2026-07-20

LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

👍 175

07/16 08:00

A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment. The gap is especially important for AI agents, whose

中文介绍 提出LongStraw，在固定GPU预算下实现超过2M token的长上下文RL后训练。通过优化训练策略与内存管理，突破长度泛化限制，使AI agent等场景获得更长上下文推理能力。

VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

👍 145

07/16 08:00

Recent advances in video understanding have spanned motion, long video, and streaming interaction, driving this field toward real-world applications. Despite this progress, current open-source models remain limited in several ways. They often struggle to generalize across diverse video types, making

中文介绍 VideoChat3为全开源视频多模态大模型（Video MLLM），支持高效通用视频理解。克服现有开源模型对多样视频类型泛化能力弱的问题，覆盖运动、长视频与流式交互场景。

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

👍 126

07/14 08:00

We introduce Boogu-Image-0.1, an open-source unified multimodal understanding and generation model family, comprising Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in high-quality text-to-image generation, fast inference, instruction-based editing, and bilingual (Ch

中文介绍 Boogu-Image-0.1模型族（Base/Turbo/Edit/Edit-Turbo）开源统一多模态理解与生成。在文本到图像生成、指令编辑和双语支持上表现强劲，兼顾推理速度与生成质量。

Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

👍 91

07/14 08:00

Reinforcement learning with verifiable rewards without human-annotated data, often referred to as zero RL, has emerged as a powerful paradigm for eliciting chain-of-thought reasoning. However, due to computational constraints, existing studies are largely restricted to small models, leaving the trai

中文介绍 Ring-Zero将零强化学习（zero RL）扩展至万亿参数规模，基于可验证奖励激发涌现式思维链推理。突破计算约束，验证大规模模型上RL训练能催生复杂推理能力。

SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

👍 85

07/16 08:00

Large language models are increasingly trained as interactive agents for long-horizon tasks involving multi-turn interaction, tool use, and environment feedback. Outcome-based reinforcement learning (RL) provides a practical optimization paradigm, but its sparse trajectory-level rewards offer limite

中文介绍 SEED提出自进化在线策略蒸馏用于Agent强化学习。针对长程交互任务中稀疏轨迹奖励问题，通过在线策略蒸馏自演进来提升多轮工具使用与环境反馈学习效率。

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

👍 60

07/16 08:00

Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-ag

中文介绍 SearchOS-V1增强开放域信息搜索代理协作鲁棒性。应对交互历史增长导致的任务跟踪困难与搜索失败场景，新框架提升证据整合与进度追踪能力。

BadWAM: When World-Action Models Dream Right but Act Wrong

👍 46

07/16 08:00

World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety,

中文介绍 BadWAM揭示世界-动作模型（WAM）中的隐患：未来世界预测正确（「dream right」）但动作生成错误（「act wrong」）。表明耦合可能带来虚假鲁棒性，需警惕推理与执行的不一致。

AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities

👍 38

07/15 08:00

As Large Language Models (LLMs) evolve into autonomous agents, the need for unified evaluation infrastructure becomes critical. However, current evaluation pipelines remain highly fragmented and tightly coupled, hindering reproducibility and causing redundant engineering. To address this, we introdu

中文介绍 AgentCompass提供统一评估基础设施，解决LLM Agent评估碎片化、紧耦合、难以复现的问题。支持标准化能力评测，减少冗余工程，提升研究可复现性。

KeyFrame-Compass: Towards Comprehensive Evaluation of Keyframe-Conditioned Video Generation

👍 36

07/15 08:00

Video generation increasingly relies on keyframe-based workflows, where creators specify a sequence of reference images to guide generation. Although recent models support multi-keyframe conditioning, it remains unclear whether they can faithfully reproduce the prescribed keyframes while maintaining

中文介绍 KeyFrame-Compass全面评估关键帧条件视频生成，衡量模型对指定帧的忠实复现与视频连贯性。填补现有基准缺乏多关键帧保真度评价的空白。

PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails

👍 35

07/07 08:00

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes.

