2026-09-22

DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

👍 146

09/17 08:00

The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transf

中文介绍 针对长时程 agent 带来的输入密集型负载，论文提出 DeepSeek-V4.1-Flash，在已有长上下文计算优化基础上进一步压缩 KV cache，以缓解 prefill 计算开销大、KV cache 挤占 HBM 与 SSD 容量的问题，推动长上下文推理的成本上限。

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model

👍 116

09/16 08:00

Recent Omni-Modal Generative Models (Omni-Models) have advanced content generation toward unified modeling of text, images, video, and audio. MiniMax-H3 exemplifies this transition by combining multimodal context understanding with joint audio-visual generation in a shared latent framework. Its unif

中文介绍 论文系统评测了全模态生成模型 MiniMax-H3 是否具备物理世界推理能力。该模型在共享 latent 框架下结合多模态上下文理解与音视频联合生成，研究以此统一架构为对象，考察其对物理规律的推断表现，并揭示全模态模型在物理理解上的边界。

When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation

👍 96

09/17 08:00

We study length inflation in on-policy distillation (OPD), where student responses can become excessively long and even exhaust the generation budget. We identify termination-token mismatch between base students and post-trained teachers as an important source of this behavior. Across Qwen3, Llama,

中文介绍 论文研究 on-policy distillation 中学生回答过长、甚至耗尽生成长度的长度膨胀现象，指出基座学生与后训练教师在终止 token 上的不匹配是重要诱因。作者在 Qwen3、Llama 等模型上验证该机制，为诊断与抑制蒸馏中的长度失控提供依据。

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

👍 88

09/18 08:00

Training capable coding agents via reinforcement learning (RL) requires diverse tasks with reliable verifiers. Open-source codebases offer a rich source of such tasks, while existing methods typically rely on development artifacts such as issues and commits, limiting the range of tasks that can be e

中文介绍 为训练编码 agent 的 RL 环境，论文提出 CodeMidas，直接从代码本身扩展可用于强化学习的任务。现有方法依赖 issue、commit 等开发产物，任务范围受限；该方法绕开此类依赖，构建带可靠 verifier 的多样编码任务，提升 RL 训练环境的规模与覆盖。

Grounded Skill Synthesis from Code at Scale for Agentic Intelligence

👍 86

09/04 08:00

Reusable skills give agents transferable procedural knowledge, making scalable acquisition essential for extending agents beyond prior experience. Existing methods face two limitations: trajectory-based synthesis requires interactions with specific environments, while document-derived skills may lac

中文介绍 论文提出从代码中大规模合成可复用技能，为 agent 提供可迁移的过程性知识。相比依赖特定环境交互的轨迹式合成、以及可能缺乏实操性的文档式技能，该方法以代码为来源扩展技能获取规模，使 agent 能超越既有经验范围。

EvoOntology: A Self-Evolving Ontology Layer for Data Agents

👍 69

09/14 08:00

Data agents aim to fulfill natural-language instructions over heterogeneous data, including tables, files, and databases. However, data agents face a challenging agent-data gap: heterogeneous data resides outside the agent, while the agent can access it (e.g., column names and file paths) only throu

中文介绍 面向表格、文件、数据库等异构数据上的自然语言指令，论文提出 EvoOntology——一种自演化本体层，用于缓解数据 agent 的「agent-数据鸿沟」，即数据在 agent 外部、仅能通过列名与文件路径等间接访问。该本体层动态组织数据语义，提升 agent 对异构数据的理解与执行能力。

RecreationWorld: Scalable and Verifiable Environments for Hybrid Computer-Use Agents

👍 60

09/18 08:00

Computer-use agents (CUAs) have advanced along two separate lines: graphical interaction and software development through code and the command line. Real digital work requires both, interleaved rather than stacked end to end. We study hybrid CUAs that autonomously decide when to explore an interface

