2026-09-25

The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

👍 119

09/22 08:00

LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agen

中文介绍 针对 LLM 智能体在长程任务中的决策质量问题，本文提出对「品味」（taste）的度量与改进方法，即选择测试哪条假设、基于哪个实现继续推进的能力。工作构建了评估基准并给出训练/推理侧的改进手段，证明提升决策品味能显著改善长程任务整体成功率。

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

👍 79

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 面向多方对话的长期记忆，SpeakerMem-R1 提出以说话人为中心的双轨记忆机制，除检索相关内容外，还建模谁说、涉及谁、彼此认知、群体共享信息与状态变化。通过双轨记忆结构与 R1 式训练，提升多方场景下的记忆准确性与角色一致性。

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

👍 43

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不足的问题，Spatial-Interactor 让模型通过与可观测物理世界交互来学习：感知物体运动与视角变化引起的局部状态转移，并在长轨迹上整合以维持更新的世界状态表示，从而提升动态场景下的空间推理与行动能力。

HappyWorld-Bench

👍 39

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 HappyWorld-Bench 是一个评估世界模型的综合基准，关注生成世界的质量，以及在探索、交互与修改下的一致性与响应性。它检验生成世界在被智能体操作后是否仍保持可靠，为世界模型的能力评测提供统一可比的测试集合。

The Past Frames the Future: Memory for Autoregressive Video Generation

👍 36

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 针对自回归视频生成中随序列延长出现的误差累积与一致性退化瓶颈，本文提出以「过去帧决定未来」为核心的内存机制，为 AR 视频生成引入历史帧记忆，支持长时程生成、交互式世界建模与演化视觉环境，改善了长视频的保真度与一致性。

Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

👍 36

09/21 08:00

In this report, we introduce Ovis-Embedding, a state-of-the-art omni-modal embedding family built on native integration of text, image, video, and audio. Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common repres

中文介绍 Ovis-Embedding 是一组全模态嵌入模型，不采用分离的模态塔，而是用共享多模态骨干把文本、图像、视频、音频编码到统一表示空间。该原生集成方式在跨模态检索等任务上达到领先水平，推动通用全模态嵌入的边界。

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

👍 33

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 现有智能体记忆多在写入时固化反思、工作流或技能，检索时难以适配新任务。本文提出 Just-in-Time Memory，让模型在任务执行时按需筛选与组织自适应记忆，实现任务自适应的记忆管理，提升 LLM 智能体复用过往经验的效果。

Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models

👍 27

09/21 08:00

Language models can be adapted by changing the computations applied to individual tokens. Quantum circuits offer one such approach, but evaluating wider circuits inside a large model can be computationally demanding. Here we introduce HyperQ, which adds token-conditioned quantum residual branches to

中文介绍 针对在大型模型中直接计算宽量子电路代价过高的问题，HyperQ 通过超网络为扩散语言模型注入以 token 为条件的量子残差分支，避免整体电路评估，从而在可控开销下获得量子增强的 token 级计算，提升语言建模表现。

Bellman Policy Optimization

👍 27

09/14 08:00

Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models (LLMs). We introduce Bellman Policy Optimization (BPO), a critic-free method derived from Policy Mirror Descent (PMD). For autoregressive generation with terminal rewards, BPO uses the

中文介绍 BPO（Bellman Policy Optimization）是一种无 critic 的 RLVR 方法，由策略镜像下降 PMD 推导而来。针对带终端奖励的自回归生成，它利用贝尔曼式目标构造 token 级优势或价值信号，在不训练价值网络的前提下提升 LLM 推理能力，兼顾效果与训练开销。

JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

👍 26

09/22 08:00

LLM-as-a-judge enables evaluation across diverse tasks, but inference cost and confidence reliability become critical at scale. We study whether a decision-only judge can provide an economical first pass and identify when stronger evaluation is needed. Comparing jev-as-a-judge with sixteen generativ

中文介绍 针对 LLM-as-a-judge 规模化时的推理成本与置信度可靠性问题，本文提出 jev-as-a-judge：仅做决策的轻量评判器作为经济的第一遍筛，在自信时直接接受、不确定时升级到更强评估。与 16 个生成式评判器对比，验证了该分层策略的性价比。

StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training

👍 25

09/22 08:00

Vector Quantization (VQ) is fundamental to discrete visual tokenizers that power modern autoregressive and masked image generation models. While recent shared-projection codebook methods have substantially advanced codebook utilization, training stability remains a critical and underexplored challen

