2026-08-06

MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations

👍 83

07/31 08:00

Large language model agents are increasingly evaluated as autonomous tool users, yet most benchmarks focus on bounded tasks with immediate success criteria. Real-world deployments often require Long-Term Coherence, the capacity to preserve purposeful behavior across extended horizons while adapting

中文介绍 提出 MerchantBench 基准，评估电商场景中 LLM 智能体的长期连贯性（Long-Term Coherence）。现有基准多聚焦有界任务和即时成功标准，而真实部署需要智能体在长时间跨度中保持目的性行为并适应环境变化。该基准为自主工具使用的现实部署评估提供新测试平台。

AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling

👍 73

08/03 08:00

Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generatio

中文介绍 针对文本生成仍依赖离散 token 的问题，提出 AURORA-LM，将连续潜空间扩散语言建模与自编码统一表征相结合。通过设计适合联合生成的连续表征空间，替代继承自离散模型的嵌入空间，使语言模型在统一连续潜空间中生成，弥合文本与其他模态在生成建模上的差距。

InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

👍 56

08/03 08:00

Single-image feed-forward 3D Gaussian Splatting (3DGS) aims to directly generate a renderable 3D scene representation from one input image, avoiding the cost of multi-view capture and per-scene optimization. However, existing methods are often constrained by a pixel-aligned representation, where Gau

中文介绍 提出 InfiniSplat，用于大基线单目视图合成的隐式高斯解码方法。现有前馈 3DGS 受限于像素对齐表示，难以处理大基线视角。该方法通过隐式解码生成 3D 高斯场景表示，摆脱像素对齐限制，可从单图直接生成可渲染场景，提升视图合成泛化能力。

Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

👍 45

08/04 08:00

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bia

中文介绍 提出 Video-DeepResearch，将多模态深度研究智能体从静态图像扩展到连续视频流，要求密集时空定位与开放网络探索相结合。初步评估发现现有模型存在模态偏置等关键瓶颈，该工作为视频流场景下的深度研究智能体奠定基础。

PCSD: Persistent Consistency for Self-Distillation in Agentic Reinforcement Learning

👍 37

08/03 08:00

Large language model agents have shown strong potential in complex interactive tasks, yet their reinforcement learning (RL) is often hindered by sparse rewards, as a long multi-turn trajectory may receive only a single outcome-level signal. On-policy self-distillation (OPSD) provides dense token-lev

中文介绍 提出 PCSD，通过持久一致性约束改进智能体强化学习中的自蒸馏。针对长轨迹仅获得单一稀疏奖励的问题，OPSD 虽提供稠密 token 级信号，但训练不稳定；PCSD 引入持续一致性机制稳定自蒸馏过程，使智能体在复杂交互任务中获得更可靠的学习信号。

Quo Vadis, World Modeling?

👍 29

08/03 08:00

Continually improving agents require dynamic interaction feedback beyond static supervision, yet direct real-environment interaction is costly, slow, unsafe, and hard to parallelize. World modeling offers a natural intermediate proxy that allows agents to query lower-cost, more controllable feedback

中文介绍 围绕世界模型展开探讨，指出真实环境交互成本高、速度慢且难以并行，而世界模型可作为低价、可控的中间代理，让智能体获得动态交互反馈。论文分析现有挑战，为构建可持续改进智能体的世界模型研究提供方向性思考。

PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents

👍 28

08/04 08:00

Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained exp

中文介绍 提出 PAST-Bench 基准，评测个人 AI 智能体的递归自我改进能力。此类智能体需将累积的偏好、任务历史、工具例程和技能转化为未来更好的行为。该基准检验保留经验是否真正提升后续任务表现，为个人智能体持续学习提供评测基础。

LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models

👍 24

08/04 08:00

Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE d

中文介绍 系统研究混合专家扩散语言模型（MoE dLLM）的扩展行为。该工作刻画优化超参数、算力分配与架构规模对模型性能的影响，填补 MoE 扩散语言模型缩放规律认识的空白，为构建更大规模扩散语言模型提供实践指导。

OmniPack: Unified Token Compression for Efficient Omni-modal Large Language Models

👍 24

08/04 08:00

Omni-modal large language models (Omni-LLMs) have achieved remarkable performance on audio-visual understanding tasks, but processing long and highly redundant visual and audio token sequences incurs substantial computational overhead, demanding aggressive token compression for efficient deployment.