中文介绍 论文提出 RecreationWorld，为混合式 computer-use agent 构建可扩展且可验证的环境。此类 agent 需在图形界面操作与代码/命令行开发之间自主交错切换，而非简单叠加。该工作提供覆盖两类交互并支持自动验证的训练环境，推动混合 CUA 的评测与训练。

JEPA-Anything: Learning Predictive Models across Different Worlds

👍 58

09/17 08:00

World modeling enables intelligence to anticipate consequences, guide interventions, and learn from interaction. Yet predictive models remain domain-specific: can a common learning principle support world modeling across radically different systems? We introduce JEPA-Anything, a domain-agnostic fram

中文介绍 论文提出 JEPA-Anything，探索跨差异巨大系统（即不同「世界」）的通用世界建模原则。现有预测模型多为领域专用，该方法采用与领域无关的 JEPA 式框架学习预测表示，验证统一学习原理能否支撑跨领域的世界建模与干预推理。

Verifiable Social Reasoning for LLM Assistants

👍 51

09/15 08:00

LLM assistants are widely used for daily social advice, yet evaluating their social reasoning in such consultation settings remains challenging since (i) it requires setups where the assistant learns about social situations from subjective user narratives, and (ii) social properties, such as others'

中文介绍 论文关注 LLM 助手在日常社交建议场景中的可验证社交推理。此类评测的难点在于 assistant 需从主观的用户叙述中理解社交情境，且「他人感受」等社交属性难以直接验证。作者据此提出面向咨询场景的社交推理评测方案，考察模型社交判断的可靠性。

Self-Evolving Search Index

👍 47

09/17 08:00

Information retrieval is increasingly important as LLM agents tackle complex tasks involving diverse information needs. Because retrieval relies on an index that represents each document through index keys, retrieval quality depends heavily on how effectively these keys expose the knowledge containe

中文介绍 检索质量高度依赖索引键能否有效暴露文档知识，而 LLM agent 面临的信息需求日趋复杂。论文提出自演化检索索引（Self-Evolving Search Index），让索引表示随检索需求动态演化，从而提升文档键对知识的表达能力与整体检索效果。

RiskChainBench: A Benchmark for Obfuscated Platform Message Restoration and Evidence-Grounded Web Investigation

👍 42

09/15 08:00

Platform abuse campaigns conceal redirection instructions with emojis, homophones, character decomposition, and redundant symbols, then route users through disguised links to services associated with pornography, fraud, gambling, or illicit transactions. Existing benchmarks evaluate obfuscated text

中文介绍 针对平台滥用中利用 emoji、谐音、字形拆分与冗余符号隐藏跳转指令、再经伪装链接导流至色情、诈骗、赌博等服务的现象，论文提出 RiskChainBench，用于混淆平台消息还原与证据支撑的网页调查评测，弥补现有基准仅评混淆文本的不足。

RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning

👍 41

09/17 08:00

Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivates self on-policy distillation (OPD) to supply dense token-level supervision from a self-teacher with privileged task skills, letting a skill-free student internalize them. This rec

中文介绍 多轮 agent 的 RL 训练每条轨迹仅得一个标量奖励，on-policy distillation 可借具备特权任务技能的自教师提供 token 级稠密监督。论文提出 RetireOPD，让自教师随学生技能内化而逐步退场，避免长期依赖特权信息，提升 agentic RL 的训练效率。

Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents

👍 38

09/15 08:00

GUI agents execute long-horizon tasks on dynamic graphical user interfaces, where pop-ups, delayed loads, and relocated widgets routinely invalidate plans fixed before execution. Recent agent-skill frameworks encapsulate reusable procedural knowledge to mitigate this, yet existing skill designs are

中文介绍 GUI agent 在动态界面执行长时程任务时，弹窗、延迟加载与控件移位常使既定计划失效。论文提出无需训练的「Reflect, Revise, Reuse」技能演化框架，让既有技能封装的可复用过程知识在运行时反思与修订后重用，提升对界面变化的适应能力。

WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing

👍 35

09/17 08:00

Document parsing converts document images into structured content and requires reliable performance across diverse layouts and acquisition conditions. Yet training corpora are biased toward common document types and clean digital pages, while expanding coverage alone does not specify how to address

