2026-07-10

Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning

👍 71

07/08 08:00

Structure-property relationships are foundational to biology, chemistry and materials science, where function, reactivity and physical response emerge from spatial, chemical and periodic organization. Mechanistically explaining these relationships requires interpreting structural evidence through sc

中文介绍 提出Deep Native Structural Reasoning方法，通过深度神经网络对空间、化学和周期组织进行结构推理，实现跨学科、透明且准确的结构-性质关系理解，适用于生物学、化学和材料科学。

Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling

👍 55

07/03 08:00

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate c

中文介绍 提出层次稀疏注意力机制，通过改进的chunk-wise稀疏注意力解决LLM长上下文中的二次计算开销和长度外推问题，在保持全注意力准确性的同时显著降低计算成本，走向无限上下文建模。

Dual Latent Memory in Vision-Language-Action Models for Robotic Manipulation

👍 44

07/08 08:00

Mainstream Vision-Language-Action (VLA) models predict actions primarily from the current observation under a Markovian assumption, thus struggling with long-horizon, temporally dependent tasks. Existing memory-augmented VLAs either expand the observation window or retrieve history from the memory b

中文介绍 提出Dual Latent Memory机制增强VLA模型，通过双潜记忆存储短期和长期历史信息，无需扩大观测窗口即可有效利用时序上下文，提升机器人长程操作任务的性能。

Gemma 4 Technical Report

👍 39

07/02 08:00

We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family. Designed to advance compute efficiency and reasoning, the Gemma 4 model suite features dense and Mixture-of-Experts architectures, ranging from 2.3B to 31B parameters. Alongside impr

中文介绍 Gemma 4是新一代开源多模态语言模型，采用dense和MoE架构，参数规模2.3B至31B。在计算效率和推理能力上取得进步，保持开放权重，适用于多种多模态任务。

Vision as Unified Multimodal Generation

👍 38

07/07 08:00

We formulate computer vision as unified multimodal generation, where heterogeneous visual tasks are expressed in the native text and image generation spaces of a unified multimodal model, without task-specific architectures. Under this formulation, SenseNova-Vision uses natural-language instructions

中文介绍 提出将计算机视觉视为统一多模态生成任务，SenseNova-Vision通过自然语言指令将各种视觉任务转化为文本和图像生成，无需任务特定架构，实现视觉任务统一处理。

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

👍 25

07/06 08:00

Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token depend

中文介绍 DSpark提出置信度调度的推测解码，结合半自回归生成，通过动态调度置信度驱动的草稿生成并捕获token间依赖，缓解并行草稿生成中接受率快速衰减问题，加速LLM推理。

Infinite Worlds with Versatile Interactions

👍 21

07/08 08:00

We present LingBot-World 2.0 (also known as LingBot-World-Infinity), an advanced iteration of LingBot-World featuring four distinct upgrades. (1) Our model achieves an unbounded interaction horizon while maintaining consistent output quality, benefiting from a carefully crafted causal pretraining pa

中文介绍 LingBot-World 2.0实现无限交互世界，具备无界交互时域、因果预训练、一致输出质量等四项升级，支持丰富交互场景，提升长期任务表现稳定性。

LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

👍 20

07/05 08:00

Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corp

中文介绍 提出LLM-as-a-Tutor方法，针对非可验证指令遵循RL训练，通过策略感知的提示自适应动态调整训练提示，使LLM裁判的评估标准与当前策略同步进化，提升奖励信号有效性。

Parallelized Autoregressive Decoding for Omni-Modal Dense Video Captioning

👍 18

07/03 08:00

Dense video captioning aims to generate temporally grounded descriptions of video events, benefiting both event-level video understanding and generation. In this domain, autoregressive video large language models have emerged as a prevalent paradigm due to their strong generative and cross-modal mod

