2026-08-15

OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution

👍 252

07/31 20:00

AI agents operate in persistent environments where early state changes can influence decisions far into the future. Unlike conventional language-model interactions, agent behavior is mediated through a shared state that is repeatedly modified and reused across long-horizon workflows. Current safety

中文介绍 提出了 OpenART，旨在通过开放式环境演化扩大代理的红队测试。与传统语言模型交互不同，代理行为通过共享状态进行调节，该状态在长时间工作流中反复修改和重用。这项研究关注代理在动态环境中的决策过程，强调安全性和决策策略的演化。意义：这为 Agent 研究与安全性评估提供了新的方向。

LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers

👍 92

08/06 20:00

No single large language model (LLM) is optimal across all queries and budget constraints, making model routing essential for cost-effective deployment. Existing routers adopt diverse formulations and implementations, making fair comparison and extension difficult. We present a unified formulation o

中文介绍 提出 LLMRouter，这是一个统一的基础设施用于开发、评估和部署 LLM 路由器。目前，没有单一的 LLM 在所有查询和预算限制下都是最佳的，因此模型路由成为有效部署的关键。LLMRouter 提供了一种统一的形式，使得比较和扩展模型路由变得更加简单和有效。意义：该框架可促进 LLM 尤其是多任务学习的应用优化。

DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation

👍 82

08/12 20:00

We present DreamX-Phi 1.0, an action-conditioned video world model for robotic manipulation that, given an observed frame, a language instruction, and a prescribed action sequence comprising end-effector poses and gripper states, predicts the resulting future observations. Yet realism alone does not

中文介绍 提出 DreamX-Phi 1.0，这是一个针对机器人操作的动作条件视频世界模型。该模型能根据观察到的帧、语言指令和一系列规定的动作序列预测未来的观察结果，从而提高机器人在操作任务中的表现。此模型的引入将推动机器人操作的应用和实际场景中的适应能力提升。意义：将助力机器人的任务执行和交互能力的提升。

Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence

👍 82

08/11 20:00

AI models have achieved remarkable success across diverse domains, yet the mechanisms underlying their capabilities and the risks they may pose remain poorly understood. As AI development becomes faster and increasingly automated, mechanistic exploration remains largely manual, widening the gap betw

中文介绍 本研究探讨 AI 作为科学工具在发现智能机制中的作用，旨在理解 AI 模型成功的基础与潜在风险。文章强调在快速和自动化的 AI 发展背景下，机制探索仍以手动方式为主，导致理论与应用之间的差距不断扩大。意义：为理解 AI 的机制提供了新视角，促进更安全有效的智能系统发展。

SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries

👍 74

08/05 20:00

Large Language Models (LLMs) increasingly act as agents whose procedural knowledge is stored in reusable skill packages and loaded at inference time. As skill libraries grow, a central challenge is to expose the smallest sufficient executable context under a limited context budget. Existing systems

中文介绍 提供了一种名为 SkillZip 的方法，通过图压缩实现可扩展代理技能库的合同保护。随着大语言模型（LLMs）作为代理执行更复杂任务，如何在有限的上下文预算下有效暴露可执行环境变得至关重要。该方法旨在提升技能库的使用效率。意义：对个性化和可扩展智能代理应用有重要影响。

DarwinX: Evolving Agent Harnesses Through Natural Selection

👍 64

07/30 20:00

An LLM agent's capability depends not only on model weights but on its harness: prompts, tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage search is path-dependent and local wins often regress other tasks. We introduce DarwinX, which treats self-evolu

中文介绍 引入 DarwinX，一种演化代理工具，能够通过自我改进循环来优化代理的能力。该方法考虑到提示、工具和控制流对任务性能的影响，从而减小单一谱系搜索的路径依赖性。此设计使得在不同任务间的综合表现大幅提高。意义：有助于推进自适应和演化智能系统的研究。

Intern-S2-Preview: Scientific Agentic Foundation Model

👍 43

08/12 20:00

Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models

中文介绍 提出了 Intern-S2-Preview，一系列科学代理基础模型，旨在应对科学发现中多元化证据的推理和交互能力。模型能维持长任务周期内的进展，适应科学工具与环境。此研究将助力跨学科科学研究的自动化程度提高。意义：拓展了智能体与科学研究的结合领域。

How Can Rhetoric Reward-Hack AI Reviewers? Dissecting Rhetorical Sensitivity in AI-Based Peer Review

