绍 Latent Space 客座文章提出「铸造厂与导航者」之分：在科学领域，思考的成本已很低，但实际执行仍昂贵，这种不对称正悄然重塑研究公司的运作方式。

Accelerating vision-language models with LFM2.5-VL-DSpark

Hugging Face Blog · 09/24 22:08

中文介绍 LiquidAI 在 Hugging Face 博客介绍 LFM2.5-VL-DSpark，用于加速视觉语言模型。

[AINews] Meta Connect 2026: Muse glasses, voice, video, and Charm

Latent Space · 09/24 16:12

Team Zuck is absolutely on fire.

中文介绍 Latent Space 的 AINews 汇总 Meta Connect 2026 大会内容，包括 Muse 智能眼镜、语音、视频以及 Charm 相关发布。

Gemini TTS 🗣️, Claude’s novel enzyme 🧬, Google private memory 🔒

TLDR AI · 09/24 08:00

中文介绍 TLDR AI 当日快讯汇总：Gemini 文本转语音（TTS）、Claude 相关的新型酶，以及 Google 的私密记忆功能。

How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows

Hugging Face Blog · 09/24 02:41

中文介绍 英伟达在 Hugging Face 博客发布教程，介绍如何使用 NVIDIA Warp 与 MjWarp 加速机器人仿真与学习工作流。

Advancing Private AI Compute with secure, server-side memory

DeepMind Blog · 09/24 00:00

Introducing private, server-side memory to Private AI Compute for personal AI.

中文介绍 DeepMind 为面向个人 AI 的「Private AI Compute」引入私有服务器端内存，以推进隐私保护下的 AI 计算。

Two years of OpenAI Academy

OpenAI News · 09/24 00:00

Marking two years of OpenAI Academy and bringing AI skills to even more communities.

中文介绍 OpenAI 纪念 OpenAI Academy 成立两周年，并表示将继续把 AI 技能带给更多社区。

Gemini 3.8 text-to-speech says hello

DeepMind Blog · 09/23 23:25

中文介绍 DeepMind 发布 Gemini 3.8 文本转语音（TTS）功能。