中文介绍 提出并行自回归解码方法用于全模态密集视频描述，通过并行化生成加速视频事件描述，同时保持跨模态建模能力，提升密集视频描述任务效率。

JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications

👍 13

06/29 08:00

JD.com, one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher opera

中文介绍 京东Oxygen AIIC V1是以LLM/VLM为核心的工业级商品理解与管理解决方案，处理数十亿SKU，通过多模态大模型提取结构化商品知识，提升消费者体验、降低管理成本并提高运营效率。

TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training

👍 13

07/07 08:00

On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify two key inefficiencie

中文介绍 TurnOPD改进在线策略蒸馏，使其对turn感知，通过解决两个关键低效问题，使蒸馏过程更适配多轮交互场景，加速长程智能体训练中的策略收敛。

From Foundation to Application: Improving VLA Models in Practice

👍 12

07/07 08:00

Despite recent progress of VLA foundation models, the disparity between laboratory conditions and real-world applications continues to impede their practical implementation. To bridge this gap, we present LingBot-VLA 2.0, which advances LingBot-VLA through improvements in three functional domains. (

中文介绍 LingBot-VLA 2.0在三个功能领域改进VLA模型，解决实验室与现实的差距，增强机器人操作等任务中的鲁棒性和泛化性，推动基础模型向实际应用转化。

MentalThink: Shaping Thoughts in Mental SVG World

👍 11

07/03 08:00

We introduce MentalThink, a visual-symbolic reasoning paradigm that equips Multimodal LLMs (MLLMs) with an executable mechanism for "mental" visualization. The core of MentalThink is a think-with-SVG pipeline, where the model learns to generate, render, and interpret scalable vector graphics (SVG) c

中文介绍 MentalThink提出视觉-符号推理范式，通过“think-with-SVG”流水线让多模态LLM在思维中生成、渲染和解释SVG图像，实现内部可视化推理，增强多模态推理能力。

CanvasAgent: Enabling Complex Image Creation and Editing via Visual Tool Orchestration

👍 9

07/06 08:00

Complex image creation and editing often require more than a single generation or editing model. A user request may involve synthesizing images, localizing objects, segmenting regions, editing selected content, compositing intermediate assets, reading text, and enhancing the final result. Such tasks

中文介绍 CanvasAgent通过视觉工具编排实现复杂图像创建与编辑，将图像生成、定位、分割、编辑、合成等多步操作组织成可执行流程，自动完成用户多步骤图像任务。

PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation

👍 9

07/02 08:00

State-of-the-art single-image 3D reconstruction methods often rely on complex hybrid architectures and loss functions, or compress geometry into latent spaces in order to leverage pre-trained latent diffusion models. In this work, we show that such architectural overhead and intricate loss formulati

中文介绍 PointDiT在像素空间进行扩散用于单目几何估计，简化传统混合架构和复杂损失函数，直接在像素空间生成点云，实现高效的单图像3D重建。

CGGS: Consistency-Augmented Geometric Gaussian Splatting for Ego-centric 3D Scene Generation

👍 7

07/04 08:00

Challenges remain in ego-centric 3D scene generation due to limited view overlap and the dominant influence of individual perspectives on scene interpretation. These factors hinder the creation of viewpoint-consistent and semantically aligned visual content, as well as the construction of accurate g

中文介绍 CGGS提出一致性增强几何高斯泼溅，用于以自我为中心的3D场景生成，通过引入一致性约束解决视角重叠有限和个体视角偏差问题，生成视角一致且语义对齐的3D场景。

Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning

👍 6

07/08 08:00

Reinforcement learning (RL) is becoming increasingly important for post-training large language models (LLMs). Previous RL pipelines for LLMs were mostly synchronous and batch-interleaved, which is inefficient for long-horizon agentic tasks. Recently, asynchronous RL has emerged as a more efficient

中文介绍 提出单次rollout异步优化方法用于智能体RL训练，相比传统同步批处理，允许异步更新，提高长程代理任务中LLM后训练的效率。

RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

👍 6

07/07 08:00

Generalist robot manipulation policies have advanced rapidly, yet existing benchmarks remain limited in systematically evaluating their capabilities. Many rely on simple, short-horizon, or skill-narrow tasks with limited capability coverage, and are often conducted only in simulation or only in the

