2026-08-24

EnvHarness: Awakening Static Worlds for Agent Learning

👍 254

08/20 08:00

LLM agents learn by interacting with environments, yet these environments are hand-built and static: blind to an agent's weaknesses, and quickly left behind as it improves. While recent environment generation methods attempt to address this, they require domain-specific pipelines, rely on expensive

中文介绍 提出 EnvHarness，根据智能体当前弱点动态生成交互环境，突破手工静态环境与领域特定生成管线的局限；使 LLM 智能体在自适应环境中持续暴露短板并获得针对性训练，提升学习效率。

Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence

👍 142

08/17 08:00

Embodied agents are increasingly used to close the gap left by end-to-end policy models. Yet the agentic path has not realized closed-loop learning in physical execution: existing harnesses remain largely open-loop, following fixed skills during rollout and reflecting only after an episode completes

中文介绍 提出 Zetta ζ，一个高效闭环具身智能框架，在物理执行过程中实时感知并调整策略，改变现有 harness 回合结束后才反思的开环模式，支持智能体在真实执行中自我进化，缩小端到端策略与实际部署间的差距。

SemaPLC: A Project-Grounded, Verification-Gated Agent Harness for PLC Code Generation

👍 115

08/19 08:00

Programmable logic controllers (PLCs) run industrial plants, and large language models can already generate independent program organization units (POUs) for them. Whether such logic integrates into an existing PLC project and then runs correctly has been checked only in limited tests. We present Se

中文介绍 提出 SemaPLC，一个面向 PLC 代码生成的项目级、验证门控智能体框架。它将 LLM 生成的程序组织单元放入真实 PLC 项目集成，并经过验证后才接受，解决以往仅独立 POU 测试、缺少项目级集成与运行正确性检查的问题。

FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis

👍 114

08/19 08:00

Training terminal agents requires scalable executable supervision, yet synthesizing high-quality terminal tasks remains challenging. Each task couples an instruction, an initialized environment, a reference solution, and an executable verifier; if these artifacts are generated from inconsistent assu

中文介绍 提出 FACET，解决终端任务合成中指令、初始化环境、参考解与可执行验证器之间假设不一致的问题。通过保留源意图与可执行状态，生成高质量且可验证的终端任务，为终端智能体提供可扩展执行监督。

Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL

👍 92

08/19 08:00

Reinforcement learning (RL) has emerged as a powerful approach for improving reasoning in language and vision-language models, yet its strongest successes still depend heavily on ground-truth supervision (e.g., verifiable reward). Such annotations are costly to obtain and become increasingly scarce

中文介绍 提出 Co-RL，利用多智能体强化学习中的多样化群体实现无监督推理涌现。智能体在群体协作与竞争中相互学习，摆脱对可验证奖励等真实标注的依赖，为语言与视觉语言模型提供可扩展的推理能力提升途径。

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

👍 88

08/13 08:00

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，一个全模态、全学科的 AI 科学家系统。在假设生成、代码执行、论文撰写等工作流之外，强调对齐完整科学证据链，使智能体能够基于多模态证据进行跨学科自动发现。

SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?

👍 61

08/20 08:00

Software increasingly functions as part of the scientific instrument itself, making failures in scientific code capable of compromising not only program behavior but also the evidence underlying scientific conclusions. Yet existing evaluations of coding agents largely emphasize aggregate task succes

中文介绍 提出 SWE-bench Science 基准，评估编码智能体解决科研工程任务的能力。与只关注聚合任务成功率的现有评测不同，它强调科学代码故障对实验证据的破坏，揭示科研场景中代码智能体的真实可靠性。

SPADE: Self-Play in Adaptive Synthetic Executable Environments

👍 48

08/19 08:00

Continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals. For language agents, existing training environment pools (hand-curated, statically synthesized, or frozen-verifier) keep the goal distribution fixed as the learner scales. We introduce SPADE (Self

中文介绍 提出 SPADE，通过自适应合成可执行环境中的自我对弈，为语言智能体持续生成多样化且随能力增长的目标，打破手工或静态合成环境目标分布固定的限制，驱动智能体自我改进。

Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis

👍 32

08/19 08:00

Single-step retrosynthesis is a central component of computer-aided synthesis planning, yet its intrinsically one-to-many nature is poorly captured by single-answer evaluation and benchmarking protocols. To address this, we introduce Top-K prompting as a robust training and inference paradigm to bet

中文介绍 针对单步逆合成反应的一对多特性，提出 Top-K prompting 训练与推理范式，并结合化学合理性感知的 LLM 训练，替代单一答案评估，使模型生成更符合真实合成规划的多样候选路线。

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use

👍 31

08/20 08:00

Memory has become a key component of large language models, enabling them to retain information and learn from long-term interactions. However, existing memory benchmarks mainly evaluate whether information is correctly extracted, stored, and retrieved, while largely overlooking how retrieved memori

中文介绍 提出 MemTrapBench，用于评测 LLM 在记忆使用中的认知陷阱。现有基准只关心信息抽取、存储与检索，该基准进一步考察检索到的记忆如何误导推理与决策，为记忆机制研究提供更全面的评估。

SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback

👍 30

08/13 08:00

Agent Skills are today either hand-authored or produced in a single LLM generation pass, and consequently possess no closed loop through which they might improve from the interaction failures they actually cause. Recent work does close this loop, but derives its feedback from single-turn question-an

中文介绍 提出 SkillEvo，从多轮交互反馈中提取自更新的进化梯度，使 Agent Skill 不再依赖单次生成或手工编写，而是在实际交互失败中持续闭环改进，提升技能的稳健性与适应性。

Repo0: Design-Driven Zero-to-All Code Generation

👍 17

08/20 08:00

Large language model agents have made substantial progress in code generation, yet most existing systems assume a predefined repository architecture. This assumption does not hold in zero-to-all code generation, where an agent must construct an entire software project directly from natural-language

中文介绍 提出 Repo0，面向从自然语言直接生成完整软件项目的 zero-to-all 场景，采用设计驱动方式替代预定义仓库架构假设，使智能体在无初始代码时也能自主构建结构一致、可运行的项目级代码。

The More Popular, The Harder to Forget: Adaptive Popularity for LLM Unlearning

👍 17

08/14 08:00

Popular facts are memorised more deeply during pretraining and resist removal longer than rare ones, yet existing LLM unlearning methods apply uniform gradient pressure regardless of training-data frequency. We propose the AdaPop (Adaptive Popularity) method, which combines local token confidence wi

中文介绍 提出 AdaPop，根据训练数据流行度自适应调整遗忘压力，对高频热门事实施加更强梯度，解决现有 LLM unlearning 方法统一处理导致热门知识难以移除的问题，兼顾遗忘与保真。

FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving

👍 16

08/20 16:02

Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase. Our previous work, FlashPrefill, mitigates this cost through instantaneous pattern discovery a

中文介绍 提出 FlashPrefill V2，使用块稀疏 attention 加速长上下文 LLM 的预填充阶段，结合即时模式发现与块稀疏计算，显著降低注意力二次复杂度带来的推理瓶颈，提升长上下文服务效率。

Looped Language Models Improve Compositional Tool Calling

👍 16

08/17 08:00

Looped language models have shown promising results on reasoning benchmarks, yet their potential for agentic tool use remains largely unexplored. We study this question in compositional tool-calling settings, where models must coordinate multiple API calls, maintain intermediate state, and preserve

中文介绍 研究循环语言模型在组合式工具调用中的表现。与标准模型相比，循环结构能更好地协调多个 API 调用、维护中间状态并保持工具间依赖，在需要复杂多步工具协作的任务上带来显著提升。

FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents

👍 16

08/19 08:00

Language model agents now execute bounded tasks reliably. Whether they can sustain effective decision-making over long horizons, where actions have cumulative consequences and the environment responds to their choices, remains largely unmeasured. FM-Bench (Football Management Benchmark) measures thi

中文介绍 提出 FM-Bench 足球管理基准，评估语言模型在长时程、竞争性环境中的管理决策能力。环境中动作具有累积后果且对手动态响应，弥补现有评测仅覆盖有界任务、缺乏长期决策测度的空白。

Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning

👍 15

08/19 08:00

JEPA-style latent world models can use Euclidean distance to a goal latent as the cost for model-predictive control (MPC). Strong decoding of task variables, however, does not guarantee that this particular cost ranks candidate action sequences by real task progress. We call the latter property deci

