2026-09-08

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

👍 166

09/03 08:00

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, a

中文介绍 大型语言模型在扩展推理时KV缓存成为内存瓶颈。现有压缩方法均按重要性评分逐token淘汰，本文提出随机注意力机制，重新审视缓存逐出策略，以提升推理效率并降低内存占用。

Knowing When Not to Reuse: Conditional Experience Transfer in Autonomous LLM Post-Training

👍 149

08/27 08:00

Large language models offer broad capabilities, but adapting them to evolving domains, tools, and requirements often entails repeated post-training. Autonomous systems automate parts of this process by proposing updates, training candidates, and using evaluation feedback to select subsequent proposa

中文介绍 针对自治LLM后训练中反复适配新领域、工具导致的重复低效，提出条件性经验迁移框架，仅在适当时机复用历史经验，自主判断何时不重用，从而提升后训练效率和模型性能。

Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems

👍 94

09/02 08:00

Multi-agent LLM systems commonly use an orchestrator to decompose a task for a team of workers and then improve through textual reflection. Despite strong empirical results, these systems lack a unified account of coordination, memory improvement, and the role of external verification. We model orch

中文介绍 多智能体LLM系统常用编排器分解任务并通过文本反思改进，但缺乏统一理论。本文基于博弈论建模双层协调反思，统一协调、记忆改进与外部验证，提升多智能体协作效果，提供系统化分析框架。

Rethinking On-Policy Distillation of Large Language Models II: One Training Example

👍 81

09/03 08:00

On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. Existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. We examine this role at the data-minimal limit by training on a single query. One-

中文介绍 本文在数据最少极限下，仅用单个查询训练研究在线策略蒸馏中训练数据的作用。单样本条件下，该过程显示出特定行为，揭示OPD对数据依赖的本质，为理解蒸馏算法提供新视角。

Iris: Climbing to the Search Frontier

👍 50

09/03 08:00

We present Iris-mini and Iris-pro, two search agents trained at the 35B-A3B and 397B-A17B scales, together with the data pipeline and training recipe behind them. Tasks are reverse-constructed from the hyperlink structure of a web corpus: we author multi-hop chains over an entity graph distilled fro

中文介绍 提出Iris-mini和Iris-pro搜索智能体，规模达35B-A3B和397B-A17B。通过网页超链接结构反向构造多跳任务，在实体图上训练，并配以数据管道和训练配方，提升搜索前沿能力。

Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction

👍 46

09/03 08:00

Online 3D reconstruction models perform poorly on long videos. This happens because regressing poses relative to a fixed first-frame anchor forces extrapolation far beyond the training distribution. Small drifts accumulate and amplify into significant geometric collapse. However, we observe that per

中文介绍 在线3D重建在长视频上性能差，因相对固定首帧的位姿回归外推超出训练分布导致误差累积。提出多相对位姿查询，促进可扩展在线重建，缓解几何崩塌，提升长视频重建精度。

Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue

👍 35

08/28 08:00

An avatar that holds a conversation should decide what to say and to move while saying it, yet these abilities live in separate model families: spoken dialogue models produce speech without motion, and co-speech motion models produce motion only from audio handed to them. The standard remedy is a ca

中文介绍 对话式说话内容与肢体动作分散在不同模型家族，缺乏统一。提出Motion-Omni端到端模型，联合生成语音与全身动作，实现自然口语对话，无需级联管线，提升交互协调性。

Beyond Retrieval: Progressive Latent Memory Evolution for Streaming Video Understanding

👍 30

09/03 08:00

Streaming video understanding requires multimodal large language models (MLLMs) to process continuous visual inputs and respond to user queries under strict causality and bounded memory. Existing approaches typically compress historical observations into an external memory bank and retrieve query-re

中文介绍 流式视频理解受严格因果和内存限制，现有检索法压缩历史到外部存储。本文提出渐进潜在记忆演化，超越简单检索，在流式处理中动态更新记忆，提升长视频查询响应的准确性和效率。

RealSWE: A Compositional Evaluation of Coding Agents under Realistic User Requests

👍 29

08/31 08:00

Coding agents are now commonly evaluated on the SWE-bench family of benchmarks, whose tasks are built from curated GitHub issues: long, structured, and information-rich. Real user requests, however, are typically far shorter and less structured. To characterize this gap, we define a six-category inf