中文介绍 RoboDojo是统一的仿真与真实世界基准，全面评估通用机器人操作策略，涵盖多种长时域、多技能任务，在仿真和真实环境中系统测试策略的泛化能力。

Attending to Multimodal Generation One Token at a Time

👍 6

07/04 08:00

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during

中文介绍 逐token分析多模态语言模型的生成过程，揭示视觉与语言信息在自回归解码中的动态交互，通过细粒度注意力分析理解多模态计算中的token级行为。

Quantifying and Expanding the Theoretical Capacity of Late-Interaction Retrieval Models

👍 6

07/07 08:00

Late-interaction retrieval models that use the MaxSim similarity function have shown strong empirical performance, often outperforming single-vector dense and sparse retrieval models. Despite these empirical findings, little is known about the theoretical representation power of MaxSim and how it co

中文介绍 量化并扩展后交互检索模型（MaxSim）的理论容量，分析其表示能力及与单向量模型的关系，提供理论保证并探讨扩展途径。

The Task Economy - Data will be the next $1 Trillion Category

@EverettRandle · 19.4K 粉丝 · 484.8K 阅 · 507 赞 · 56 转

07/08 00:15

The Token Economy When we talk about AI today, tokens are king. Specifically, inference tokens have emerged as the primary proxy for tracking the growth of the AI ecosystem. Public companies report

中文介绍 讨论Token Economy，认为推理代币（inference tokens）已成为追踪AI生态系统增长的主要代理，并预测数据将成为下一个万亿美元类别。

Own the Outer Loop

@addyosmani · 405.4K 粉丝 · 396.4K 阅 · 547 赞 · 69 转

07/09 02:44

In the past year, the conversation around agentic engineering has moved to harnesses and loops, fleets and software factories. My 2c is engineers need to own the outer loop - the accountability for

中文介绍 在自主工程中，工程师应掌握外层循环（outer loop），对系统结果负责，而非仅关注内循环的工具和流程。

This prompt will change your life

@EXM7777 · 123.5K 粉丝 · 233.6K 阅 · 554 赞 · 38 转

07/07 22:31

for most people, fable has a few hours left to live... this is the smartest model that ever existed and everyone's spending its last night building one more app i made it do something 100x more

中文介绍 在Fable模型生命末期，用提示词让模型做“100倍更智能”的事，强调其价值远超简单构建应用。

How to become an applied AI engineer

@eyad_khrais · 20.2K 粉丝 · 225.8K 阅 · 508 赞 · 38 转

07/07 23:42

I wrote this as the guide I would have wanted before making the transition into Applied AI Engineering. The role broadly overlaps with traditional software engineering, but it adds a few important

中文介绍 提供成为应用AI工程师的指南，指出该角色与传统软件工程重叠但有关键差异，适合转型者参考。

The Heretic’s Guide to AI’s Stars Part III: AI Demand, Offshore Financing, & Compression Too

@michaeljburry · 2.0M 粉丝 · 209.2K 阅 · 508 赞 · 51 转

07/08 08:23

“They are just flying empty airplanes around.” - Son NVIDIA is the North Star, Orion, the whole Milky Way. NVIDIA is benefitting from strong demand, but is selling into a concentrated set of buyers

中文介绍 分析NVIDIA在AI需求中的位置，认为其受益于强劲需求但买家集中，存在类似“空飞飞机”的融资和压缩风险。

Expanding Managed Agents in Gemini API: background tasks, remote MCP and more

@GoogleAIStudio · 184.1K 粉丝 · 165.4K 阅 · 504 赞 · 62 转

07/08 00:38

Today we’re announcing new capabilities in Managed Agents in Gemini API, including background execution, remote MCP server integration, custom function calling and refreshing credentials across

