2026-09-23

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

👍 154

09/21 08:00

An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iteratively proposing and selecting component-wise edits of an agent ha

中文介绍 针对冻结主干模型下 agent 的 harness（提示词、控制流、工具、记忆、上下文管理）优化问题，提出 RRSI 正则化递归自我改进方法，迭代地提出并筛选组件级编辑。通过正则化约束避免自我改进中的退化，在保持稳定性的同时持续提升 agent 性能，为 harness 自动化设计提供了可复用的递归优化范式。

Transferring the Intelligence of VLMs to Robotic Control

👍 101

09/19 08:00

Humans can seamlessly adapt to both physical and digital worlds, suggesting that while a digital-to-real gap exists in embodiment, environment and task, human intelligence itself may transfer across this gap. This naturally raises a fundamental question: can the intelligence of vision-language model

中文介绍 探究视觉语言模型（VLM）的智能能否跨数字与真实世界的具身鸿沟迁移到机器人控制。论文提出将 VLM 的感知与语义理解能力迁移至机器人操作，通过适配训练让 VLM 输出可执行控制，验证了数字域预训练智能在实体环境与任务上的可迁移性，为通用机器人策略提供了新路径。

OmniEdu: Open Foundation Models for Learning and Teaching

👍 70

09/19 08:00

Educational foundation models must solve problems, understand curriculum structure, diagnose learner difficulties, and provide appropriate instructional support. Existing educational language models often focus on either problem solving or tutoring, with training mixtures organized by source or task

中文介绍 面向教育场景的教学基础模型需兼具解题、理解课程结构、诊断学习困难与教学支持能力，而现有教育语言模型多偏重单一方向。论文提出 OmniEdu 开放基础模型，按教学能力组织训练数据混合，统一解题与辅导任务，为学习与教学提供多能力覆盖的开源教育基础模型。

Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion

👍 48

09/21 08:00

Retrieval-Augmented Generation (RAG) systems over enterprise knowledge bases must ingest heterogeneous document formats -- PDFs, Word documents, presentations, and scans -- whose content is locked inside complex visual layouts, multi-column pages, and dense tables. Rule-based extraction and OCR dest

中文介绍 针对企业知识库 RAG 摄入 PDF、Word、演示文稿与扫描件时内容困于复杂版式、多栏与密集表格的问题，提出 D-RAC：先做 PDF 归一化，再转多模态 Markdown，实现检索感知的通用文档切分与摄入。相比规则抽取与 OCR，该方法提升了异构企业文档的检索可用性与切分质量。

VideoGen-Agent: Reinforcing Video Generation Agents

👍 35

09/21 08:00

Recent advances in video generative models have enabled high-fidelity, temporally coherent video generation. However, these models often struggle to satisfy prompts requiring specialized knowledge, specific identities, physical consistency, or ordered events. In this paper, we present VideoGen-Agent

中文介绍 视频生成模型难以满足涉及专业知识点、特定身份、物理一致性与有序事件的提示。论文提出 VideoGen-Agent，以 agent 方式强化视频生成流程，通过规划、检索与反思等智能体行为补足模型能力，在复杂提示下提升生成视频的知识准确性与事件顺序一致性。

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

👍 31

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 定义 OmniVChat 任务：用户直接以音视频输入与 omni 模型进行原生音视频对话，查询内嵌于音频与视频中，无独立文本，模型需同步接收并返回文本。论文构建了数据合成、基准评测与训练方案，推动 omni 模型在无文本提示下的原生音视频交互能力研究。

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

👍 27

09/21 08:00

We present onPanda, an interactive tool for efficiently annotating LLM alignment data and agent trajectories. onPanda adopts token-level correction as its core interaction: while reading a model response, the annotator locates the first inappropriate token and either picks a substitute from the mode

中文介绍 提出 onPanda 交互式标注工具，用于高效标注 LLM 对齐数据与 agent 轨迹。其核心是 token 级纠正：标注者定位首个不当 token 并从候选中选取替换或直接修改，大幅减少全量重写成本，在保持标注质量的同时显著提升 on-policy 对齐数据与轨迹标注效率。

Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

👍 25

09/18 08:00

Professional graphic design is a long-horizon agentic task in which structured, editable artifacts emerge from many interdependent actions, yet outcomes admit no reliable programmatic oracle. We introduce a continual adaptation framework in which a frozen frontier model operates professional design

中文介绍 专业平面设计是长程 agent 任务，产出可编辑结构化成果却缺乏可靠程序化评价标准。论文提出 Designer-RSI，让冻结的前沿模型持续适应设计工作流，从用户流量中演化程序性记忆，实现对真实使用分布下设计任务的持续改进，提升结构化设计产物的生成质量。

Grounded Action Model: 3D Grounding as a Foundation for Robotics

👍 21

09/20 08:00

Manipulation policies must know which objects matter and where they are, yet the pretrained backbones that current robot foundation models build on, from language in vision-language-action models (VLAs) to video generation in world-action models (WAMs), do not directly require this metric grounding,

中文介绍 现有机器人基础模型（VLA、world-action models）的预训练主干并不直接要求度量级的 3D 定位能力，但操作策略必须知道哪些物体重要及其位置。论文提出 Grounded Action Model，把 3D grounding 作为机器人操作的基础，使策略具备空间感知与定位，提升操作任务表现。

One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents

👍 16

09/20 08:00

Repository-level software engineering (SWE) comprises heterogeneous task categories, whose progress under pooled agentic reinforcement learning can be uneven: gains in some categories coincide with regressions in others, while aggregate resolution obscures these changes. Motivated by this category s

中文介绍 仓库级软件工程包含异质任务类别，统一池化的 agentic RL 会导致部分类别提升而另一些退化。论文提出类别感知的迭代专家训练方法，按任务类别训练并迭代整合专家，缓解类别间不平衡，在不牺牲单一类别的前提下提升 SWE agent 的整体解决率。

BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence

👍 15

09/16 08:00

Business intelligence (BI) is a cornerstone of enterprise decision-making and is widely used by enterprise users in software such as Power BI and Tableau. In traditional BI workflows, users need to prepare data by (1) identifying relevant tables, (2) performing data transformations, and (3) building

中文介绍 商业智能（BI）工作流中用户需手动选表、做数据变换并构建可视化。论文提出 BI-Agent 与 BI-Bench，面向端到端 BI 自动化：agent 自动完成表格识别、数据转换与图表构建，并配套基准评测，推动企业级 BI 流程从人工操作走向智能体自动执行。

Realtime-Venus: A full-duplex interaction system with asynchronous delegation

👍 13

09/12 08:00

Natural interaction in digital and physical environments requires continuous perception and timely responses. Spoken dialogue relies on acoustic and linguistic cues, while video interaction also requires grounding the conversation in evolving visual context. We present Realtime-Venus, a proactive fu

中文介绍 面向数字与物理环境的自然交互需要持续感知与及时响应。论文提出 Realtime-Venus 全双工交互系统，采用异步委托机制，同时处理语音与视频输入并将对话扎根于动态视觉上下文，实现主动式实时多模态交流，兼顾低延迟响应与视觉语境理解。

Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations

👍 13

09/06 08:00

Existing approaches to persona simulation with Large Language Models (LLMs) mostly rely on shallow character descriptions that fail to sustain coherent character behavior across extended interactions. We introduce Deep Persona, a psychologically grounded, three-layered architecture that organizes pe

中文介绍 现有 LLM 人格模拟多依赖浅层角色描述，难以在长程交互中保持行为一致。论文提出 Deep Persona，构建心理学基础的三层人格架构来组织角色属性，并配套评测框架，使角色扮演 agent 与仿真在长对话中维持连贯、可信的角色行为。

Harness-Zero: Harness Distillation via Agent-as-Harness

👍 12

09/21 08:00

Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent performance, but their gains remain tied to the harness at deployment. Because the best harness varies across domains, instances, and models, a general-purpose agent must either settle f

中文介绍 agent harness 能显著提升性能，但其收益绑定于部署时的具体 harness，而最优 harness 因领域、实例与模型而异。论文提出 Harness-Zero，通过 agent-as-harness 进行 harness 蒸馏，将 harness 能力内化到模型中，使通用 agent 无需依赖特定外部框架即可获得相近收益。

Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

👍 12

09/21 08:00

Agentic memory is becoming essential for long-horizon AI agents, yet many existing systems rely on autoregressive LLMs to control how memories are organized, retrieved, and used, placing expensive generation on the critical path of memory operations. We introduce \method, a new agentic memory archit

中文介绍 现有 agentic memory 多依赖自回归 LLM 控制记忆的组织、检索与使用，昂贵生成处于记忆操作关键路径上。论文提出 Jev-Mem，采用 System-One 控制的 agent 记忆架构，将记忆操作从生成中解耦，降低延迟与算力开销，提升长程 AI agent 的记忆效率。

CARE: Experience-Guided Atomic Corrective Execution for Vision-Language-Action Policies

👍 11

09/21 08:00

Vision-Language-Action (VLA) policies achieve strong performance in robotic manipulation but remain brittle once execution deviates from nominal trajectories. We propose CARE (Corrective Atomic Robotic Execution), a framework that improves recovery by learning from failures encountered during execut

中文介绍 Vision-Language-Action（VLA）策略在机器人操作中表现优异，但执行一旦偏离标称轨迹便十分脆弱。论文提出 CARE 经验引导的原子纠正执行框架，从执行中遇到的失败学习恢复策略，实现细粒度纠错，提升 VLA 在扰动与偏差情形下的恢复成功率。

1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

👍 9

09/21 08:00

Sparse on-policy distillation (OPD) allocates teacher supervision to a small subset of tokens in student-generated trajectories. However, useful teacher guidance can yield a noisy update when its gradient is estimated from a sampled next token. We study this estimation problem at a fixed prefix in i

中文介绍 稀疏 on-policy 蒸馏（OPD）仅对学生生成轨迹中的少量 token 施加教师监督，但基于采样下一 token 估计梯度会引入噪声。论文在固定前缀下研究该估计问题，分析梯度估计偏差与方差，提出改进方案，表明仅约 1% 的 token 即可提供有效监督。

ACLArena: Agent Continue Learning in Multi-stage Post-training

👍 8

09/21 08:00

Building general-purpose agents for industrial deployment requires integrating multiple capabilities, each typically acquired at a distinct stage of training. Yet there is currently no well-established recipe for Agent Continual Learning (ACL), with little understanding of the trade-offs among exist

中文介绍 工业级通用 agent 需整合在多个训练阶段分别获得的能力，但 agent 持续学习（ACL）尚缺乏成熟方案与权衡分析。论文提出 ACLArena，构建多阶段后训练中的 agent 持续学习评测与训练框架，系统比较不同策略在能力保持与迁移上的取舍。

Gricea: An Open Science Platform for Conversational AI Research

👍 8

09/18 08:00

We need studies on conversational AI (CAI) at scale to understand human behavior and shape CAI design. However, fragmented reporting of systems and study configurations hinders replication, extension, and knowledge accumulation. We present Gricea, an open-science platform representing studies as con

中文介绍 对话式 AI 研究需要大规模研究支撑，但系统与实验配置的碎片化报告阻碍复现与知识积累。论文提出 Gricea 开放科学平台，将研究表示为可组合配置，统一记录系统、被试与实验设置，支持复现、扩展与跨研究比较，促进对话式 AI 研究的规范化与累积。

MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup

👍 8

09/14 08:00

Scaling large language models efficiently has motivated sparse capacity mechanisms such as Mixture-of-Experts and, more recently, conditional memory: token-indexed embedding tables that augment the backbone with cheap parametric lookups. Existing memory-embedding methods retrieve via a deterministic

中文介绍 为高效扩展大语言模型，Mixture-of-Experts 与条件记忆（token 索引嵌入表）等稀疏机制受到关注，但现有记忆嵌入方法多采用确定性检索。论文提出 MoME，即 Mixture-of-Memory Embeddings，实现上下文感知的稀疏查找，使记忆访问更灵活，提升参数效率与语言建模表现。

Salvor v1.0.0-beta: I got tired of my projects forgetting what they learned, so I taught the repo to

@BlockchainDan · 18.5K 粉丝 · 3.2M 阅 · 501 赞 · 35 转

09/21 13:25

Soft-launching an open-source, repo-native engineering knowledge layer for coding agents. Built because of one very bad /clear. About a year into building an HFT tool with AI agents, I typed one