中文介绍 SWE-bench任务源自结构化GitHub问题，真实用户请求短且非结构化，存在差距。提出RealSWE组合评估基准，定义六类信息缺口，用于评测编码智能体在真实用户请求下的性能，填补现有基准不足。

The Attention Triangle in Audio-Video Models

👍 25

09/03 08:00

Audio-video diffusion models rely on cross-modal attention to coordinate text, sound, and visual content, yet this same mechanism can introduce subtle and systematic semantic leakage. We study these models by probing and analyzing the ``attention triangle,'' comprising the three cross-attention edge

中文介绍 音视频扩散模型依赖跨模态注意力协调文本、声音和视觉，但可能引入系统性语义泄漏。本文探测分析音频-视频-文本三者之间的注意力三角，揭示泄漏机制，为改进模型设计提供洞见。

DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

👍 25

09/03 08:00

Reinforcement Learning from Verifiable Rewards works well when a task has a programmatic checker, but most long-horizon agent domains have none. We work in the outcome-blind setting, where ground-truth success signals are not available. Multi-criteria rubrics are a popular way to supply such a rewar

中文介绍 长期任务智能体缺乏可验证奖励，多标准评分常用于提供信号。提出DRACO方法，基于动态评分标准进行细粒度信用分配，在无真实结果的环境中高效训练，提升有效监督。

CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

👍 25

09/03 08:00

MLLM-based embedding models remain limited in compositional retrieval, often failing to distinguish scenes containing the same concepts but different attribute-object bindings. Yet the same backbone can resolve such distinctions when used as a cross-attentive reranker, motivating us to distill its c

中文介绍 基于MLLM的嵌入模型在组合检索上有限，无法区分同概念不同属性绑定。利用同一骨干作为交叉注意力重排器可解决，提示将其知识蒸馏到嵌入模型，提出CORE方法，提升组合推理能力。

WorldReward: Reward Modeling for Camera-Conditioned World Models

👍 24

09/03 08:00

Camera-conditioned world models generate interactive videos in which commanded actions should induce the expected scene changes while appearance, geometry, and temporal dynamics remain coherent. Existing rewards assess these requirements separately: geometry-based rewards estimate trajectory executi

中文介绍 相机条件世界模型生成交互视频，需要动作与场景变化一致。现有奖励割裂几何与视觉连贯性，本文提出WorldReward统一建模奖励，同时评估轨迹执行与外观、几何、动态一致性，提升模型训练。

PACE: Towards Surfacing Hidden Conflicts in User Requests

👍 23

09/03 08:00

Personalized assistants should not only comply with user requests but also assess whether those requests are appropriate given the user's current circumstances. However, prior work has primarily focused on accurately executing requests, overlooking the need for assistants to account for context and

中文介绍 个性化助手需评估用户请求是否适合当前情况，现有研究仅注重执行。提出PACE方法，旨在揭示用户请求中的潜在冲突，结合上下文情境，提升助手决策的适当性与安全性。

Using Grounded Theory for Agent Behavior Analysis at Scale

👍 19

08/31 08:00

Understanding agent behavior requires methods that scale to thousands of trajectories and surface new patterns in long, often unfamiliar tasks where pre-built classifiers fall short. We propose to bring grounded theory into agent trajectory analysis: a six-decade-old qualitative method from the soci

中文介绍 分析智能体行为需扩展到数千轨迹并发现新模式。本文引入社会学扎根理论用于轨迹定性分析，系统性强，能处理长期未知任务，超越预构建分类器，提供可扩展的行为分析方法。

Enoki: Efficient Multi-Level Hallucination Detection

👍 16

09/01 10:21

Ensuring factuality remains a critical challenge for deploying LLMs in high-stakes settings. Existing hallucination detectors usually operate at a single level: claim-level methods provide interpretable factual units, while span-level methods localize unsupported text. Bridging these views is costly

中文介绍 现有幻觉检测通常单层，声明级可解释但跨度级需定位文本。提出Enoki高效多级幻觉检测，桥接声明与跨度，成本更低，在高风险场景提高事实性保障，实现可解释与定位兼顾。

Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

👍 15

09/04 08:00

Large language models (LLMs) are increasingly used to formulate optimization models from natural-language problem descriptions, yet realistic operations research (OR) requests are often incomplete: missing objectives, constraints, or business rules can change the resulting mathematical program. Exis