中文介绍 向量量化是离散视觉 tokenizer 的核心，但训练稳定性长期缺乏系统研究，即便共享投影码本方法提升了码本利用率仍如此。StableVQ 总结出一套实用的稳定训练指南，涵盖码本更新与损失设计等要点，帮助自回归与掩码图像生成模型稳定获得高码本利用率。

From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health

👍 16

09/21 08:00

The rising global prevalence of mental health conditions, together with longstanding barriers in traditional healthcare, such as limited resources, high cost, stigma, and privacy concerns, has created an urgent need for accessible and scalable support. Large Language Models (LLMs) have emerged as a

中文介绍 该综述系统梳理 LLM 在心理健康领域的应用，背景是心理疾病患病率上升与传统医疗资源有限、成本高、污名化与隐私等障碍。文章将现有工作从模式识别工具到个性化陪伴分阶段归类，讨论评估、安全与伦理挑战，并展望可信可扩展的心理健康支持方向。

PACT: From Credit Assignment to Critic Alignment

👍 15

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 针对 LLM 后训练中 token 级信用分配缺乏公认数学定义的问题，本文提出完整性等三条正则性条件来刻画信用，并由此提出 PACT，将信用分配与 critic 对齐统一起来。分析澄清了常用训练信号与信用之间的关系，并给出改进的对齐方案。

Schrödinger's Code Repository: Have LLMs Learned SWE-bench or Memorized It?

👍 14

08/21 08:00

Repository-level coding benchmarks have become the standard for evaluating coding agents, yet they inherently suffer from data leakage because they are built upon popular open-source repositories repeatedly used for training. Consequently, strong performance may reflect memorization of canonical rep

中文介绍 仓库级编程基准多基于常被用于训练的流行开源仓库，存在数据泄漏，高分可能来自记忆而非真实能力。本文以「薛定谔的代码仓库」为切入，设计扰动与对照实验区分 SWE-bench 上的学习与记忆，揭示评测中记忆成分的影响并提出更可靠的评估思路。

Agensh: Scaling Organizational Intelligence to 1,024 Agents

👍 14

09/22 08:00

A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordi

中文介绍 针对多智能体系统受中心编排器任务分配与协调能力限制、难以扩展的问题，Agensh 提出去中心化多智能体编排框架，将组织智能扩展到 1024 个智能体。通过并发执行复杂任务，在降低延迟的同时维持协调效率，验证了大规模智能体协作的可行性。

Lean Pool: An AI-Maintained Archive of Formalized Mathematics

👍 14

09/21 08:00

Lean Pool is a repository of formalized mathematics. It is grown, maintained and optimized by AI agents.

中文介绍 Lean Pool 是一个由 AI 智能体生长、维护与优化的形式化数学仓库：智能体持续编写与整理 Lean 形式化证明并改进内容组织。该工作展示了 AI 自动维护大规模形式化数学库的可行路径，为形式化数学积累提供可持续的自动化方案。

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

👍 10

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 机器人装箱是长程序贯决策问题，单次放置会改变后续可用空间，现有方法多依赖人工几何启发式或从零学的 RL 策略。PackLab 提供开发、训练与评估 MLLM 的完整框架，用多模态大模型替代手工启发式，统一了装箱任务的训练与评测流程。

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

👍 10

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 现有 VLA 策略大多只看当前观测，缺乏对整段 episode 信息的保留与利用，在依赖历史的操作任务中表现受限。MemBodied 为 VLA 引入循环式联想记忆，保存并调用历史观测，从而在历史相关操作任务上提升决策一致性。

Hunyuan-A13B Technical Report

👍 8

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 Hunyuan-A13B 是开源 MoE 大语言模型，总参数 80B、推理仅激活 13B，在能力、算力效率与部署成本间取得平衡。模型基于大规模预训练数据训练，支持长上下文等能力，报告给出架构、训练流程与评测结果。

WhatWorkedBench: Benchmarking Experimental Understanding in AI Agents

👍 8

09/23 08:00

AI research agents need reliable knowledge of how their experiments change outcomes. We introduce WhatWorkedBench to measure experimental understanding, the accuracy of predictions about component changes after budgeted experimentation. Agents inspect code, select measurements, and submit a response

中文介绍 针对 AI 研究智能体是否真正理解实验因果的问题，WhatWorkedBench 提出「实验理解」度量：在有限预算实验后，预测改动某个组件会带来何种结果变化。智能体需阅读代码、选择测量并作答，以此检验其预测实验效果的准确性与稳定性。