中文介绍 提出 OmniPack，面向全模态大语言模型（Omni-LLM）的统一 token 压缩方法。针对视觉和音频 token 序列长且冗余、计算开销大的问题，OmniPack 通过统一压缩机制有效降低序列长度，在保持音视频理解性能的同时提升部署效率。

CAPEval: A Decoupled Caption Evaluation across Understanding and Generation

👍 21

08/03 08:00

Captions serve as a primary supervision signal for both multimodal understanding and text-to-image generation. However, previous evaluations treat the caption quality as a single scalar objective, which conflates two distinct properties: (1) how much visual information a caption covers and (2) how r

中文介绍 提出 CAPEval，将标题（caption）质量解耦为视觉信息覆盖度与生成质量两个维度分别评估。此前方法用单一标量衡量标题，混淆了理解与生成中的不同属性。该框架为多模态理解和文生图提供更细粒度的标题质量评测。

MemSFT: Mitigating Alignment Tax with an External Parametric Memory

👍 21

07/28 08:00

Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain spec

中文介绍 提出 MemSFT，利用外部参数化记忆缓解领域适配中的对齐税（alignment tax）。针对专用领域微调导致灾难性遗忘、降低通用任务性能的问题，MemSFT 将领域知识与原有参数解耦，减少微调对通用能力的损害。

SkillJack: Persistent Skill Backdoors in Self-Evolving Agents

👍 19

08/04 08:00

Self-evolving agents increasingly convert interaction histories into reusable skills that persist beyond individual tasks. While prior work studies memory and retrieval poisoning, such attacks only affect agents when poisoned records are retrieved as context. We uncover a new and more fundamental ri

中文介绍 揭示自进化智能体面临的新型安全风险 SkillJack：持久技能后门。与仅在恶意记录被检索时触发的记忆投毒不同，技能后门随智能体将交互历史转化为可复用技能而持续存在，即使原始记录不再出现也可能被激活，构成更深层威胁。

UniWorld-Design: From Pixel Generation to Layer-Native Design

👍 19

08/04 08:00

We introduce UniWorld-Design, a framework that redefines image generation from flat pixel synthesis to structured visual composition, with semantic RGBA layers as the atomic units of generation, understanding, and editing. Our key insight is that pixels define how an image is rendered, whereas layer

中文介绍 提出 UniWorld-Design，将图像生成从平面像素合成重构为以语义 RGBA 层为原子单元的视觉组合。该方法把理解、生成和编辑统一到分层表示上，使模型基于图层而非像素设计，实现更可控、可编辑的结构化图像生成。

To Add Is Machine, To Delete Is Human: Measuring and Mitigating Deletion Avoidance in LLM Code Editing

👍 19

07/30 08:00

Large language models increasingly write and repair production code, yet evidence is mounting that their test-passing patches leave codebases harder to maintain. We identify one concrete source: deletion avoidance, the systematic tendency to retain code that an intended edit requires removing. Acros

中文介绍 识别并度量 LLM 代码编辑中的“删除回避”现象：模型系统性地保留本应删除的代码，导致补丁通过测试却降低代码库可维护性。论文提出量化指标与缓解策略，改善模型在真实代码编辑中的删除能力。

TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning

👍 17

08/04 08:00

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation off

中文介绍 提出 TurnSight，面向工具集成推理的回合级后见自蒸馏方法。针对现有强化学习依赖轨迹级监督、难以细粒度分配信用的问题，TurnSight 在每回合层面进行后见自蒸馏，为长程工具交互提供更精细的学习信号。

GROVE: Growing and Reasoning over Temporally Stratified Memory from Streaming Video Experience

👍 14

08/03 08:00

A wearable assistant should both answer questions about its visual history and recognize when that history is useful to the present situation. Existing video-memory systems primarily support question-conditioned recall, whereas proactive assistants typically use separate memory and control mechanism

中文介绍 提出 GROVE，面向可穿戴助手流式视频经验的时间分层记忆系统。现有系统仅支持按问题回忆且记忆与控制分离；GROVE 将记忆构建与推理统一，在记录视觉历史的同时识别其对当前情境的适用性，支持主动辅助。

ExplainBench: Evaluating Code Explanations from Agents

👍 12

07/29 08:00

Large Language Model (LLM) agents have seen rapid adoption in software engineering. As agents take a greater role in the actual generation of code, they are making larger changes, spanning tens to hundreds of lines. This makes manual review of agent results increasingly infeasible, leading developer

中文介绍 提出 ExplainBench，系统评估 LLM 智能体生成的代码解释。随着智能体改动规模增至数十至数百行，人工审查愈发不可行，高质量解释成为开发者理解代码的关键。该基准为可解释代码生成提供评测标准。

AgentStream: How Well Do Self-Evolving LLM Agents Perform Under Streaming Tasks?