中文介绍 PolicyShiftGuard针对安全策略动态变化场景，提出策略自适应图像护栏基准与改进方法。使同一图像在不同产品/政策下可动态调整安全性判定，而非固定属性。

MetaView: Monocular Novel View Synthesis with Scale-Aware Implicit Geometry Priors

👍 34

07/13 08:00

Current visual generation models are capable of producing high-quality content, yet they lack a coherent perception of the spatial structure. Existing generative novel view synthesis methods typically introduce explicit geometry priors, which enforce spatial consistency but inherently restrict gener

中文介绍 MetaView使用尺度感知隐式几何先验进行单目新视角合成。相比显式几何先验方法，在保持空间连贯性的同时不限制生成多样性，实现高质量、灵活渲染。

MultiRef-Compass: Towards Comprehensive Evaluation of Multi-Reference-to-Audio-Video Generation

👍 31

07/15 08:00

Multi-reference-to-audio-video (MR2AV) generation aims to generate coherent audio-video content conditioned on multiple references and textual instructions. Existing benchmarks mainly focus on text-driven generation, single-reference subject preservation, or isolated audio-video alignment, leaving t

中文介绍 MultiRef-Compass全面评估多参考条件音视频生成（MR2AV）。现有基准局限于文本驱动或单参考，该评测系统聚焦多参考保真度与音视频联合对齐。

Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes

👍 29

07/14 08:00

Human cognition does not separate understanding and generation. A teacher at a whiteboard speaks and draws together, each modality reshapes the other. In this paper, we bring this coupled loop to artificial systems. Masked Diffusion Models (MDMs) are ideally suited to this task, yet existing sampler

中文介绍 引入自校正耦合马尔可夫跳过程实现图像理解与生成并发执行。基于Masked Diffusion Models（MDMs），两个过程相互重塑，突破传统采样器无法同时进行理解与生成的限制。

UniVR: Thinking in Visual Space for Unified Visual Reasoning

👍 28

07/14 08:00

Learning broad world knowledge directly from raw visual data is a fundamental capability of intelligence. We introduce UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations. At its core, U

中文介绍 UniVR首次从纯视觉演示中同时学习复杂推理、细粒度物理动力学与长期规划。核心是视觉空间统一推理，无需语言标注，直接从原始视觉数据习得世界知识。

GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

👍 26

07/15 08:00

World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at infer

中文介绍 GigaWorld-Policy-0.5基于AutoResearch加速并增强世界-动作模型（WAM）。改进推理时不显式生成未来视频的设计，实现更快更强、物理接地的机器人策略学习。

Self-Improvements in Modern Agentic Systems: A Survey

👍 25

07/14 08:00

Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern self-improving agents as adaptive systems that convert experience into a

中文介绍 综述现代Agent系统的自我改进机制，将自治代理视为自适应系统，从经验中可控进化。强调最小人工干预下的能力提升，梳理当前方法、挑战与未来方向。

RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination

👍 23

07/15 08:00

Embodied cognition requires agents to connect high-level task reasoning with the physical states to be achieved. We introduce Hy-Embodied-RxBrain, an embodied cognition foundation model with joint language-visual reasoning and imagination. Unlike vision-language models that emphasize scene understan

中文介绍 Hy-Embodied-RxBrain具身认知基础模型，联合语言-视觉推理与想象。超越传统VLM场景理解，将高层次任务推理与物理状态连接，实现认知-行动闭环。

Spectral Rewiring for Exploration, Purification, and Model Merging

👍 22

07/03 08:00

Reinforcement learning has become a standard post-training recipe for large language models, but dense full-parameter updates create two deployment-relevant bottlenecks: suppressed reasoning performance, often reflected by premature saturation of test-time scaling, and interference when consolidatin

中文介绍 提出频谱重连（Spectral Rewiring）方法用于LLM RL后训练的探索、净化与模型合并。解决全参数更新导致的推理性能抑制与模型合并干扰问题，提升测试时扩展效率。