How To Code Almost Forever for $20/Month With Codex + GPT-6 Luna

@sairahul1 · 147.5K 粉丝 · 785.8K 阅 · 506 赞 · 61 转

09/23 17:41

Yesterday OpenAI released GPT-6 Luna. And it might completely change the economics of AI coding. Here is what most people are missing. $20/month ChatGPT Plus plan with Codex now gives you access to

中文介绍 博主拆解如何用 20 美元/月的 ChatGPT Plus 搭配 Codex 与 GPT-6 Luna 实现低成本长期写代码。核心观点是这套组合正在改变 AI 编程的经济学，多数人没注意到 Plus 订阅已能直接用上 Codex。

How to automate SEO with Opus 5.5 (Full Course)

@EXM7777 · 146.5K 粉丝 · 316.1K 阅 · 552 赞 · 33 转

09/23 21:54

I'm going to walk you through the SEO & AEO setup i run with AI agents, from the lazy autopilot version to the full Claude Code build on Opus 5.5, step by step... because search changed shape a while

中文介绍 一份用 Opus 5.5 自动化 SEO 与 AEO 的完整教程。博主分两档讲解自己跑的 AI agent 方案：从最省事的 autopilot 版本，到基于 Claude Code 的全搭建流程，逐步演示，背景是搜索形态已发生变化。

I Think I’ve Been Using GPT-6 Astra Wrong

@BradGroux · 7.6K 粉丝 · 174.3K 阅 · 501 赞 · 5 转

09/23 01:49

I think I’ve been using GPT-6 Astra wrong this entire time. I kept giving it the same job I’ve given the strongest model available for the past several years: understand everything, plan everything,

中文介绍 博主反思自己一直用错 GPT-6 Astra：习惯性把最强模型当成「全懂全规划」的通用大脑。这条属于使用心得与踩坑分享，讨论应如何针对模型特点调整任务分工，而非沿用旧的工作方式。

The $4 Trillion Revolution: How AI Agents Will Rewire Finance and Spark an Economic Big Bang

@joechalom · 18.9K 粉丝 · 105.3K 阅 · 505 赞 · 113 转

09/23 20:00

The convergence of new onchain financial primitives and unlimited agent attention promises unprecedented value creation for consumers. The battle over control has begun. Over $1 trillion of annual

中文介绍 偏宏观论述：链上金融原语与 agent 的无限注意力交汇，可能为消费者创造空前价值，围绕控制权的争夺已开始。文中提到每年超 1 万亿美元规模，属于趋势预判而非教程。

How to train your own Jev for $17

@nutlope · 100.4K 粉丝 · 89.8K 阅 · 518 赞 · 35 转

09/24 06:02

TLDR: We just launched our own Jev-like classifier, together/Tev1-4B-experimental, on top of Qwen3.5 4B. In this blog post we’ll show you how to fine-tune your own version! Jev has quickly become one

中文介绍 博主发布自研 Jev 类分类器 together/Tev1-4B-experimental，基于 Qwen3.5 4B 微调，成本约 17 美元，并附博客教程讲解如何微调出自己的版本。对想低成本做垂直分类模型的人有参考价值。

Incubating The Horowitz Andreessen Academy

@eriktorenberg · 163.4K 粉丝 · 53.5K 阅 · 502 赞 · 55 转

09/22 22:18

The internet made it possible to know anything. AI makes it possible to do anything. We live in an age of infinite knowledge, and we are entering an age of infinite doing. One in which we are

中文介绍 博主宣布筹建 Horowitz Andreessen Academy，立意是互联网让「知道任何事」成为可能，AI 让「做到任何事」成为可能，人类正从无限知识时代进入无限行动时代。属于教育项目孵化的理念宣告。

Building a Custom Harness with Pi and Jev

@omarsar0 · 321.0K 粉丝 · 45.0K 阅 · 555 赞 · 46 转

09/23 22:09

An AI agent is a language model that works in a loop. It reads the task, uses a tool such as "read this file" or "delete that file", looks at the result, and keeps going until the job is done. Each

中文介绍 科普并实操如何用 Pi 和 Jev 搭建自定义 harness。开篇先把 agent 讲清楚：语言模型在循环里读任务、调工具、看结果、直到完成。适合想理解 agent 执行框架并动手搭一套的读者。