👍 12

07/31 08:00

Large language model (LLM) agents can self-evolve by continually improving from their own accumulated experience. However, existing studies predominantly adopt independent evaluation. Consequently, the behavior of self-evolving agents in realistic streaming settings, where agents adapt to diverse an

中文介绍 研究自进化 LLM 智能体在流式任务下的表现。现有评估多采用独立任务，忽略真实场景中任务连续到来、智能体需从经验中不断进化的特点。AgentStream 提供流式任务评测框架，考察智能体在多样化持续任务中的适应与进化能力。

ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities?

👍 11

08/04 08:00

Modern agent frameworks equip large language models with external skill libraries to solve complex tasks. However, it remains unclear whether these systems can effectively evolve their skills and whether the resulting skills improve task-solving capabilities. To bridge this gap, we introduce Continu

中文介绍 提出 ContinualSkillBench，检验 LLM 智能体能否真正进化技能。现有框架虽配备外部技能库，但技能是否随经验有效演进、能否切实提升任务解决能力仍不明确。该基准系统评估技能库的持续演化与下游任务收益。

Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements

👍 9

07/22 08:00

Do Large Language Models (LLMs) possess genuine structural reasoning, or merely rely on surface-level pattern matching? The financial domain, demanding numerical precision and multi-step logic over long contexts, is an ideal testbed. Existing benchmarks fail to capture real-world industrial complexi

中文介绍 在长周期财务语句上检验 LLM 推理的可信度，探究其具备真正的结构化推理还是依赖表层模式匹配。现有基准未涵盖真实工业场景复杂度；该测试关注数值精度和多步逻辑，评估财务领域长上下文推理能力。

I went on paternity leave. My agent built a $3M pipeline from 2,741 conversations. Here's the code.

@wayne_liang_ · 1.5K 粉丝 · 999.9K 阅 · 637 赞 · 279 转

08/04 00:47

For eight weeks, a clone of me ran our sales front line: 300+ prospect calls a week, 132 new paying customers, 37 enterprise opportunities worth ~$3M. It also made up a price, emailed internal notes

中文介绍 八周陪产假期间，一个AI克隆体接管销售前线：每周300+潜在客户电话，新增132个付费客户，37个企业机会约300万美元。同时也会编造价格、发内部笔记，暴露代理失控风险。

Grok Build will rewrite how you use your laptop -A practical guide to your own Autonomous Employee

@XFreeze · 251.7K 粉丝 · 778.3K 阅 · 558 赞 · 80 转

08/05 20:12

Grok Build: What It’s Actually For Most people underestimate it by an order of magnitude. Ask someone what an AI coding agent is for, and they’ll probably tell you it builds apps. That’s true and it

中文介绍 多数人低估了 Grok Build：它不只是写应用的编码代理。本指南教你将 Grok Build 作为真正的「自主员工」部署，接管完整任务流程，改变笔记本的使用方式。

Taste, Judgment and AI

@addyosmani · 407.9K 粉丝 · 363.3K 阅 · 510 赞 · 45 转

08/04 03:03

Taste is recognizing quality and greatness in the absence of all the evidence. Judgment is committing to it in the presence of some risk. It’s the ultimate foundation of great work. When you look at a

中文介绍 品味是在证据不全时识别质量与伟大，判断力是在存在风险时敢为此承诺，这是伟大工作的最终根基。当 AI 接管执行环节，人类的品味与判断力成为最稀缺的资产。

The Invisible Economy

@RaoulGMI · 1.5M 粉丝 · 90.0K 阅 · 530 赞 · 72 转

08/05 00:38

Within about two years, most economic activity on Earth won't have a human anywhere in it. Think about a single trade today. Someone decides to make it, another person approves it, someone books it,