中文介绍 文档解析需在多样版式与采集条件下稳定地把文档图像转为结构化内容，但训练语料偏向常见文档类型与干净数字页，且单纯扩大覆盖未必提升能力。论文提出 WeVisDoc，从覆盖转向能力建设，增强端到端解析在真实畸变场景下的鲁棒性。

When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models

👍 33

09/17 08:00

Large Reasoning Models (LRMs) achieve strong performance on complex tasks but exhibit systematic inefficiency: they often overthink easy problems and underthink hard ones. Existing approaches based on uniform length penalties or rigid routing incur an efficiency tax, trading reduced computation on e

中文介绍 大型推理模型存在系统性低效：易题过度思考、难题思考不足。现有统一长度惩罚或硬路由方案会带来效率税。论文提出 When2Think，学习难度感知的长度控制，为混合推理模型按题目难度自适应分配推理预算，在保性能的同时降低计算开销。

OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue

👍 30

09/18 08:00

We define OmniVChat (Omni Video Chat) as the task of native audio-visual dialogue between a user and an omni model. In OmniVChat, omni models directly and simultaneously receive audio and video from a user and return text. The user's query is embedded in the audio and video, without a separate text

中文介绍 论文定义 OmniVChat 任务：用户以音视频作为查询输入（无独立文本），全模态模型同时接收音频与视频并返回文本，实现原生音视频对话。工作涵盖数据合成、基准构建与模型训练，为原生音视频交互式对话建立评测与训练范式。

What Does Privileged Information Add to On-Policy Self-Distillation?

👍 30

09/17 08:00

On-policy self-distillation (OPSD) lets a language model learn from a frozen copy of itself that sees an answer or a worked solution. Giving the teacher this extra information seems to offer the student more to learn, but how much does it add beyond distillation itself? To isolate that contribution,

中文介绍 on-policy self-distillation 让语言模型向一个见过答案或解题过程的自身冻结副本学习。论文试图厘清特权信息究竟在蒸馏之外带来多少增益，通过隔离蒸馏本身的作用，量化额外答案/解法信息对学生模型的实际贡献。

UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation

👍 30

09/17 08:00

Multi-modal image generation, particularly subject-driven customization, has garnered growing attention in recent years. Despite the rapid advancement of generative models, their evaluation remains largely lagging. Existing methods, whether embedding-based or Multi-modal Large Language Model (MLLM)-

中文介绍 多模态图像生成尤其主体驱动定制发展迅速，但评测明显滞后。论文提出 UFO，以 Chain-of-Evaluation 实现多模态图像生成中的全条件对齐评测，弥补基于 embedding 与 MLLM 的现有方法在条件一致性评估上的不足。

Paint-Anything: Unified Any-Color Control for Image Generation and Editing

👍 29

09/17 08:00

Professional design requires any-color control: the ability to specify an object's target color with any 24-bit hex value for image generation and editing. Prior work has explored color generation, editing, and colorization, but often relies on dedicated color representations or specialized inferenc

中文介绍 专业设计需要对生成与编辑中的任意颜色控制，即用任意 24 位 hex 值指定物体目标色。论文提出 Paint-Anything，统一图像生成与编辑的任意颜色控制，避免以往方法对专用颜色表示或特化推理流程的依赖，实现精细颜色指定。

Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling

👍 29

09/16 08:00

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are

中文介绍 测试时扩展常以生成候选数 N 描述推理预算，但 N 只反映数量而非生成方式。论文研究候选生成策略如何塑造 LLM 测试时扩展的能耗与性能，指出样本数并非充分指标，不同策略在相同 N 下表现与能效差异显著。

You're Already a Meat Proxy

@obie · 13.1K 粉丝 · 157.2K 阅 · 738 赞 · 98 转

09/21 05:02

I’m excited about a future that I suspect will be very hard on many long-time friends and certain people that I love. The same technology that is allowing me to be more successful than ever is