中文介绍 LLM从自然语言制定优化模型常遇需求不完整，导致目标或约束缺失。本文提出交互优化中先询问再动态预表述澄清，识别并补全缺失信息，提高数学模型准确性，增强实用性。

Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference

👍 15

09/04 08:00

Layer dropout (a.k.a. stochastic depth) has been shown to enable faster training, higher accuracy, and robustness to zero-shot layer pruning in both language and vision transformers. However, as models and datasets have scaled, dropout - particularly layer dropout - has largely disappeared from larg

中文介绍 层丢弃（随机深度）曾带来更快训练和更高准确率，但大型模型已少用。本文研究层稀疏优化，证明不丢弃丢弃可在训练和推理中提升效率与性能，恢复该方法在大模型中的价值。

Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs

👍 15

09/03 08:00

Long-video language models cannot look at every frame: an hour sampled once per second is 3,600 images, and a system keeps only a small fixed slice of that pool. Which frames survive that slice is usually treated as a preprocessing detail; we test whether it should be. Published selectors make the c

中文介绍 长视频模型不能看每帧，常保留固定帧子集。视觉token分配常被视为预处理细节，本文控制性研究选择、压缩、重投资，检验选择器影响，优化视觉token分配，提升长视频理解性能。

QCell: Recombining and Aligning Cell Queries for Overlapping Instance Segmentation

👍 14

08/29 08:00

Instance segmentation of overlapping cells in microscopy remains challenging due to semi-transparent structures that produce weak boundaries and mixed visual evidence in overlap regions. Existing methods address this through local regions of interest or shape priors but lack global reasoning across

中文介绍 显微图像中重叠细胞半透明致弱边界，现有方法缺乏全局推理。提出QCell方法，通过重组和对齐细胞查询，实现重叠实例的精确分割，有效处理弱边界与混合证据，提升效果。

Claude Code → Codex (and keep everything you already built)

@charliejhills · 14.1K 粉丝 · 120.8K 阅 · 504 赞 · 53 转

09/06 18:44

For months people have been telling me Codex is better than Claude Code. I ignored all of it. I have built my whole business inside Claude Code (every pipeline, every system) and I did not fancy

SARVAM AI Interview Experience and Prep - Backend Engineer (On-Device AI) Internship:

@buzzy_bit · 767 粉丝 · 97.5K 阅 · 500 赞 · 20 转

09/05 15:30

Disclaimer: This post is written purely to share my experience and the way I prepared. It is not a leaked question bank, I'm intentionally not reproducing exact interview questions, and I'm not naming

Grok Bot Marketplace is live!

@ericzakariasson · 87.8K 粉丝 · 58.2K 阅 · 522 赞 · 49 转

09/07 19:05

We launched the Grok Bot Marketplace last week on Grok Bot · Bot Marketplace! It's a shelf of teammates other people already built for real jobs. Pick one, import it into your sidebar, and you're

The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it)

Latent Space · 09/08 05:32

Our first Astra project dives into AEO trends, a top asked topic from founders and DX leaders we talk to.

中文介绍 Latent Space 发布首个 Astra 项目，聚焦 AEO（AI 引擎优化）趋势，此为创始人和 DX 领导者最常询问的话题之一。

Supporting independent journalism in Ukraine

OpenAI News · 09/07 08:00

OpenAI, AIRPPU and WAN-IFRA launch an AI program to help Ukrainian news organizations strengthen innovation, resilience, and independent journalism.

中文介绍 OpenAI 与 AIRPPU 及 WAN-IFRA 合作，在乌克兰启动 AI 项目，以帮助当地新闻机构增强创新、韧性和独立新闻报道能力。

Claude proves Fermat 🧮, automated AI researcher 🔬, Z1 efficiency chip ⚡

TLDR AI · 09/07 08:00

中文介绍 TLDR AI 报道：Claude 证明费马大定理、自动化 AI 研究员出现、Z1 高效芯片问世。

An Alien Mind

OpenAI News · 09/06 17:00

Jakub Pachocki reflects on increasingly capable AI and the challenge of keeping it aligned. He calls for stronger safeguards and international coordination.

