2026-10-06

RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations

👍 256

10/01 08:00

A companion that talks with a person for months should come to understand them. It should remember what they said, infer who they are, and know when the past bears on the message in front of it. Testing this requires a real person's record, and such records are private, so benchmarks generate the pe

中文介绍 针对AI伴侣长期陪伴中需理解用户的问题，提出RealCompanion基准，用真实人物跨越数月的对话记录检验模型能否记忆所言、推断身份并判断历史是否与当前消息相关，解决了此类隐私数据难以获取、现有基准只能合成的问题。

Does Learning Protein Folding Generalize to Broader Reasoning?

👍 113

09/30 08:00

Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them. Protein folding is a natural testbed, because one solved structure yields thousands of exactly checkable spatial and topological statements. We ask: can lea

中文介绍 探究蛋白质折叠任务学到的能力能否泛化到更广推理。因一个已解结构可导出数千条可精确校验的空间与拓扑陈述，作者将其作为天然测试床，考察模型能否借此摆脱对人类文本表层答案的依赖，获得结构化空间逻辑。

MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation

👍 89

09/29 08:00

Vision-language-action (VLA) models have advanced robotic manipulation, but their zero-shot generalization in new tasks and environments remains limited, and their reliance on specialized training keeps them from benefiting directly from rapidly advancing general-purpose vision-language models (VLMs

中文介绍 针对VLA模型在新任务与环境零样本泛化有限、且依赖专用训练而无法直接受益于通用VLM进展的问题，MotorMind以脚手架方式将通用VLM用于机器人操作，无需专门训练即可实现零样本操作，打通通用视觉语言能力与动作执行。

Latent-MOPD: Latent Multi-Teacher On-Policy Distillation

👍 54

10/01 08:00

On-policy distillation (OPD) trains a student on the responses it generates. Existing LLM multi-teacher OPD transfers what specialists predict through their output distributions. We introduce Latent-MOPD, to our knowledge the first representation-level multi-teacher OPD method for LLMs. It integrate

中文介绍 现有LLM多教师on-policy蒸馏只通过输出分布传递专家预测。Latent-MOPD提出首个表示级多教师OPD方法，在学生自生成响应上整合多个专家模型的隐层表征，实现更有效的知识迁移，突破输出层监督的局限。

HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents

👍 51

10/02 08:00

We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging. Each question targets a concise, pub

中文介绍 推出HyperBrowseComp，一个多语言多模态网页浏览基准，含423道由母语或高熟练度者人工编写并经人工校验的题目，覆盖13种语言。题目设计极为困难，每题指向一个简洁且可公开验证的答案，用于压力测试浏览智能体。

Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models

👍 50

10/02 08:00

Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and sh

中文介绍 掩码扩散语言模型（dLM）可并行推理，但去噪中少量确定性选择会迅速降低其余掩码位置的不确定性，造成信用分配难题。Pivot-SD提出高效自蒸馏方法应对该问题，提升掩码扩散模型在复杂推理上的训练效率与表现。

VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks

👍 46

10/01 08:00

As LLM agents undertake increasingly complex, long-horizon tasks, verifying their outputs becomes increasingly challenging. We study how verification capability can be strengthened with a fixed base model, without access to reference answers or grading rubrics at test time. Repeated sampling yields

中文介绍 随着LLM智能体承担越来越复杂的长程任务，验证其输出愈发困难。VeriHarness研究在固定基座模型、测试时无参考答案与评分标准的条件下如何增强验证能力，指出重复采样可提供验证信号，并给出扩展智能体验证的方案。

SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation

👍 46

10/01 08:00

Existing Simulink benchmarks mainly evaluate whether generated models compile, execute, or resemble a reference model. These criteria do not establish whether a model satisfies its engineering requirements. We introduce SimuVerity, a benchmark of 101 text-to-executable Simulink model-generation task

中文介绍 现有Simulink基准只评估生成模型能否编译、执行或与参考模型相似，无法判定是否满足工程需求。SimuVerity提出含101个文本到可执行Simulink模型生成任务的基准，从工程规范角度衡量智能体生成模型的质量。

Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory

👍 44

10/01 08:00

Long-video generation and world models have shown strong potential for interactive entertainment and embodied simulation by predicting future observations conditioned on user actions and historical memory. However, as memory sequences grow longer and their structures become increasingly complex, man

中文介绍 长视频生成与世界模型依赖历史记忆预测未来观测，但记忆序列变长、结构变复杂后管理困难。该工作提出空间记忆智能，为世界模型引入由理解驱动的长期记忆机制，在用户动作条件下生成未来观测，提升长程一致性与交互仿真能力。

Multilingual GSM-Symbolic: What determines capability transfer across languages?