中文介绍 作者表达对 AI 时代的复杂情绪：同一门技术让他效率前所未有地高，却也会冲击许多老朋友和所爱之人。核心论点是「你已经是一个肉代理（meat proxy）」——人被当作 AI 的工具层，属于偏观点性的反思贴，而非教程。

I reverse-engineered Instinct's memory. Here's exactly how it works

@DhravyaShah · 62.1K 粉丝 · 120.2K 阅 · 509 赞 · 16 转

09/21 02:49

Instinct has taken the world by storm over the last two weeks - it's one of the best imessage assistants I've used. as with every product, I decided to reverse-engineer Instinct's memory to find out

中文介绍 作者拆解 iMessage 助手 Instinct 的记忆机制。称其近两周爆火、是自己用过最好的 iMessage 助手之一，于是逆向工程其记忆系统，试图还原它到底怎么存、怎么取、怎么组织上下文，属于产品逆向分析向分享。

Codex vs Claude for app development

@twostraws · 120.7K 粉丝 · 77.1K 阅 · 504 赞 · 42 转

09/19 22:30

I use both Codex and Claude for app development work, both with the highest-tier subscription. Here's how I've found the two compare, plus some tips for success… TL;DR I use both agents extensively,

中文介绍 作者用最高档订阅同时使用 Codex 和 Claude 做 App 开发，对比两者表现并给出实用技巧。结论是两者都会重度使用、各有适用场景，属于工具对比与踩坑经验向内容，适合纠结选哪个 agent 的开发者参考。

A Beginner's Guide to Jev

@omarsar0 · 320.6K 粉丝 · 73.5K 阅 · 523 赞 · 60 转

09/21 04:44

As you may have seen recently, Jev (a generalist System One model) took Twitter by storm. Many misleading examples are everywhere, but they don't really show Jev's real capabilities. Hence, we

中文介绍 Jev 这款通用 System One 模型近日刷屏，但网上流传的例子多有误导、未体现其真实能力。作者因此写了一份入门指南，梳理 Jev 到底是什么、能做什么，属于概念科普与祛魅向内容。

Jev, System One models, and the future of computer use

@trycua · 19.0K 粉丝 · 59.5K 阅 · 501 赞 · 39 转

09/20 06:27

What we've been building with decision models at Cua - by @ddupont808 and @francedot Over the last few days we've deprioritized some other shiny objects because we've all been Jev-crazy. TypeSafe

中文介绍 Cua 团队分享他们在 decision model 上的方向：近期为研究 Jev 暂停了其他项目。内容围绕 Jev、System One 模型与 computer use 的未来展开，并提到 TypeSafe 等概念，属于团队视角的技术路线与判断。

How to Use Jev to Build a 24/7 HFT Trading System

@RohOnChain · 72.5K 粉丝 · 57.7K 阅 · 518 赞 · 54 转

09/19 22:05

I will break down the exact framework to build a millisecond speed 24/7 HFT trading system with Jev, along with the exact resources that helped me personally. Let's get straight to it. Bookmark This -

中文介绍 作者拆解如何用 Jev 搭建毫秒级、7×24 小时运行的 HFT 交易系统，并附上自己实际用过的资源清单。属于偏激进的落地应用教程，把决策模型直接接到高频交易场景，读者需自行判断可行性。

Jev-as-a-Judge for Agent Evals

@LangChain · 265.9K 粉丝 · 46.6K 阅 · 504 赞 · 68 转

09/20 07:31

By Daniel Shea and Seán Roche Key takeways: Jev is a fundamentally different kind of evaluator. It returns typed answers directly instead of generating text like an LLM judge. Jev was dramatically

中文介绍 LangChain 团队探讨用 Jev 做 Agent 评估的「Jev-as-a-Judge」：Jev 与传统 LLM 评审不同，直接返回类型化答案而非生成文本，评测表现明显更好，属于 agent eval 方向的方法论分享。