中文介绍 针对 JEPA 潜在世界模型以潜空间距离作为 MPC 代价时与真实任务进展不一致的问题，提出决策度量对齐的诊断方法及动作条件化目标函数，确保候选动作排序反映实际任务收益。

Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

👍 14

08/14 08:00

High-quality creative writing data for large language models (LLMs) remains dominated by story-centric data, limiting models' ability to follow the structural and functional conventions of diverse creative formats. We propose an attribute-guided genre expansion framework for scaling creative writing

中文介绍 提出属性引导的体裁扩展框架，根据结构、功能等属性生成多样化创意写作数据，突破故事中心数据的限制，使 LLM 学会遵循不同非叙事体裁的规范，扩展创意写作能力。

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

👍 13

08/18 21:09

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the sc

中文介绍 在低资源语言上对三款前沿 MoE 模型进行 SFT 与 RL 微调，发现准确率基准几乎无变化且随机种子即可显著影响得分，说明精度指标在此尺度接近噪声；SFT 构建基础能力，RL 修复行为，现有准确率无法揭示真实差异。

The Embedder's Dilemma: LLMs Are Better, but at What Cost?

👍 11

08/13 08:00

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pa

中文介绍 通过受控、成本感知的对比，评估 10 个 LLM 与 26 个 embedding 模型（118M–14B）在 37 个任务上的表现，揭示用 LLM 取代嵌入管线带来的质量提升及其计算成本，为文本嵌入选型提供量化依据。

AI Engineering Skills Map: Building and Deploying AI Applications

@AndrewYNg · 1.8M 粉丝 · 62.4K 阅 · 1.2K 赞 · 189 转

08/22 00:37

I previously wrote about our AI Engineering Skills Map, with the highest level skills being (i) Building and deploying AI applications, (ii) Software engineering fundamentals, (iii) Using coding

中文介绍 吴恩达再次提及他此前发布的AI工程技能图谱，将构建和部署AI应用、软件工程基础、使用编码列为最高层技能，并暗示后续将详解这些能力的培养与进阶路径。

[AINews] 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over

Latent Space · 08/22 15:36

Did you think RSI stopped at model training?

中文介绍 文章探讨模拟技术正在接管AI领域，称其相比真实环境“性能差10%、成本低100倍、速度快10000倍”，并指出递归自我改进（RSI）的应用不仅限于模型训练阶段。

The Evolution of the Agent Harness

Latent Space · 08/22 15:30

Models keep absorbing the harness into their weights — soon, it will be a harness for human attention rather than for the model.

中文介绍 文章讨论“代理框架”（Agent Harness）的演变，指出模型正不断将外部框架内化到自身权重中，未来该框架或将服务于人类注意力而非模型本身。

Simulation: the new Scaling Law — Joon Sung Park, Simile AI

Latent Space · 08/22 07:37

Simile’s CEO about his journey from the viral Generative Agents to creating 8 Billion Digital Twins of every living human... and why it’s gone from fun exploration to very serious business.

中文介绍 Simile AI CEO Joon Sung Park 分享从“生成式智能体”到为每个人创造80亿数字孪生的历程，认为“模拟”正成为新的扩展定律，并已从趣味探索转变为严肃商业。

From Atari to EVE Online: Building on 15 Years of AI Research in Games

DeepMind Blog · 08/21 19:59

Google DeepMind partners with game studios to prototype breakthrough AI gameplay.

中文介绍 Google DeepMind 宣布与游戏工作室合作，在游戏中原型验证突破性AI玩法。文章回顾了其从Atari到EVE Online、长达15年的游戏AI研究积累。

[AINews] Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B, employees go for $6B, Infraco scaling to 7GW neocloud

Latent Space · 08/21 13:45

Yes, we’re confused too.