Demystifying On-Policy Distillation: Roles, Pathologies, and Regulations

👍 18

07/15 08:00

On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the role of OPD as an exploration catalyst: it steers the student t

中文介绍 系统研究在线策略蒸馏（OPD）在LLM后训练中的角色、病理与调控。明确OPD作为探索催化剂引导学生学习，并揭示训练动态中的问题与调控策略。

Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation

👍 18

07/15 08:00

While recent advances in 3D generation have enabled impressive visual synthesis, existing methods often rely on 2D diffusion supervision without explicit mechanisms for geometric consistency, leading to spatial hallucinations such as duplicated structures and misaligned geometry. These issues become

中文介绍 Hallo4D缓解多模态幻觉以实现一致时空生成。针对3D生成因缺乏几何一致性导致的重复结构、几何错位等问题，提出显式机制保证时空连贯性。

Maybe Intelligence Ain't All That

@CliffordSosin · 16.4K 粉丝 · 197.0K 阅 · 514 赞 · 49 转

07/19 05:36

Superintelligence arrived. You probably didn't notice, because it turned out to be kind of incremental. Don't believe me? Run the test. Talk to Fable 5 for an hour, then talk to your ten smartest

中文介绍 超级智能已悄然到来，但效果增量式。作者建议测试Fable 5模型（或游戏）一小时后，再与你最聪明的十个人对话，结果可能出乎意料。对AI能力实际进展的反思。

How to build your first team of agents:

@EXM7777 · 126.7K 粉丝 · 173.6K 阅 · 521 赞 · 51 转

07/18 21:48

I'm going to show you how to build your first team of AI agents... and how to wire them into a self improving loop inside a shared workspace the team we're building has 5 agents, one for each function

中文介绍 手把手教程：构建首个AI智能体团队。包含5个职能不同的智能体，在共享工作区内形成自我改进的循环。适合入门多智能体协作与工作流设计。

Kimi K3 on KernelBench

@elliotarledge · 40.7K 粉丝 · 69.3K 阅 · 506 赞 · 45 转

07/17 17:24

This is the Kimi K3 post you guys have been waiting for. I got some early access to this model and have been testing it on kernels, and even before seeing the benchmark scores I was impressed with its

中文介绍 抢先测试Kimi K3模型在KernelBench上的表现。作者获早期访问权，在未见基准分前就对内核任务处理能力印象深刻。期待后续正式分数。

From Loop Engineering to Graph Engineering?

@IntuitMachine · 62.3K 粉丝 · 50.7K 阅 · 552 赞 · 47 转

07/18 18:00

What the shift in AI agent architecture is really about Peter Steinberger just posted nine words that gathered thousands of likes: https://x.com/steipete/status/2078277297791189132 "Are we still talking loops or did we

中文介绍 AI智能体架构正从「循环工程」转向「图工程」。引用Peter Steinberger的经典九字帖引发热议：我们还在讨论循环，还是已进化到图？探讨架构趋势。

[AINews] not much happened today

Latent Space · 07/18 12:30

a quiet day

中文介绍 AI新闻简报指出当天市场消息平淡，无重大事件发生。

Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers

Hugging Face Blog · 07/17 23:57

中文介绍 NVIDIA与Hugging Face合作，通过NeMo Automodel和Diffusers库实现视频与图像模型的大规模微调。

A scorecard for the AI age

OpenAI News · 07/17 18:00

Sarah Friar, CFO of OpenAI, introduces a practical AI scorecard to measure ROI through useful work, cost per successful task, dependability, and return on compute.