中文介绍 Gemini API的Managed Agents新增后台执行、远程MCP服务器集成、自定义函数调用和凭证刷新等功能。

Model and effort in Claude Code: knowing more vs. trying harder

@ClaudeDevs · 564.6K 粉丝 · 95.9K 阅 · 1.5K 赞 · 100 转

07/09 00:55

Claude Code gives you two settings that both seem to "make the answer better": the model, and the effort level. But what do these actually do to the output? And how do you know whether to reach for a

中文介绍 解释Claude Code中模型和努力水平两种设置的区别：模型升级增加知识，努力提升增加推理尝试，指导如何按需选择。

You can go from zero to hireable AI engineer in 4 months. Here's the exact path.

@free_ai_guides · 11.1K 粉丝 · 59.1K 阅 · 503 赞 · 114 转

07/08 00:32

AI engineering is one of the best-paid, fastest-growing jobs in tech right now. And the door to it just got wider than it's ever been. Most guides get this wrong. They hand you a wall of theory, tell

中文介绍 声称4个月路径可从零基础达到可雇佣的AI工程师，提供具体路线，强调实践而非理论。

How to Distill a $10M+ AI Researcher

@JIACHENLIU8 · 2.9K 粉丝 · 48.9K 阅 · 508 赞 · 28 转

07/08 02:16

tldr: Research Process > Research Outcomes Everyone says the best researchers have great intuition. They seem to know which questions are worth asking before the field recognizes them as important.

中文介绍 主张优秀研究人员的关键在于研究过程而非成果，拥有直觉知道哪些问题值得问，分享如何提炼出这种能力。

Unicity Labs and Quant Partner to Bring Verifiable Execution to Agentic Trading

@unicity_labs · 127.3K 粉丝 · 4.5K 阅 · 520 赞 · 827 转

07/08 22:22

Today we're announcing a partnership between Unicity Labs and @tryquantio to build the secure foundation for agentic finance. Quant - the AI intelligence and execution layer for the agentic future

中文介绍 Unicity Labs与Quant合作，为代理金融构建可验证执行的安全基础，专注于代理交易中的AI执行层。

A Google Knowledge Panel Isn't a Vanity Badge. Here's What It Actually Is - and How I Earned Mine.

@krudhrah · 1.2K 粉丝 · 1.3K 阅 · 517 赞

07/08 00:11

Search my name — "Rudhrah Keshav" — and Google doesn't just show you links. It shows you a panel: my photo, my profession, my date of birth, my education, my podcast, my profiles. All assembled by

中文介绍 分享如何获得Google知识面板，解释它不仅仅是虚荣标志，而是由算法汇编的个人信息展示，并给出实际获取方法。

Anthropic found a hidden space where Claude puzzles over concepts

MIT Tech Review AI · 07/10 04:22

The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving. Researchers at the company built a tool called the

中文介绍 Anthropic公司开发了一种新技术，得以揭示大型语言模型在回答问题或执行任务时的内部运作机制，发现的结果既有平凡也有令人不安之处。

GPT-5.6 is now the preferred model in Microsoft 365 Copilot

OpenAI News · 07/09 21:00

Learn how GPT-5.6 powers Microsoft 365 Copilot with stronger AI capabilities across Word, Excel, PowerPoint, Chat, and Cowork for faster, higher-quality work.

中文介绍 GPT-5.6成为Microsoft 365 Copilot的首选模型，在Word、Excel、PowerPoint、Chat和Cowork中提供更强AI能力，实现更快、更高质量的工作。

ChatGPT is now a partner for your most ambitious work

OpenAI News · 07/09 18:00

ChatGPT Work is an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.

中文介绍 ChatGPT Work作为智能体，可在各类应用和文件中执行操作，持续跟进项目数小时，将目标转化为完成的工作。

GPT-5.5 Bio Bug Bounty

OpenAI News · 07/09 18:00

Details about the OpenAI Bio Bounty program

中文介绍 OpenAI公布GPT-5.5生物安全漏洞赏金计划详情。

GPT-5.6: Frontier intelligence that scales with your ambition

OpenAI News · 07/09 18:00

More intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work.