👍 39

08/09 20:00

As large language models increasingly participate in scientific evaluation, we investigate a potential form of reward hacking: how rhetorical choices shape AI-review judgments when reported scientific content is preserved and how these effects vary across evaluation conditions. We construct a contro

中文介绍 通过研究修辞学在 AI 基于同行评审中的影响，探讨了修辞选择如何在保持科学内容不变的情况下，影响 AI 评审的判断。论文构建了一个控制实验，系统性分析这些影响，期望为科学评价过程中引入AI的公正性与透明性提供参考。意义：为增强 AI 在科学评价中的合理性提供了新思路。

AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

👍 35

08/12 20:00

Transforming multimodal sources into condensed and structured media outputs can be fundamentally conceptualized as a long-horizon agentic process centered on a model-harness system. While an ideal harness system should align with human design priors and accumulate reusable experience through empiric

中文介绍 AutoDesign 提出了元控制优化方法，它能够将多模态源转化为紧凑且结构化的媒体输出，构建在模型-工具系统上。该方法强调设计优先级与累积可重用经验的重要性。意义：探讨长期代理设计的高效性与实用性，对多模态 AI 的发展具有推动作用。

PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives

👍 35

08/12 20:00

Video world models simulate future states conditioned on current observations and user actions. Recent systems have demonstrated impressive video consistency and action controllability over long sequences. However, fairly comparing these interactive models remains challenging. In practice, a human p

中文介绍 PlayWorld 旨在基于代理玩家通过长期目标对世界模型进行基准评估。尽管近期系统展示了良好的视频一致性与交互控制能力，然而公平比较这些交互模型仍然存在挑战。该研究提供了新的评估标准，将推动交互式世界模型的发展。意义：为未来的多模态能力发展奠定基础。

Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence

👍 30

08/12 20:00

Spatial intelligence is becoming a foundation for embodied agents, robotic planning, and multimodal assistants. To improve the spatial reasoning ability of VLM agents, existing work has mainly followed two lines. One line uses post-training methods, such as supervised fine-tuning and reinforcement l

中文介绍 提出了一种新的 Spatial Memory Agent，它基于经验的程序记忆来提升空间智能，针对 VLM 代理的空间推理能力进行了增强。该方法在空间推理的训练中结合了监督微调和强化学习的方式。意义：为机器人规划和多模态助手提供了新的技术支持。

Can LLM Agents Stick to the Script? A Benchmark for Long-Horizon Consistency in Interactive Narratives

👍 26

08/07 20:00

The rapid advancement of Large Language Models (LLMs) is revolutionizing AI for Games by enabling open-ended and fluid interactive storytelling. However, existing research has largely overlooked the critical challenge of maintaining long-horizon logical consistency and narrative integrity against un

中文介绍 本研究探讨了大语言模型（LLMs）在互动叙事中的长期一致性问题，开发了一个基准以评估其在开放式与流畅叙述中的逻辑一致性及叙事完整性。通过分析不同设定对模型表现的影响，论文强调了此项研究在游戏与叙事生成中的重要性。意义：推动 AI 在游戏叙事中的深入应用。

Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

👍 19

08/11 20:00

We present the first systematic study of Massive activations (MAs) in layer-interleaved HLA LLMs and uncover two architecture-aligned morphologies: MAs consistently spike immediately before full attention layers, forming pre-attention spikes (PAS), and can persist through intervening linear attentio

中文介绍 研究了混合线性注意力大语言模型中的巨大激活（MAs）现象，发现了与模型架构一致的两种形态：激活在全注意力层前急剧上升，并能够在随后通过线性注意力层延续。实验揭示了激活在模型无效信息转移中的作用，推动了对大语言模型性能提升的理解。意义：为深度学习模型的性能优化提供了新见解。

Self-Evolving Embodied Agents via Skill-Harness Evolution

👍 11

08/10 20:00

Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model. While supervised fine-tuning and reinforcement learning can adapt agents to

中文介绍 提出了一种自我演化的实体代理模型，该模型不仅依赖于模型权重，还关注技能、上下文和操作接口的演变。研究强调了技能演化在机器人等实际应用中的潜力，改善了代理在不同任务中的表现。意义：推动了在实体智能体中的技能和智能演化应用。

Full-bandwidth transformer

👍 10

08/08 20:00

Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth. Dense attention gives each token broad horizontal access to the past, but the vertical feedback channel between decoding steps remains narrow: only the sampled token returns