👍 41

10/02 08:00

We understand little about how capabilities acquired in one language carry over to another, or what governs this transfer: evaluations rely on incomparable, saturation-prone datasets and rarely examine its determinants jointly. Identifying what predicts transfer would let us avoid exhaustive evaluat

中文介绍 针对能力跨语言迁移机制不明的问题，指出现有评测依赖不可比且易饱和的数据集，且很少联合考察迁移的决定因素。作者构建多语言GSM-Symbolic，系统分析何种因素决定一种语言习得的能力向另一语言迁移，以减少穷举式评测。

Language Models that Play Chess and Explain Their Moves

👍 27

10/02 08:00

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations. We introduce Queen,

中文介绍 现代国际象棋引擎棋力超人类却不给解释，语言模型能生成看似合理的解释但棋力薄弱。作者提出Queen，将二者结合，使模型既能下棋又能说明走子理由，兼顾棋力与可解释性，提升解释在实战中的实用性。

Local Support Learning

👍 23

10/01 08:00

We explore catastrophic forgetting in the context of large pre-trained models. By considering forgetting as a geometric problem in the input space of each weight matrix, we uncover a natural retention objective under which updates produced by gradient-based optimizers are suboptimal. Following this

中文介绍 从大模型灾难性遗忘出发，把遗忘视为各权重矩阵输入空间中的几何问题，据此揭示梯度优化器产生的更新在该保留目标下并非最优。作者沿此思路提出局部支撑学习，构造更优的更新方向以缓解微调中的遗忘。

Tail-Influence Sampling for CVaR Policy Evaluation

👍 21

09/29 08:00

Policies with similar mean returns can differ sharply in rare failures, yet estimating lower-tail conditional value-at-risk (CVaR) accurately can require many costly rollouts. When different conditional components of a stochastic workflow can be queried separately, we ask how to allocate a fixed eva

中文介绍 均值回报相近的策略在罕见失败上可能差异巨大，而准确估计下尾CVaR需要大量昂贵采样。该工作研究当随机工作流各条件组件可分别查询时，如何把固定评测预算分配到最能影响尾部估计的组件上，提出尾影响采样方法，提高CVaR评估效率。

FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution

👍 21

10/02 08:00

LLM-guided evolutionary methods, such as AlphaEvolve, have emerged as powerful approaches for challenging computational optimization problems, such as circle packing. However, prior work typically optimizes performance gain over a fixed number of iterations. We argue that practical optimization shou

中文介绍 AlphaEvolve等LLM引导的进化方法在圆填充等计算优化问题上表现强大，但通常只优化固定迭代次数内的性能增益。FrugalEvo提出成本感知的程序进化，在优化效果与LLM调用开销之间权衡，使实际优化过程更节省算力与费用。

PointWAM: 3D World Action Modeling for Dexterous Robotic Manipulation

👍 19

10/02 08:00

World action models jointly learn to forecast world dynamics and predict robot actions, such that the learned internal world dynamics guide accurate actions. Existing approaches typically represent the world as RGB frames or latent counterparts while predicting actions as end-effector poses or joint

中文介绍 世界动作模型联合预测世界动态与机器人动作，但现有方法多将世界表示为RGB帧或隐变量，动作表示为末端位姿或关节值。PointWAM以3D点云进行世界与动作建模，让学到的内部动态引导灵巧操作，提升动作预测的准确性。

4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes

👍 19

10/02 08:00

We introduce 4DCodeBench, a benchmark for 4D inverse graphics through code generation, in which agents reconstruct dynamic scenes from video as executable graphics programs. To accomplish this, agents must translate visual observations into compact representations of scene structure and dynamics, by

中文介绍 提出4DCodeBench，通过代码生成考察4D逆图形能力：智能体需从视频重建动态场景，输出可执行的图形程序，将视觉观测转化为场景结构与动态的紧凑表示。该基准评估智能体对动态三维世界的逆向建模能力。