中文介绍 AI初创公司Poolside获得NVIDIA 120亿美元的“反向高管聘用”：创始人留任获10亿美元，员工获60亿美元，基础设施公司Infraco将扩展至7GW规模的新云（neocloud）。

not much happened today

Smol AI News · 08/21 13:44

**Ox Alpha** emerged as a mystery model with strong coding and agentic performance, likely a **Zhipu/GLM-family** model such as **GLM-5.3 Vision**. Analysts suggest its gains come from post-training and infrastructure improvements rather than sheer size, based on the **743B base** of **GLM-5.2** wit

中文介绍 神秘模型Ox Alpha在编码和智能体任务上表现强劲，据推断可能为智谱GLM系列（如GLM-5.3 Vision）。基于743B基座模型，分析认为其提升主要来自后训练与基础设施优化，而非规模扩大。

Measuring benchmark optimization in speech recognition

Hugging Face Blog · 08/21 08:00

中文介绍 Hugging Face博客探讨如何衡量语音识别（ASR）中的基准优化，分析模型在标准测试集上的性能提升及其评估方法，旨在避免过度优化基准带来的偏差。

ChatGPT Apple Messages 💬, Anthropic’s meeting recorder 💼, Mistral Agentic Search 🔍

TLDR AI · 08/21 08:00

中文介绍 本期AI资讯要点：ChatGPT集成苹果Apple Messages；Anthropic推出会议记录工具；Mistral发布智能体搜索（Agentic Search）。

The /wayfinder Skill: Navigating the “Fog of War” of Planning

Latent Space · 08/21 04:59

Matt Pocock tells us about his /wayfinder skill, for greenfield projects or for when the way forward is unclear.

中文介绍 Matt Pocock 介绍其/wayfinder技能，专为全新项目或前路不明的情况设计，帮助在规划中穿越“战争迷雾”，找到前进方向。

Up to 3.2x Faster Inference with LFM2.5-DSpark

Hugging Face Blog · 08/21 00:52

中文介绍 Hugging Face博客发布Liquid AI的LFM2.5-DSpark模型，宣称相比原有版本推理速度最高提升3.2倍，并说明了其优化方法与适用场景。

Debates over AI consciousness are a trap

MIT Tech Review AI · 08/20 23:42

“Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators. Prominent tech leaders such as Demis Hassabis, Dario Amodei, and Sam Altman push for regulation of these seemingly “superhuman”

中文介绍 MIT Technology Review文章认为，围绕AI意识的争论是一个陷阱。当前“失控”“叛逆”“自主”等修辞让人误以为AI智能体已觉醒并对创造者不满；Demis Hassabis、Dario Amodei、Sam Altman等科技领袖也在推动相关监管。

Unlocking hidden revenue streams with market models

MIT Tech Review AI · 08/20 17:47

Each day, an airline transports tens of thousands of passengers on hundreds of flights. Often these are not straightforward point-to-point routes, with passengers requiring multiple connections. The airline can consider potentially hundreds of variables to price each of these journeys: demand, seaso

中文介绍 MIT Technology Review文章探讨利用市场模型解锁隐藏收入流，以航空公司为例：每天运输数万乘客、数百航班，许多行程涉及中转，定价需考虑数百个变量，市场模型可优化此类决策。

Introducing AI Futures

OpenAI News · 08/20 15:00

Introducing AI Futures, a new OpenAI blog exploring how transformative AI could reshape power, governance, the economy, and individual freedom.

中文介绍 OpenAI宣布推出新博客“AI Futures”，旨在探讨变革性AI可能如何重塑权力、治理、经济和个人自由等宏观议题。

not much happened today

Smol AI News · 08/20 13:44

**OpenAI** and **Anthropic** expanded their agent platforms with new desktop features, collaborative editing, and composable APIs like Skills and Files API. **OpenAI** rolled out memory and workflow features in the EEA, UK, and Switzerland. **AT&T** revealed that 40% of employee AI usage routes to o

中文介绍 OpenAI与Anthropic扩展智能体平台，新增桌面功能、协作编辑，以及Skills和Files API等可组合接口；OpenAI还在欧洲经济区、英国和瑞士推出记忆与工作流功能。

[AINews] Death of Params: Z.ai CEO Jie Tang on GLM 5.3 and the new Post-training Scaling Law

Latent Space · 08/20 13:17

Every lab CEO is on X now

中文介绍 Z.ai CEO唐杰（Jie Tang）在访谈中讨论GLM 5.3与新的后训练扩展定律，认为“参数规模”增长路线正在终结，业界重心转向后训练优化。