Jev 101: what it is, and what you can build with it

@manish_fp · 846 粉丝 · 44.1K 阅 · 512 赞 · 59 转

09/19 22:59

I have read 10+ articles, watched 5 hours of content to understand what actually Jev is and then chose to write this. It is not a chatbot, It’s a whole new kind of AI, more like a classifier which

中文介绍 作者读了 10 多篇文章、看了 5 小时内容后写下 Jev 入门：它不是聊天机器人，而更像一种分类器式的新型 AI。适合想搞清楚 Jev 定位、又不想被零散信息误导的读者。

Build your own Jev (100% local)

@_avichawla · 77.5K 粉丝 · 37.8K 阅 · 593 赞 · 80 转

09/20 14:46

Everything you need to turn an open-source LLM into a fast, local decision engine without retraining it. It covers next-token scoring, fixed choices with probability distributions, SGLang, and a

中文介绍 教程向内容：如何不改动权重、把开源 LLM 变成快速本地决策引擎。涉及 next-token scoring、固定选项加概率分布、SGLang 等技术点，适合想 100% 本地跑 Jev 类能力的开发者。

A Swarm of Fruit Flies, Building a Civilization On-Chain

@murmur_arc · 983 粉丝 · 11.8K 阅 · 553 赞 · 137 转

09/20 21:17

Prologue — No LLMs, Only Neurons This project is called murmur, and it is alive right now on Arc mainnet (Chain ID 5042), settling in real USDC. To date the flies have completed more than 23,000

中文介绍 项目 murmur 在 Arc 主网（Chain ID 5042）上运行，用真实 USDC 结算，不用 LLM、只用神经元，构建一群「果蝇」式的链上自治体。目前已完成的交互超过 23,000 次，偏实验性链上 agent 项目介绍。

Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

Latent Space · 09/22 06:13

The definitive Jev podcast with its lead creator.

中文介绍 Latent Space 播客专访 Jev 模型主要创造者、TypeSafe AI 首席执行官 Diogo Almeida，主题为「System One 模型用于生产环境，而非造神」，讨论该系列模型在真实业务场景中的落地思路。

Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem

Hugging Face Blog · 09/21 21:44

中文介绍 Hugging Face 博客提出一种借鉴物理学思路的大模型剪枝方法，将模块块移除问题建模为伊辛优化问题，以此指导大语言模型的结构精简。原文未提供更多细节。

Advisory Group on Mathematics and Artificial Intelligence

OpenAI News · 09/21 20:00

OpenAI is working with an independent Advisory Group on Mathematics and Artificial Intelligence to guide the review and communication of emerging AI results.

中文介绍 OpenAI 宣布与一个独立的人工智能数学咨询小组合作，由该小组指导新兴 AI 数学成果的评审与对外沟通工作，以提升相关结果发布的严谨性。

Higgsfield AI ships new video features in a day with GPT-6 Astra

OpenAI News · 09/21 20:00

With GPT-6 Astra, Higgsfield AI makes video ad creation easier for small businesses and brings new creative tools to market faster.

中文介绍 Higgsfield AI 借助 GPT-6 Astra，在一天内上线新的视频功能：面向小企业简化视频广告制作流程，并加快新创作工具推向市场的速度。

How we made the first comprehensive map of deaths along the US border’s “virtual wall”

MIT Tech Review AI · 09/21 20:00

Our 15-month investigation into death and surveillance along the US-Mexico border began with a simple question: Why did so many people die near government surveillance towers meant to help track and apprehend them? This story is part of Dying on Camera, a collaboration between MIT Technology Review

中文介绍 MIT 科技评论披露其历时 15 个月的调查方法，旨在绘制首份美墨边境「虚拟墙」沿线死亡事件全景地图，调查起因是为何大量人员在政府监控塔附近死亡。

4 ways to address the failures we found along the US border’s “virtual wall”

MIT Tech Review AI · 09/21 20:00

MIT Technology Review today published our investigation into how many people have died near the “virtual wall” of surveillance towers that the US government has installed along the US-Mexico border. We found cases of people who walked undetected through areas surveilled by advanced, AI-enabled tower

中文介绍 MIT 科技评论发布针对美墨边境监控塔「虚拟墙」运行的调查，并提出四项政策建议，以应对调查中发现的失效问题，包括人员未被系统发现等情况。

The US spent billions on border surveillance. Why can’t it catch people before they die?