中文介绍 提出 Full-bandwidth transformer，对于自主回归 transformer 模型，其在生成令牌时横向和纵向的计算方式进行了优化，增强了每个令牌对过去信息的访问能力，改进了模型的解码效率。意义：为相关模型的效率提升提供了新的思路，有助于多个AI应用的推动。

Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

👍 10

08/09 20:00

The world evolves following its dynamics, i.e., its laws of motion. However, leading video diffusion models largely fit the pixels without modeling how the pixels transit over time. Thus, they render visually plausible frames but may not accurately obey the laws. To capture the dynamics purely from

中文介绍 研究了演绎视频世界模型的潜力，通过潜在动态推理捕捉画面演变过程。当前主流视频扩散模型主要侧重像素拟合而忽略时间动态，本文希望通过新的方法覆盖更准确的运动规律，提升模型的时间预测能力。意义：为未来视频生成和理解提供基础。

An AI4AI Framework for Visual Token Pruning

👍 7

08/06 20:00

Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. As pruning objectives, budgets, and model architectures diversify, manually navigating t

中文介绍 提出 AI4AI 框架，以优化视觉令牌剪除，从而显著降低多模态大语言模型的推理成本。与现有方法相比，新框架减少了对手工设计的依赖，增强了针对不同目标的剪除效果，有助于在多样化的模型架构中实现更多优化。意义：推动高效 AI 对视觉信息处理的研究。

The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images

👍 7

08/05 20:00

The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. They may also repeatedly crop irrelevant regi

中文介绍 研究“图像思考”模式，探讨为多模态 LLM 引入主动视觉操作的有效性。尽管该方法可提升 AI 的表现，但实验表明相较直接推理，其提升效果有限且代价高昂。论文分析了这一现象并提出未来的优化建议。意义：对多模态智能系统中的视觉处理提供了反思与指导。

AVA-Encoder: Towards Agent-Native Video Representation Learning

👍 6

08/11 20:00

Creative agents still lack an effective way to learn from high-quality human films, limiting their ability to produce cinematic-grade videos. A key challenge is the absence of a structured video representation that is both faithful to film content and directly usable for agentic reasoning and manipu

中文介绍 提出 AVA-Encoder 模型，旨在增强代理从高质量人类影片中学习的能力，以提升其创作电影级视频的能力。关键在于建立一种同时忠实于影片内容并可用于代理推理的结构化视频表示。意义：为 AI 在影视创作中的应用开辟了新前景。

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

👍 6

08/12 20:00

Recent advances in foundation models have enabled AI scientists to automate increasingly complete research workflows, from hypothesis generation and code execution to manuscript preparation. Yet workflow coverage alone does not provide access to the full evidence on which scientific discovery depend

中文介绍 提出 OmniScientist，一个跨模态、跨学科的 AI 科学家，能够自动化完整的研究工作流。虽然工作流覆盖面逐渐扩大，但当前的挑战在于如何保证科学发现的全面性与准确性。此研究为科学研究的自动化提供了新维度。意义：推动 AI 在科学发现领域的深入应用。

Cursor + SpaceXAI: the fastest iterating team wins

@sarahdingwang · 18.0K 粉丝 · 1.2M 阅 · 514 赞 · 73 转

08/14 10:28

By @BornsteinMatt, @martin_casado, and myself. AI coding could be the biggest addressable market of our lives. And, unless you think we’ve fully addressed it (which no one thinks; we are more like 1%

中文介绍 博主强调，AI 编程市场仍有巨大潜力，当前开发者的挖掘程度仍不到 1%。同时，分享了对 Cursor 与 SpaceXAI 的合作观点，指出快速迭代团队在这一领域的优势。

Agent Plugins are the future of Agent Skills

@GoogleCloudTech · 1.3M 粉丝 · 56.7K 阅 · 505 赞 · 76 转

08/12 22:49

Agent Plugins is an open, vendor-neutral standard for packaging Agent Skills and the MCP servers they depend on into one portable folder that any compatible client can load. Google is joining the

中文介绍 博主介绍了 Agent Plugins 作为包装 Agent Skills 的新标准，可以提升工作流的兼容性和便捷性。这一趋势表明，行业正在朝着开放和中立标准发展，推动多方协作。

The Elon–Jensen "Understanding"

@TeslaLarry · 53.1K 粉丝 · 56.5K 阅 · 507 赞 · 54 转

08/13 12:20

After the SpaceX earnings call on August 4, several pieces started lining up for me in a way that is hard to ignore. Elon didn’t just praise Nvidia hardware, he declared an exclusive commitment to