WEFT: Scaling Tool-Use Post-Training for General-Purpose Agents

👍 16

09/29 08:00

Recent efforts to scale tool-use post-training have largely centered on the synthesis of executable environments, which constitute only one component of a broader agentic interaction system comprising the environment, task, agent harness, and evaluator. Scaling environments in isolation, however, do

中文介绍 近期工具使用后训练多聚焦于合成可执行环境，但环境只是环境、任务、智能体框架与评估器构成的更大交互系统的一部分，单独扩展环境效果有限。WEFT提出协同扩展上述组件的后训练方案，提升通用智能体的工具使用能力。

Equal Ranking Quality, Different Decisions: Measuring and Reducing Order Dependence in LLM Scorers

👍 16

09/26 08:00

In passage reranking, response ranking and multi-document question answering, LLMs can score several candidate documents or responses together in one prompt, each still receiving its own score. Such scorers are selected on ranking quality, but their scores determine a decision: what a score threshol

中文介绍 在段落重排、回答排序与多文档问答中，LLM可在同一提示内为多个候选项打分。这类打分器按排序质量挑选，但分数决定阈值决策，因而对候选项顺序敏感。该工作度量并降低LLM打分器的顺序依赖性，在不损失排序质量下稳定决策。

Collective Bias Mitigation via Model Routing and Collaboration

👍 14

10/02 08:00

Large language models (LLMs) are increasingly deployed in public health, finance, and governance, requiring both accuracy and societal value alignment. Despite recent advances, LLMs often perpetuate or amplify bias embedded in their training data, posing challenges to fairness. While self-debiasing

中文介绍 LLM在公共卫生、金融与治理中需兼顾准确性与社会价值对齐，但常延续或放大训练数据中的偏见。该工作提出基于模型路由与协作的集体去偏方案，让多个模型分工配合以缓解偏见，而非依赖单一模型的自我去偏。

MinkowskiPE: Minkowski Positional Encoding for Spatiotemporal Perception

👍 13

09/27 08:00

Modeling spatiotemporal coupling is a key challenge in building physical intelligence across scales, from microscopic to macroscopic. Existing models capture such structure broadly through physics-motivated dynamical formulations or learning-motivated architectures. The former provide stronger prior

中文介绍 跨尺度物理智能需建模时空耦合，现有方法或借助物理启发的动力学公式提供更强先验，或依赖学习型架构。MinkowskiPE提出基于闵可夫斯基度量的位置编码，将时空结构统一编码，为时空感知任务提供兼具归纳偏置与灵活性的表示。

How To Hire A Grok Bot: Experience From The First Zero Human Company.

@BrianRoemmele · 489.9K 粉丝 · 1.1M 阅 · 533 赞 · 85 转

10/05 02:36

You do not build a Grok Bot. You hire one. The difference is the whole article. A chatbot answers a question and forgets the room. A Grok Bot has a name, a job, a conversation that persists, and a

中文介绍 作者提出「聘用」而非「搭建」Grok Bot 的差异：聊天机器人答完就忘，而 Grok Bot 有名字、职位和持续对话记忆。他以「首家零人类公司」的实践为例，讲如何为业务配置有身份的 AI 员工。

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 播客作者谈 AI 仍处「第 0.000 天」的极早期阶段，类比掌控算力者掌控一切，认为接下来变化只会更剧烈。偏宏观趋势判断，提醒观众别因当下进展已足够惊人就低估后续空间。

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 LiveAvatar 推广 AI 语言学习方案，主张人人都能拥有耐心的一对一外教。内容围绕口语这一最难环节展开，讲会给学习者、教师和学校带来什么变化，属于产品软文性质。

VIBE MANUFACTURING IS HERE

@gregisenberg · 724.2K 粉丝 · 44.1K 阅 · 569 赞 · 53 转

10/04 21:25

For the FIRST TIME in HISTORY, you can describe a physical product in one sentence and, a week later, hold it in your hands, cut from metal and built to your exact measurements. That's VIBE

中文介绍 Greg Isenberg 提出「Vibe Manufacturing」概念：用一句话描述实体产品，一周后即可拿到按精确尺寸切割的金属成品。强调这是历史上首次实现从自然语言到实体制造，属于概念造词与趋势判断。

Evals: how to know whether an AI system actually works

@sermakarevich · 549 粉丝 · 33.0K 阅 · 520 赞 · 64 转

10/04 02:38

An article for everyone who ships, buys, or signs off on software built on large language models (LLMs): engineers, product managers, and the CEO. Written in plain language, from the big picture down

中文介绍 一篇面向工程师、产品经理和 CEO 的 LLM 评估科普文，用大白话从整体框架讲到细节，讨论如何判断一个基于大模型的系统是否真的可用。切入点偏方法论，适合需要验收或采购 AI 软件的人。

LIST OF HACKATHON FOR OCTOBER.