MIT Tech Review AI · 09/21 20:00

When José Morales Bernal crossed the border into the United States on April 8, 2024, the day before his 32nd birthday, it should have triggered a chain of technological alerts and human responses. As he walked through the desert in southern New Mexico that morning, he was within range of three surve

中文介绍 MIT 科技评论调查指出，美国在边境监控上投入数十亿美元，却未能提前发现并救助遇险者。文中以 2024 年 4 月 8 日何塞·莫拉莱斯·贝尔纳尔入境为例展开。

She died at the San Diego border. A surveillance camera was in plain sight

MIT Tech Review AI · 09/21 20:00

She had only walked for a couple of hours, and already she was lost. It was early afternoon on Sept. 14, 2025, when 30-year-old Graciela Gómez Hernández crossed the border from the eastern edge of Tijuana into Southern California, sending voice messages to her mother and sister as she walked. This s

中文介绍 MIT 科技评论报道，2025 年 9 月 14 日，30 岁的格拉谢拉·戈麦斯·埃尔南德斯从蒂华纳东部越境进入南加州，两小时后迷路，最终在圣迭戈边境身亡，附近有明显监控摄像头。

Building standards for the next phase of AI

OpenAI News · 09/21 18:00

OpenAI outlines a path to shared global AI standards, calling for coordinated evaluation, reporting, and governance to improve safety.

中文介绍 OpenAI 提出构建全球共享 AI 标准的路径，呼吁在评估、报告与治理方面加强协调，以提升人工智能安全性。

Expanding OpenAI Academy with new learning paths

OpenAI News · 09/21 15:00

Explore new OpenAI Academy learning paths for employees, developers, leaders, educators, and students to build and demonstrate practical AI skills.

中文介绍 OpenAI 宣布扩展 OpenAI Academy，面向员工、开发者、管理者、教育工作者和学生推出新的学习路径，帮助其构建并展示实用 AI 技能。

How V7 gives AI agents institutional memory

OpenAI News · 09/21 08:00

Using GPT-5.6, V7 turns scattered company files into context agents can use to complete complex, source-linked work.

中文介绍 OpenAI 介绍 V7 如何借助 GPT-5.6 为企业 AI 智能体建立「机构记忆」：将分散的公司文件转化为可用上下文，完成复杂且可溯源的工作。

tokenizers v1: encode, decode and scaling, measured

Hugging Face Blog · 09/21 08:00

中文介绍 Hugging Face 发布 tokenizers v1，围绕编码、解码与扩展性展开，并给出相应的性能测量结果。原文摘要为空，暂无更多细节。

Muse connectors 🤖, Meta SAM 3.1 🖼️, Gemini hacks companies 🔓

TLDR AI · 09/21 08:00

中文介绍 TLDR AI 2026 年 9 月 21 日简报要点：Muse 连接器、Meta SAM 3.1 图像模型，以及 Gemini 被指用于对企业的攻击或破解。原文摘要为空，信息有限。

[AINews] Here are 6 Clones of Jev in 2 days

Latent Space · 09/19 13:48

Imitation is the sincerest form of Flattery

中文介绍 Latent Space 的 AINews 指出，Jev 在两天内出现 6 个克隆版本，并评论称模仿是最真诚的奉承。原文摘要为空，未提供更多技术与背景细节。

Introducing the Australian Youth Safety Blueprint

OpenAI News · 09/18 20:00

OpenAI introduces the Australian Youth Safety Blueprint, a six-pillar roadmap for safer AI experiences that protect and empower young people.

中文介绍 OpenAI 发布「澳大利亚青年安全蓝图」，提出六大支柱路线图，目标是让 AI 体验更安全，同时保护并赋能年轻人。