中文介绍 GPT-5.6提供每token更多智能、更强性价比，按需增强能力以应对最艰巨任务。

[AINews] SpaceXAI launches Grok 4.5, first Opus-class model post Cursor acquisition

Latent Space · 07/09 14:05

SpaceXAI continues to move faster than any other frontier lab on earth.

中文介绍 SpaceXAI推出Grok 4.5，这是收购Cursor后首个Opus级模型，该公司持续以超过任何前沿实验室的速度推进。

not much happened today

Smol AI News · 07/09 13:44

**OpenAI** launched the **GPT-5.6** family including **Sol, Terra, and Luna** models, integrated across **ChatGPT, Codex, and API** with immediate rollout. The release emphasized improved **performance-per-dollar** with pricing matching GPT-5.5 but better capabilities, introducing new features like

中文介绍 OpenAI推出GPT-5.6系列模型，包括Sol、Terra和Luna，即时集成至ChatGPT、Codex和API。定价与GPT-5.5持平但性能更强，强调性价比提升。

Grok 4.5 🤖, GPT-Live 🎙️, SWE-1.7 👨‍💻

TLDR AI · 07/09 08:00

中文介绍 TLDR AI当日简报涵盖Grok 4.5、GPT-Live和SWE-1.7等AI进展。

Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO

Latent Space · 07/09 06:55

2 years after our first coverage, we return with Modal's other cofounder to explore why Agent Experience is working now, and everything they have learned building the new agent cloud.

中文介绍 Modal联合创始人兼CTO Akshat Bubna探讨AI基础设施如何为Agent体验演进，分享构建新型智能体云的经验与洞察。

Data for Agents

Hugging Face Blog · 07/09 01:16

中文介绍 NVIDIA博客文章讨论面向智能体的数据准备话题。

Our approach to government and national security partnerships

OpenAI News · 07/08 21:30

Learn how OpenAI approaches government and national security partnerships, with principles for responsible AI use, democratic accountability, and public safety.

中文介绍 OpenAI阐述其政府及国家安全合作方针，强调负责任AI使用、民主问责与公共安全原则。

Separating signal from noise in coding evaluations

OpenAI News · 07/08 21:00

A new analysis from OpenAI reveals issues in SWE-Bench Pro, a popular coding benchmark, raising concerns about reliability and accuracy in evaluating AI models.

中文介绍 OpenAI新分析指出流行编程基准SWE-Bench Pro存在问题，引发对AI模型评估可靠性与准确性的担忧。

Helping K–12 educators build practical AI skills

OpenAI News · 07/08 18:00

OpenAI Academy and the Walton Family Foundation are bringing hands-on AI Skills Jams to help K–12 educators build practical AI skills for the classroom.

中文介绍 OpenAI Academy与沃尔顿家族基金会合作举办AI Skills Jams工作坊，帮助K-12教育工作者培养课堂实用AI技能。

not much happened today

Smol AI News · 07/08 13:44

**xAI** publicly launched **Grok 4.5**, a new coding-and-agents-focused frontier model emphasizing capability-per-dollar rather than benchmark supremacy. Elon Musk described it as "Opus-class" but faster, more token-efficient, and lower cost, with a **1.5 trillion parameter** size, making it 3x larg

中文介绍 xAI发布Grok 4.5，专注于编程和智能体，强调每美元能力而非基准领先。Elon Musk称其为“Opus级”但更快、token效率更高且成本更低，参数规模达1.5万亿。

[AINews] Lilian Weng summarizes 35 papers on Harness Engineering for RSI

Latent Space · 07/08 10:20

a quiet day lets us read some condensed insight

中文介绍 Lilian Weng总结35篇关于Harness Engineering for RSI的论文，提供浓缩洞见。