中文介绍 分享开源项目 Salvor v1.0.0-beta，为 coding agent 打造 repo 原生的工程知识层。作者做 HFT 工具约一年，因一次糟糕的 /clear 导致项目遗忘所学而开发，主打让 agent 记住仓库内的历史经验。

Jev is INSANE for Marketing

@dsqjaffa · 737 粉丝 · 217.8K 阅 · 510 赞 · 37 转

09/21 23:15

You can't escape it. Literally EVERYWHERE you look on your feed is filled with something about "jEv is iNsAnE"... and it's generated tens of millions of views on X in less than a week. I'm guilty of

中文介绍 复盘 Jev 营销为何爆火：一周内 X 上铺天盖地刷屏、累计数千万播放。作者分析这种病毒式传播的机制与打法，属营销案例拆解视角，非技术教程。

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 一篇观点思考：AI 让人更成功的同时，也将深刻冲击人际关系。作者自嘲已是「人肉代理」（meat proxy），探讨技术进步对亲密关系与友情的长期影响，偏哲学随笔。

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 反向工程 iMessage 助手 Instinct 的记忆机制，拆解它如何存储与调用上下文。作者称其为近期最好用的 iMessage 助手之一，属产品机制逆向分析，给做记忆层的开发者参考。

Jev + graphical models: a paradigm shift?

@fdellaert · 16.7K 粉丝 · 74.9K 阅 · 504 赞 · 61 转

09/21 08:19

Could Jev + graphical models fundamentally change how we build systems that reason under uncertainty? Unless you've been off the grid this week, you now know that Jev is @typesafeai’s new model for

中文介绍 探讨 Jev 与概率图模型（graphical models）结合的可能，认为这可能改变构建「不确定性推理系统」的方式。Jev 是 @typesafeai 的新模型，作者从学术建模角度切入，属前瞻性讨论。

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 入门指南：作者指出网上流传的示例多有误导，未能展示真实能力，于是整理一份更靠谱的科普。Jev 被称为通用 System One 模型，本周在 X 刷屏。

Jev Explained for Normies

@matthewcanham · 10.7K 粉丝 · 57.1K 阅 · 516 赞 · 45 转

09/22 00:46

If you’ve been hanging out on the internet the past few days, you’ve probably come across Jev, a new AI model that’s taken the AI community by storm. Jev is not a traditional LLM. It’s built for

中文介绍 面向普通人的 Jev 科普：解释这不是传统 LLM，而是为特定用途构建的新模型类型。适合想快速搞懂 Jev 到底特殊在哪、为何在 AI 社区爆火的人。

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：无需重训练，用开源 LLM 本地搭一个 Jev，做成快速本地决策引擎。涉及 next-token 打分、固定选项配概率分布、SGLang 等技术要点，主打 100% 本地运行。

Build a Jev Judge

@akshay_pachaar · 289.5K 粉丝 · 32.1K 阅 · 501 赞 · 59 转

09/22 01:26

We use LLMs to write answers, then call another LLM to judge them. But when the judgment is a handful of bounded decisions, do we need another round of text generation? An agent answers a customer

中文介绍 教程：用 Jev 思路做「裁判」（Judge）。当判断只是少数有界决策时，不必再走一轮文本生成。以 agent 回答客户为例，展示更省成本的判定实现方式。

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 介绍项目 murmur：不用 LLM，只用神经元，一群「果蝇」构成的链上集群，已在 Arc 主网（Chain ID 5042）用真实 USDC 结算，累计完成逾 23000 次任务，属链上自治实验。

🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science

Latent Space · 09/23 05:07

We talked to Google’s Oscar winning “Giganerd” about automating science, solving climate change, and how future generations can contribute to science in the age of superintelligent AI

中文介绍 Latent Space 对话谷歌研究员 John Platt，他曾获奥斯卡奖、两颗小行星以其命名，著有 sklearn 中广为使用的高斯过程等算法。访谈围绕用 AI 实现科研自动化、应对气候变化，以及超级智能时代后代如何参与科学展开。

Better prompt caching for GPT-6

OpenAI News · 09/23 05:00

Learn how GPT-6 improves prompt caching with higher cache hit rates, new diagnostics, explicit breakpoints, and controls that reduce latency and costs.