Gemini 3.8 text-to-speech says hello

@GoogleAIStudio · 205.4K 粉丝 · 41.9K 阅 · 519 赞 · 50 转

09/24 00:20

Today, we’re introducing two new text-to-speech models to the Gemini family, transforming voice generation from static presets into a dynamic creative studio. These models enable creators, developers,

中文介绍 Google 官方发布 Gemini 家族两款新文本转语音模型，把语音生成从静态预设升级为动态创作工作室，面向创作者与开发者。属于产品发布消息，具体能力与调用方式需看后续文档。

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，新增实时虚拟形象（Live Avatar）功能，可在实时交互中配合数字人形象使用，具体能力与上线范围未在摘要中说明。

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 客座文章指出，在科学研究中思考成本已大幅下降，但实验执行成本依然高企，这种不对称正在悄然重塑科研公司的运作方式。

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

Hugging Face Blog · 09/24 22:08

中文介绍 LiquidAI 在 Hugging Face 发布 LFM2.5-VL-DSpark 方案，聚焦视觉语言模型的加速，具体性能数据未在摘要中披露。

[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 Meta Connect 2026 大会发布 Muse 智能眼镜，并围绕语音、视频与 Charm 展开，Latent Space 称 Meta 团队表现强势。具体产品参数与售价未在摘要中披露。

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

TLDR AI · 09/24 08:00

中文介绍 TLDR AI 当日要点包括：Gemini 推出文本转语音、Claude 应用于新型酶研究、Google 上线私有记忆功能，具体细节未在摘要中披露。

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 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，具体性能提升数据未披露。

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 Private AI Compute 引入安全的服务端私有记忆功能，用于个人 AI 场景，在服务端实现数据私有化存储与处理。

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，称其持续向更多社区提供 AI 技能培训。具体参与人数与地区未在摘要中披露。

Gemini 3.8 text-to-speech says hello

DeepMind Blog · 09/23 23:25

中文介绍 DeepMind 发布 Gemini 3.8 文本转语音功能，具体音色、语言支持与开放范围未在摘要中说明。

🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

Latent Space · 09/23 21:27

Radical Numerics is using biological chain-of-thought and multimodal perception to keep up with the bio-defense arms race, design new genomes and gain insights into biology itself.

中文介绍 Radical Numerics 首席执行官 Eric Nguyen 表示，生物安全是一场 AI 军备竞赛。该公司利用生物链式思维与多模态感知，推进生物防御、设计新基因组并加深对生物学本身的理解。

OpenAI extends cyber access to Ukraine for civilian defense

OpenAI News · 09/23 21:00

OpenAI is extending access to its Daybreak program to the Government of Ukraine to support the cyber defense of civilian infrastructure.

中文介绍 OpenAI 宣布将其 Daybreak 计划的使用权限扩展至乌克兰政府，用于支持民用基础设施的网络防御。

Sam Altman’s remarks at the United Nations Security Council

OpenAI News · 09/23 20:00

OpenAI CEO Sam Altman discusses AI safety, human control, and international cooperation in remarks to the United Nations Security Council.

中文介绍 OpenAI 首席执行官 Sam Altman 在联合国安理会发言，讨论 AI 安全、人类对 AI 的控制以及国际合作等议题。

Harvey turns legal context into stronger drafts with GPT-6 Astra

OpenAI News · 09/23 20:00

GPT-6 Astra produces more structured, context-aware legal documents, freeing lawyers to focus on strategy.

中文介绍 法律 AI 公司 Harvey 借助 GPT-6 Astra 生成结构更清晰、更具上下文感知的法律文件，使律师可专注于策略性工作。具体效率提升数据未披露。

How invideo improves color grading 3x with GPT‑6 Astra

OpenAI News · 09/23 20:00

With GPT‑6 Astra, invideo plans edits with greater precision, improves color correction and grading threefold, and produces 50 custom effects in one day.

中文介绍 视频平台 invideo 借助 GPT-6 Astra 提升剪辑规划精度，调色与色彩校正效率提升 3 倍，并可在一天内产出 50 个自定义特效。

Ringg’s AI agents resolve up to 65% of customer calls with OpenAI

OpenAI News · 09/23 20:00

Using GPT-5.6, Ringg powers multilingual agents across voice, chat, WhatsApp, and web for 90% less cost vs. GPT-4.1.

中文介绍 Ringg 基于 GPT-5.6 打造多语言 AI 客服代理，覆盖语音、聊天、WhatsApp 与网页渠道，可解决最多 65% 的客户来电，成本较 GPT-4.1 降低 90%。