中文介绍 博主从 SpaceX 财报解析出市场趋势，特别提到 Elon 对 Nvidia 硬件的重视，暗示这一先进硬件将在未来 AI 开发及转换中扮演关键角色。

/show-me: compact visual representations for coding agents

@dexhorthy · 30.0K 粉丝 · 42.9K 阅 · 635 赞 · 31 转

08/12 11:59

tl;dr make your agent converse visually instead of in walls of prose. Lighter and faster than HTML, good enough for most dev-work shaped problems. Coding agents are pretty much unreadable The

中文介绍 博主提出让编程代理使用视觉对话的构想，替代传统的文字描述，将代码的可读性和交互性提升到一个新水平，表明行业对开发工具优化的方法有新的探索。

Preparing GLM-5.3 for Open Release: A Responsible Path to Cyber Defense

@Zai_org · 136.2K 粉丝 · 42.7K 阅 · 583 赞 · 50 转

08/14 11:04

When GLM-5.2 helped Hugging Face investigate an incident in which an AI autonomously bypassed its own safeguards, it highlighted a broader shift. AI is becoming part of both cyber offense and cyber

中文介绍 博主探讨 GLM-5.3 的开放发布，强调其在网络安全领域的重要作用。这突显 AI 在网络攻防中逐渐起到关键角色，预示着新的安全防护思路正在形成。

The AI Engineering Skills Map

@AndrewYNg · 1.8M 粉丝 · 41.2K 阅 · 716 赞 · 89 转

08/14 12:29

I am delighted to present The AI Engineering Skills Map. AI allows us to build software very differently today than in 2022, and everyone with the skills to take advantage of this shift has numerous

中文介绍 博主发布 AI 工程技能地图，指出 AI 技术使软件开发方式发生了根本变化，具备相关技能的人员将拥有更广阔的职业发展前景，这一变化在行业中引起广泛关注。

Grok 4.6 – A field guide

@ericzakariasson · 80.4K 粉丝 · 32.2K 阅 · 719 赞 · 52 转

08/12 11:46

Grok 4.6 is out! I've used it for a few weeks as my daily driver across the normal mix of coding and knowledge work, and built a few projects with it specifically to push on where it holds up. It's

中文介绍 博主分享 Grok 4.6 的使用体验，指出其在日常编程与知识工作中的表现，并测试了多个项目，反映出开发者对此工具堪用性的重视。这标志着开源工具的持续进化。

What's new in Gemini 3.7 Flash

@GoogleAIStudio · 192.9K 粉丝 · 31.5K 阅 · 535 赞 · 34 转

08/14 11:00

Gemini 3.7 Flash (gemini-3.7-flash) is generally available (GA) and ready for production use. It is our most intelligent workhorse model yet for software engineering, and agentic workflows. This guide

中文介绍 博主揭示 Gemini 3.7 Flash 的发布，这一高效的智能模型专为软件工程与代理工作流设计，表明 AI 工具在提升生产力方面的重大进展。

Introducing Gemini 3.7 Flash

@GoogleAIStudio · 192.9K 粉丝 · 30.7K 阅 · 566 赞 · 60 转

08/13 13:25

Today, we’re building on the progress of our widely used Flash series by introducing Gemini 3.7 Flash, our most intelligent workhorse model yet for coding and agents. This release comes just three

中文介绍 博主介绍 Gemini 3.7 Flash 的新特性，是其 Flash 系列中的最新迭代，旨在进一步提升编码与智能代理的效率，预示着 AI 工具的快速发展与应用深化。

Qwen3.8-27B Practical Guide: Control Reasoning Depth and Extend Context to 1M Tokens

@QwenDevs · 8.0K 粉丝 · 25.6K 阅 · 559 赞 · 51 转

08/14 11:40

Qwen3.8-27B is here — a compact, deployment-friendly dense model built on the Qwen3.5 architecture, with native vision-language understanding and flexible thinking control. Two developer-facing

中文介绍 博主推出 Qwen3.8-27B，强调其在视觉-语言理解及灵活推理控制方面的新特性，此模型的发布展示了行业在 AI 模型优化和应用场景扩展中的新方向。

[AINews] Gemini 3.7 Flash brings GDM back to the forefront

Latent Space · 08/14 01:30

Down, but not out!