@onlyonealexia · 1.6K 粉丝 · 26.5K 阅 · 509 赞 · 53 转

10/04 10:08

A List of Web3, AI & Web2 Hackathons Compiled by @onlyonealexia There are a lot of hackathon lists that look impressive until you actually open the links. I compiled a list of ongoing and upcoming

中文介绍 作者整理了一份 10 月 Web3、AI 与 Web2 黑客松清单，吐槽很多列表看着唬人、点开链接才发现无效。本人筛选了正在进行和即将开始的活动，并保留报名链接，供开发者挑选参赛。

All games are open source now

@rehan_shei · 11.9K 粉丝 · 24.9K 阅 · 520 赞 · 49 转

10/04 08:52

A week ago I pointed Opus 5.5 at Terraria before bed. I woke up to a fully working, reverse engineered version of the game. Then I added a homing missile launcher, a tactical nuke and a bunch of new

中文介绍 作者睡前让 Opus 5.5 分析 Terraria，醒来已得到可运行的反编译版本，随后还加入了追踪导弹、战术核弹等新内容。借此论证游戏逆向与改造门槛已被 AI 大幅拉低，二手游戏源码几乎等于开放。

Reinforcement Learning from scratch #1

@goyal__pramod · 11.4K 粉丝 · 24.7K 阅 · 504 赞 · 59 转

10/03 12:45

Now I am aware you must have been looking forward to the second part of my CUDA blog (if you haven't read it yet, check it out!), but hey! A man can have varied interests. And I believe if you are

中文介绍 CUDA 系列博主转向新主题，开写「从零实现强化学习」第一篇。作者自称兴趣广泛，承诺后续仍会补 CUDA 第二篇，属于面向初学者的教学连载开篇。

Connecting AI agents to enterprise knowledge

MIT Tech Review AI · 10/05 23:47

For all the data that AI systems continually amass and analyze, enterprise AI agents often suffer from a curious shortcoming: a lack of knowledge. More than data, knowledge is the understanding of what the data means in the context of individual organizations. AI agents need this understanding to re

中文介绍 MIT Tech Review 探讨企业 AI 智能体的知识短板：其真正缺的不是数据，而是理解数据在具体组织语境中含义的「知识」。文章聚焦如何将 AI 智能体与企业知识相连接。

Our approach to EU text provenance rules

OpenAI News · 10/05 23:00

How OpenAI is approaching text watermarking under EU rules. Learn where watermarks apply, how detection works, and why access starts with researchers.

中文介绍 OpenAI 介绍其在欧盟文本溯源规则下对文本水印的应对方式，说明水印的适用范围、检测机制，以及为何初期仅向研究人员开放相关访问权限。

Bringing predictive analytics to the agentic AI era

MIT Tech Review AI · 10/05 21:29

In 2026, the question for enterprise AI is no longer whether predictive models can outperform statistical forecasts—that argument is settled. The big question now is how to enable predictive systems to act on their own conclusions without drifting from business intent. The frontier has moved from pr

中文介绍 MIT Tech Review 指出，2026 年企业 AI 的问题已不再是预测模型能否超越统计预测，而是如何让预测系统依据自身结论自主行动，同时不偏离业务意图。

Building advertising for the way people use AI

OpenAI News · 10/05 18:00

OpenAI introduces a new visual ad format in ChatGPT and expands measurement tools, attribution partnerships, and brand suitability for advertisers.

中文介绍 OpenAI 在 ChatGPT 中推出全新可视化广告格式，并扩展广告效果衡量工具、归因合作伙伴关系及面向广告主的品牌适配性功能。

People really hate AI, so why can’t they get enough?