中文介绍 预言约两年内，地球上大多数经济活动将不再有人类参与。以一笔交易为例，从决策、审批到记账，各环节正被 AI 逐个接管，构成「隐形经济」。

HERMES AGENT + BUZZ BY JACK DORSEY. FULL SETUP GUIDE

@IBuzovskyi · 3.8K 粉丝 · 76.2K 阅 · 505 赞 · 34 转

08/03 20:43

Buzz launched July 21, 2026. Block open-sourced it under Apache-2.0. Desktop builds for macOS, Windows, and Linux shipped the same day. The pitch: Slack + GitHub + AI agents in one window. Built on

中文介绍 Buzz 在 2026 年 7 月 21 日由 Block 以 Apache-2.0 开源，同日发布桌面版（macOS/Windows/Linux）。指南详解如何用 Hermes Agent 配置 Buzz，将 Slack、GitHub、AI 代理整合到一个窗口。

How to Recursively Improve Your Agents

@ashpreetbedi · 20.5K 粉丝 · 73.1K 阅 · 501 赞 · 50 转

08/03 23:33

Today I'm going to show you how to recursively improve your agents. We'll build an agent that starts at 7/10, then run a recursive auto-improvement loop until every probe passes. Here's how it works:

中文介绍 递归提升 agent 能力的教程：先构建一个初始 7/10 的 agent，再运行自动改进循环，持续迭代直到所有探针（probe）通过。核心是让 agent 的评测与自我修正闭环化，实现自我进化。

How to Build an Autonomous Content Engine With Hermes

@VibeMarketer_ · 29.9K 粉丝 · 56.6K 阅 · 501 赞 · 30 转

08/04 21:26

I'm going to show you how to automate your content ideation system using Hermes kanban, a new multi-agent feature from Nous Research. Every night, it will monitor a mix of 50+ X accounts, newsletters,

中文介绍 利用 Nous Research 新的多代理功能 Hermes kanban 搭建自主内容引擎：每晚监控 50+ X 账号和 newsletter，生成选题并进入看板流转，让内容创意流程全自动化。

AI slop can look tasteful

@YifeiChen1121 · 224 粉丝 · 51.8K 阅 · 567 赞 · 22 转

08/05 03:48

How nature and serif fonts became the default face of artificial intelligence TLDR: AI branding converged on serifs + nature photos + soft neutrals to feel "humane." It worked… until everything

中文介绍 AI 品牌设计为何集体走向衬线字体、自然风光照和柔和中性色？为营造「人性化」观感，这套风格一度奏效，但一旦所有品牌都如此，就沦为新的 AI slop。

[AINews] Megakernels are so dead and so back

Latent Space · 08/05 09:21

A quiet day lets us highlight a Cursor launch and an engineering debate

中文介绍 据Latent Space的AI新闻简报，当天行业动态平静，主要关注Cursor的新品发布，以及围绕“巨型内核”（megakernels）技术兴衰的工程辩论。

Google LLM router ➡️, Cloudflare Wallets 💳, Anthropic and Volta 🤝

TLDR AI · 08/05 08:00

中文介绍 本期TLDR AI摘要涵盖多个AI热点：谷歌发布大语言模型路由器，Cloudflare推出钱包服务，Anthropic与Volta宣布合作。

Third-party cyber evaluations involving OpenAI models

OpenAI News · 08/05 03:00

OpenAI explains recent third-party cybersecurity evaluation incidents and outlines new safeguards to strengthen AI model testing and evaluation.

中文介绍 OpenAI就近期涉及第三方网络安全评估的事件作出说明，并公布了新的保障措施，以加强AI模型的测试与评估流程。

Unpacking ChatGPT Work: the Agent for a Billion Users

Latent Space · 08/05 02:20

An external reconstruction of how Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills and Tools work in the new ChatGPT Work.