中文介绍 OpenAI首席财务官Sarah Friar推出AI记分卡，通过有用工作、每成功任务成本、可靠性和计算回报率衡量AI投资回报。

The risk of weather data sabotage is rising

MIT Tech Review AI · 07/17 16:57

Every morning, airline dispatchers, grid operators, and farmers around the world make decisions based on the same thing: a weather forecast. While these forecasts are something that most people glance at for two seconds, weather predictions influence major strategic decisions in many industries, wit

中文介绍 天气数据破坏风险日益增加，全球航空调度、电网运营和农业等依赖天气预报的行业面临威胁。

not much happened today

Smol AI News · 07/17 13:44

**Moonshot's Kimi K3 release** has sparked a reassessment of **Chinese open-weight models**' proximity to the frontier, with strong performance in coding, agentic tasks, and long-horizon knowledge work. The strategic focus has shifted from a "compute moat" to an "efficiency stack" involving **MoE ro

中文介绍 Moonshot发布Kimi K3模型，引发对中国开源模型前沿水平重新评估，其在编码、智能体任务和长周期知识工作表现出色，战略重心从算力护城河转向效率堆栈。

[AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing

Latent Space · 07/17 09:46

a great week for open models continues.

中文介绍 Moonshot发布Kimi K3模型，参数规模2.8T，采用A50B架构，为史上最大开源模型，性能达到Opus 4.8级别，定价与Sonnet 5相当。

Kimi K3 🌕, Gemini 3.5 delayed ⏳, crushing ARC-AGI 3 🤖

TLDR AI · 07/17 08:00

中文介绍 今日AI快讯：Moonshot发布Kimi K3模型；谷歌Gemini 3.5延迟发布；ARC-AGI 3基准被攻克。

Why teens deserve access to safe AI

OpenAI News · 07/17 00:00

Learn how OpenAI is making ChatGPT safer for teens with age-appropriate protections, learning tools, parental controls, and expert partnerships.

中文介绍 OpenAI致力于为青少年提供安全AI使用体验，通过年龄适配保护、学习工具、家长控制和专家合作提升ChatGPT安全性。

🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez-Bombarelli, Lila Sciences

Latent Space · 07/16 21:30

Lila is betting that science, not the internet, is the last untapped source of training data. We went to find out what that actually looks like in a room full of robots.

中文介绍 Lila Sciences认为科学是最后未被开发的数据源，未来实验室应如数据中心，机器人自动化驱动科学研究。

Newer Models, Same Advantage

Hugging Face Blog · 07/16 19:49

中文介绍 文章探讨新模型在性能提升的同时保持原有优势，但具体内容未提供。

Our approach to bioresilience

DeepMind Blog · 07/16 17:30

Google DeepMind and Isomorphic Labs are sharing our joint approach to bioresilience and AI models.

中文介绍 Google DeepMind与Isomorphic Labs公布其在生物弹性与AI模型方面的联合方法。

[AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B)

Latent Space · 07/16 14:18

Thinky's first full LLM release is a banger and bonus: it's open weights!

中文介绍 Thinky发布首个完整LLM模型Inkling，975B-A41B参数多模态，为美国Apache 2.0开源最佳模型，另有276B-A12B小型版本，均开放权重。

not much happened today

Smol AI News · 07/16 13:44

**Moonshot AI** launched **Kimi K3**, a frontier-class open-weights model with **2.8T parameters**, **1M-token context window**, and **native multimodal input**. It features novel **Kimi Delta Attention (KDA)** enabling up to **6.3x faster decoding** and **Attention Residuals** for **~25% higher tra

中文介绍 Moonshot AI发布Kimi K3开源模型，参数2.8T，支持100万token上下文窗口和原生多模态输入，采用Kimi Delta Attention技术实现6.3倍解码加速及Attention Residuals机制。

How Cars24 scales conversations and builds faster with OpenAI

OpenAI News · 07/16 08:00

Cars24 uses OpenAI-powered voice and chat agents to handle 1M+ monthly conversation minutes, recover 12% of lost leads, and bring agentic workflows to teams across the company.

中文介绍 Cars24借助OpenAI语音和聊天代理处理每月超100万分钟对话，挽回12%流失线索，并在全公司推广智能体工作流。

Security incident disclosure — July 2026

Hugging Face Blog · 07/16 08:00

中文介绍 Hugging Face披露2026年7月安全事件，但未提供具体细节。