中文介绍 OpenAI 介绍 GPT-6 的提示缓存改进：更高的缓存命中率、新增诊断功能、显式断点以及多种控制选项，可降低延迟与成本。

Introducing GPT-6 Sol and Luna

OpenAI News · 09/23 02:00

Meet GPT-6 Sol and Luna, two models that bring frontier intelligence to everyday work with different balances of capability and cost.

中文介绍 OpenAI 发布 GPT-6 Sol 与 Luna 两款模型，将前沿智能引入日常工作场景，二者在能力与成本之间提供不同的平衡方案。

Roundtables: The Deadly Failures of The Virtual Border Wall

MIT Tech Review AI · 09/22 21:42

The US has spent billions building a “virtual wall” of surveillance towers along its southern border over the past 25 years, promising they will help detect and apprehend border crossers and save lives. But a groundbreaking investigation by MIT Technology Review has documented over a thousand people

中文介绍 过去 25 年美国斥资数十亿美元在南部边境建设由监控塔组成的「虚拟墙」，宣称可协助侦测与逮捕越境者并挽救生命。MIT Technology Review 的调查记录显示该计划存在致命失效。

Parallel cut research time and cost in half with GPT‑6 Astra

OpenAI News · 09/22 20:00

GPT‑6 Astra allowed Parallel’s agents to research and synthesize labor-market data in half the time and at half the cost vs. prior models.

中文介绍 Parallel 借助 GPT‑6 Astra，让智能体进行劳动力市场数据的调研与综合，相较此前模型将研究时间和成本各削减一半。

Don’t be fooled by this summer of AI hype

MIT Tech Review AI · 09/22 19:04

It’s been a busy few months for AI hype. At the end of April, Anthropic claimed that its model Claude Mythos is better at finding software vulnerabilities than most security experts. Then we had the OpenAI–Hugging Face hacking incident, after which Anthropic (proudly) and Meta (reluctantly) disclose

中文介绍 MIT Technology Review 提醒勿被今夏 AI 炒作误导：4 月底 Anthropic 称其模型 Claude Mythos 在发现软件漏洞上强于多数安全专家，随后发生 OpenAI–Hugging Face 黑客事件，Anthropic 与 Meta 也相继发声。

[AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M

Latent Space · 09/22 14:30

crowning a new Chinese frontier lab

中文介绍 小米发布 MiMo-V2.6-Pro 1T-A42B，据称训练成本约 300 万美元，成为新的顶级开源权重模型，标志着又一家中国前沿实验室崛起。

Priorities and principles for effective third party assessments

OpenAI News · 09/22 08:00

OpenAI outlines priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.

中文介绍 OpenAI 阐述其对前沿模型与安全防护措施开展严格、安全、独立第三方评估的优先事项与原则。

How UK AISI and EvalEval Are Making Benchmark Results Reproducible

Hugging Face Blog · 09/22 08:00

中文介绍 Hugging Face 博客介绍英国 AI 安全研究所（UK AISI）与 EvalEval 如何推动基准测试结果的可复现性。

Transformers now runs llama.cpp quants

Hugging Face Blog · 09/22 08:00

中文介绍 Hugging Face 博客宣布 Transformers 现已支持运行 llama.cpp 的量化格式（quants）。

Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community

Hugging Face Blog · 09/22 08:00

中文介绍 Hugging Face 宣布 oMLX 创建者兼维护者 Jun Kim 加入，以支持 MLX 社区。

Opus 5.5 imminent ⏳, Grok 4.7 🚀, MiMo v2.6 🤖

TLDR AI · 09/22 08:00

中文介绍 TLDR AI 当日要点：Opus 5.5 即将发布、Grok 4.7 与 MiMo v2.6 等模型动态。

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 播客对话 TypeSafe AI 首席执行官、Jev 主要创作者 Diogo Almeida，主题为面向生产环境而非「神级」的系统一模型。

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 博客提出以物理学视角剪枝大语言模型，将模块移除问题建模为伊辛优化问题。

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，一天内上线新的视频功能，让小型企业更易制作视频广告，并更快将新创意工具推向市场。