MIT Tech Review AI · 10/05 16:00

Over the summer I talked to the CEO of Springboards, a startup building an LLM that’s designed to come up with a wider variety of responses than its mainstream rivals do. At the start of the call, he said something that’s been stuck in my head since: “We often say that we’re a self-loathing AI…

中文介绍 MIT Tech Review 作者回顾与初创公司 Springboards CEO 的对话，该公司正构建一款旨在比主流对手产出更多样回答的 LLM。文章探讨人们一边反感 AI、一边又难以割舍的矛盾现象。

EmTech Future 2026: When AI Meets Everything

MIT Tech Review AI · 10/05 12:00

Yossi Matias, Vice President & Head of Google Research, explores how AI is beginning to reshape biology, infrastructure, manufacturing, and science, and why its greatest impact may come when it intersects with other fields. Step inside the newsroom with our MIT Technology Review editors for sharp an

中文介绍 在 EmTech Future 2026 上，Google Research 副总裁兼负责人 Yossi Matias 探讨 AI 如何开始重塑生物学、基础设施、制造业与科学，并指出其最大影响或在与其他领域交汇时出现。

Prime Inference 📈, agent population 📈, Claude academy 🎓

TLDR AI · 10/05 08:00

中文介绍 TLDR AI 2026 年 10 月 5 日简报，关键词包括 Prime Inference、智能体数量增长以及 Claude 学院。

The Agent Said It Was Done. The Database Disagreed.

Hugging Face Blog · 10/04 06:56

中文介绍 Hugging Face 博客文章，标题意为「智能体说它已完成，数据库却不同意」，主题涉及智能体执行结果与实际数据状态之间的不一致问题。

[AINews] not much happened today

Latent Space · 10/03 16:45

a quiet day.

中文介绍 Latent Space 的 AINews 栏目当日报导称人工智能领域消息较为平淡，是安静的一天。

A model guide for the GPT-6 family

OpenAI News · 10/03 00:15

Learn how startups can choose GPT-6 models, tune reasoning effort, improve prompts and skills, coordinate tools, and prepare workflows for production.

中文介绍 OpenAI 发布 GPT-6 系列模型使用指南，介绍初创公司如何选择 GPT-6 模型、调节推理强度、优化提示与技能、协调工具调用，并为生产环境准备工作流。

Redefining enterprise intelligence with autonomous AI

MIT Tech Review AI · 10/02 23:49

Enterprise AI is no longer a future ambition. It is in full operational flight. Model capabilities are advancing faster than most organizations can absorb, while the cost of performance continues to fall. Globally, AI investment is set to reach $2.5 trillion in 2026, up 44% from the previous year. F

中文介绍 MIT Tech Review 指出企业 AI 已全面进入运营阶段，模型能力提升速度超过多数组织的吸收能力且性能成本持续下降。2026 年全球 AI 投资预计达 2.5 万亿美元，同比增长 44%。

Open-sourcing AstaBrief, the fast report-generation model in Asta

Hugging Face Blog · 10/02 23:19

中文介绍 Allen AI 在 Hugging Face 博客宣布开源 AstaBrief，这是 Asta 中用于快速生成报告的开源模型。

Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience

Latent Space · 10/02 22:04

After leading Meta’s Llama models, Ahmad Al-Dahle is now transforming Airbnb with AI — from how its teams develop products to how it serves guests.

中文介绍 Latent Space 报导，曾主导 Meta Llama 模型的 Ahmad Al-Dahle 现投身 Airbnb，正以 AI 重塑该公司，涵盖团队产品开发方式与面向房客的服务体验。

Don’t be fooled—LLMs don’t reason

MIT Tech Review AI · 10/02 16:00

On an afternoon in Seoul in March 2016, I watched a program I helped build put a stone on the fifth line of a Go board in what looked like a gift to its human opponent. Move 37 in game two of the five-game match looked so absurd that some commentators thought it was a…

中文介绍 MIT Tech Review 作者以 2016 年 3 月首尔一场围棋赛第 37 手为例——该手看似荒诞，部分评论者一度以为失误——强调不应被表象迷惑，大语言模型并不真正具备推理能力。

[AINews] Pi 1.0, Pi Durable, and AIE NYC

Latent Space · 10/02 14:40

the minimalist harness goes stable... and TypeScript!

中文介绍 Latent Space 的 AINews 报导，极简式智能体框架 Pi 发布 1.0 稳定版并推出 Pi Durable，同时采用 TypeScript；此外还有 AIE 纽约相关动态。