中文介绍 Gemini 3.7 Flash 的发布使 GDM 再次成为焦点，尽管面临下降风险，但仍未被淘汰。

State of Open Models: Summer 2026 Observations

Hugging Face Blog · 08/13 20:00

中文介绍 关于2026年夏季开放模型的观察结果尚未公开详细信息。

Gemini 3.7 🤖, GPT-5.6 Sol Ultrafast ⚡, Anthropic $2T IPO 💰

TLDR AI · 08/13 20:00

中文介绍 Gemini 3.7 和 GPT-5.6 将成为 AI 技术中的重要发展，期待带来显著改进。

Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

Hugging Face Blog · 08/13 13:16

中文介绍 Strands Agents 和 LeRobot 将允许从一个地方记录、训练和部署，具体细节尚未披露。

Introducing Gemini 3.7 Flash

DeepMind Blog · 08/13 13:04

中文介绍 DeepMind 宣布推出 Gemini 3.7 Flash，相关功能与特性暂无详细介绍。

Flock is tightening its rules in response to a growing surveillance backlash

MIT Tech Review AI · 08/13 09:41

The police-tech giant Flock is announcing today that it will change officers’ access to its nationwide network of license plate readers, in an apparent effort to quell a growing backlash and win back contracts lost amid concerns about mass surveillance and police abuse. Several changes aim directly

中文介绍 Flock 宣布将更改警察对全国车牌读写器网络的访问，以回应日益增长的监控反对压力。

The builder’s guide to GPT‑5.6

OpenAI News · 08/13 07:00

Learn how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.

中文介绍 OpenAI 为创业公司提供指导，介绍如何利用 GPT-5.6 更快捷且高效地构建 AI 代理。

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed

OpenAI News · 08/13 06:00

Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14× faster. Powered by Cerebras, it delivers up to 750 output tokens per second.

中文介绍 OpenAI 发布 GPT-5.6 Sol 的超快速模式，速度比之前快最多 14 倍，提供每秒最多 750 个输出令牌。

OpenAI appoints Dali Rajic as Chief Revenue Officer

OpenAI News · 08/13 05:00

OpenAI appoints Dali Rajic as Chief Revenue Officer to lead its global revenue organization and help businesses realize the full value of AI.

中文介绍 OpenAI 任命 Dali Rajic 为首席营收官，负责全球营收组织，助力企业充分利用 AI。

How kids feel about AI, in their own words

MIT Tech Review AI · 08/13 05:00

When we set out to talk to kids about artificial intelligence, we thought we knew what we’d hear. We expected some to tell us they were using it to cheat a little, the way Millennials and Gen Xers opened up CliffsNotes or programmed formulas into their TI-82s, and others to share inspiring ways they

中文介绍 关于儿童对人工智能的看法，相关采访显示他们的态度与以往几代人相似。

not much happened today

Smol AI News · 08/13 01:44

**Google** rapidly released **Gemini 3.7 Flash** just three weeks after 3.6 Flash, targeting coding, web development, knowledge work, and agentic workflows with a 50% introductory price cut and improved benchmark scores like **DeepSWE 65.3%** and **Code Arena Elo 1588**. The update quickly integrate

中文介绍 谷歌在短短三周内发布了 Gemini 3.7 Flash，推出价格减少 50% 的优惠，并在多个基准测试中取得了 DeepSWE 65.3% 和 Code Arena Elo 1588 的成绩。

[AINews] SpaceXAI Grok 4.6 and Grok @Bot

Latent Space · 08/12 21:53

AI teammate category just had its most significant new entrant yet

中文介绍 SpaceXAI Grok 4.6 和 Grok @Bot 成为 AI 伙伴类别中最新的重要参与者。

What We Learned by Reproducing 2,200 papers from ICML

Hugging Face Blog · 08/12 20:00

中文介绍 关于重现 2200 篇 ICML 论文的学习结果尚无进一步信息。

Claude Chrome Cowork 🌐, Grok 4.6 🚀, DeepSeek v4-Pro-0813 🧠

TLDR AI · 08/12 20:00

中文介绍 Claude Chrome Cowork 和 Grok 4.6 的发布标志着 AI 技术的快速发展，具体内容待定。

Scaling AI agents with trustworthy data

MIT Tech Review AI · 08/12 12:51

Business and technology leaders need no convincing that the time of agentic AI is here. Organizations are rapidly adopting agents, and few executives doubt the technology’s potential to transform work. But many organizations find that realizing the desired return on investment (ROI) from AI hinges o

中文介绍 随着代理 AI 的迅速普及，企业需关注信任数据在实现投资回报率中的关键作用。