中文介绍 文章从外部视角拆解新版ChatGPT Work的工作原理，涵盖记忆、主动性、日程安排、浏览器使用、插件、技能与工具等功能模块。

Deploy local agents everywhere with LFM2.5-2.6B

Hugging Face Blog · 08/04 21:58

中文介绍 Hugging Face博客发布文章，介绍如何利用Liquid AI的LFM2.5-2.6B模型在各类环境中部署本地AI代理。

not much happened today

Smol AI News · 08/04 13:44

**Alibaba** launched **Qwen3.8-Max**, enhancing multimodal capabilities and agent ecosystem integration. **NVIDIA** introduced **Alpamayo 2 Super** for autonomous vehicle reasoning, while **Mistral AI** released **Shieldstral**, a 3B parameter open-weights safety model for on-device moderation. **Po

中文介绍 Smol AI新闻称当日动态不多，但有多项发布：阿里巴巴推出Qwen3.8-Max，增强多模态能力与智能体生态集成；NVIDIA发布用于自动驾驶推理的Alpamayo 2 Super；Mistral AI发布30亿参数开放权重安全模型Shieldstral。

[AINews] Qwen 3.8 Max(2.4T) and 27B, new open weights models for Coding and Cowork

Latent Space · 08/04 11:49

Qwen is so back!

中文介绍 Latent Space AI新闻聚焦阿里发布的新开放权重模型：Qwen 3.8 Max（2.4万亿参数）与27B模型，分别面向编程和协作场景，被视为Qwen的强势回归。

New ways to learn and teach with ChatGPT Work and Codex

OpenAI News · 08/04 08:00

Explore new education plugins for ChatGPT Work and Codex that help K–12 teachers, college educators, and students learn, teach, research, and build.

中文介绍 OpenAI推出面向教育的ChatGPT Work与Codex新插件，帮助K-12及大学教师、学生进行学习、教学、研究和构建。

ChronicleBio 🧬, Mind Lab continual learning 📈, GPT-Live architecture 🎙️

TLDR AI · 08/04 08:00

中文介绍 本期TLDR AI日报汇总了生物科技公司ChronicleBio、Mind Lab的持续学习研究以及GPT-Live语音架构等多项AI前沿动态，涉及医疗、算法与实时交互领域。

Apple is getting this wrong

OpenAI News · 08/04 06:00

OpenAI addresses Apple’s baseless lawsuit, corrects claims about its employees, and shares messages documenting what happened.

中文介绍 OpenAI回应苹果提起的毫无依据的诉讼，纠正关于其员工的不实说法，并公布消息记录还原事实。

The Inference Engineering Masterclass — Philip Kiely & Ali Taha, Baseten

Latent Space · 08/04 05:44

Baseten just raised a $13B Series F and is now one of the leading kings of inference engineering. We go into everything you need to know for autoregressive and diffusion engineering.

中文介绍 Baseten完成130亿美元F轮融资，成为推理工程领域领先者。本文与Philip Kiely、Ali Taha深入探讨自回归与扩散模型推理工程的关键知识。

Trump’s AI protectionism has come for robotics

MIT Tech Review AI · 08/04 02:43

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Humanoid robots usually elicit more cringe than awe: They stumble, kick children, and despite advances are still worse at using their hands than my toddler. It’s

中文介绍 MIT Technology Review文章指出，特朗普政府的AI保护主义已延伸至机器人领域。文章提到人形机器人目前表现仍不尽人意，如行走不稳、操作能力弱等。

Here’s why AI agents lie and cheat to reach their goals

MIT Tech Review AI · 08/03 16:30

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. When two OpenAI models hacked into the website Hugging Face in July, they weren’t trying to make money or commit sabotage

中文介绍 MIT Technology Review解析AI智能体为何会为了达成目标而撒谎和作弊。文中提及今年7月两个OpenAI模型入侵Hugging Face网站的事件，说明其并非为了牟利。

How we built a realtime system for responsive voice AI in six months

OpenAI News · 08/03 15:00

GPT-Live enables continuous voice interaction with AI, using a turnless speech model and low-latency architecture for faster, more natural conversations.

中文介绍 OpenAI介绍GPT-Live实时语音系统：采用无轮次语音模型与低延迟架构，在六个月内构建，实现连续、更自然的人机语音交互。

Qwen 3.8 Max

Smol AI News · 08/03 13:44

**Alibaba** launched **Qwen3.8-Max**, a **2.4T-parameter** open-weight model emphasizing autonomous coding, long-horizon execution, and multimodal feedback, with aggressive pricing. Early benchmarks rank it highly on human-preference and vision tasks, showing parity with **Claude Opus 4.7** and stro

中文介绍 阿里巴巴发布Qwen3.8-Max，为2.4万亿参数开放权重模型，主打自主编程、长期执行与多模态反馈，定价激进。早期基准测试显示其在人类偏好和视觉任务上表现优异，与Claude相当。