中文介绍 OpenAI 研究员 Jakub Pachocki 撰文反思日益强大的 AI 及其对齐挑战，呼吁加强安全防护和国际协作。

Research acceleration: The view inside OpenAI

OpenAI News · 09/06 16:00

Inside OpenAI, coding agents are reshaping AI research. Explore early data on agent usage, experiment velocity, task complexity, and research acceleration.

中文介绍 OpenAI 内部数据显示，编码代理正重塑 AI 研究，加速实验进程、提升任务复杂性和研究速度。

OpenClaw Power, MacBook Simplicity: Five Days With Grok Bot

Latent Space · 09/05 23:01

SpaceXAI’s Grok Bot has the same level of programming power as OpenClaw, but it’s programmable at a different level of abstraction.

中文介绍 Latent Space 评测 Grok Bot：其编程能力与 OpenClaw 相当，但抽象层级不同，兼具 MacBook 的简洁性。

Architecting memory and storage in the AI era

MIT Tech Review AI · 09/05 02:39

The era of AI inference has arrived. Imagine a healthcare system analyzing millions of data points in real time to accelerate life-saving medical research, or an intelligent assistant instantly resolving thousands of complex customer needs at once. These real-world breakthroughs rely on advanced inf

中文介绍 MIT 科技评论探讨 AI 推理时代的存储架构，医疗系统可实时分析数百万数据点，加速医学研究，智能助理可同时处理海量客户需求。

Data from drones in Ukraine is fueling a new Wild West marketplace

MIT Tech Review AI · 09/04 17:25

Battlefields in Ukraine are littered with the remnants of drones, which are now firmly established as a critical weapon of modern warfare. But behind all that wreckage, there’s a new gold mine for the defense sector. The data drones generate will far outlast the wars in which they are used to fight,

中文介绍 MIT 科技评论：乌克兰战场无人机残骸背后的数据成为国防领域新金矿，催生数据交易的新兴市场。

[AINews] GPT-6 Astra: OpenAI’s biggest LLM launch of all time

Latent Space · 09/04 13:18

new SOTA computer use and coding, 2.5x pricier per token, but WAY cheaper per task, less monitorable. overall, a very successful launch of OpenAI’s new frontier model class.

中文介绍 AINews 报道 GPT-6 Astra 发布，为 OpenAI 史上最大规模 LLM 发布，在计算机使用和编码上达到新 SOTA，每 token 价格贵 2.5 倍，但每任务成本更低。

GPT-6 Astra 🚀, Grok Bot Enterprise 🤖, Runway world model 🌎

TLDR AI · 09/04 08:00

中文介绍 TLDR AI 简报：GPT-6 Astra 发布、Grok Bot 企业版推出、Runway 世界模型亮相。

GPT-6 Astra: an automated AI Engineer you can hire for <$6 an hour

Latent Space · 09/04 05:09

We spent 20B+ tokens of GPT-6 Astra to explore everything. Here’s our learnings.

中文介绍 Latent Space 花费 200 亿+ token 测试 GPT-6 Astra，称其为可雇佣的自动化 AI 工程师，时薪低于 6 美元。

Introducing WeatherNext 3, our most advanced and accurate global weather AI model

DeepMind Blog · 09/03 23:02

中文介绍 Google DeepMind 推出 WeatherNext 3，宣称是其最先进、最准确的全球天气 AI 模型。

Daybreak for Frontline Defenders: $1B to protect essential services

OpenAI News · 09/03 21:15

OpenAI introduces Daybreak for Frontline Defenders. A $1 billion commitment expands access to frontier cyber AI, training, and support for essential services.

中文介绍 OpenAI 推出 Daybreak for Frontline Defenders 计划，承诺投入 10 亿美元扩大前沿网络 AI 及培训，保护关键服务。

NeoMME: an efficient Multimodal-native and Multilingual Encoder

Hugging Face Blog · 09/03 21:13

中文介绍 Hugging Face 博客发布 NeoMME，一种高效的多模态原生及多语言编码器。

Legora reviewed 41 documents in minutes with GPT-6 Astra

OpenAI News · 09/03 20:00

Legora used GPT-6 Astra to review 41 documents in minutes, find all four planted errors, and improve performance by nearly 40% in this financial-review workflow.

中文介绍 Legora 使用 GPT-6 Astra 在数分钟内审阅 41 份文件，找出全部四个预设错误，财务审查效率提升近 40%。