AI 情报文章归档

High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry. While effective, this design fragments continuous surfaces into many local…

We show that multimodal models possess strong reasoning abilities and that an appropriate harness can unlock their potential to solve tasks across diverse interactive environments. We introduce VISTA, a visual harness that gives a general-…

Full-parameter fine-tuning of large language models (LLMs) incurs substantial optimizer state memory overhead, limiting the model sizes that fit on modern GPUs. Existing approaches either compress optimizer state, abandon first-order gradi…

Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions…

Discrete diffusion language models offer a compelling alternative to autoregressive generation for tasks demanding bidirectional reasoning and global constraint satisfaction. Yet they share a structural bottleneck: when decoding in paralle…

The generalized Schrödinger bridge on a graph moves mass between two distributions while charging a cost for the states visited. It has been approached by learning the rates of a controlled continuous-time Markov chain, with a temporal-dif…

While Large Language Models (LLMs) have demonstrated striking capabilities on frontier mathematical problems, it remains unclear whether they possess the structural mathematical understanding underlying their solutions. In this paper, we t…

Step-size selection remains a central challenge in large-scale neural network optimization; conservative steps slow convergence, while aggressive steps can destabilize it. We combine \textbf{Z}ero-and-\textbf{F}irst-\textbf{O}rder optimiza…

试试 Grok @Bot！

Intrinsically disordered protein regions (IDRs) play central roles in cellular processes such as transcriptional regulation, signal transduction, and subcellular localization, yet their functional design remains challenging. Structure-base…

Distribution Matching Distillation (DMD) trains a few-step student from the difference between separately estimated target and student scores, so it must keep an auxiliary diffusion model fitted to the student's evolving distribution at ex…

Molecular learning models are strongly shaped by their underlying representations. Yet standard sequential and graph formalisms struggle to explicitly encode higher-order topology, such as ring systems and recurring motifs. Existing higher…

Looped Transformers achieve parameter efficiency by repeatedly executing a shared block across recurrent loops. Each loop yields an intermediate representation decodable for the same next token, yet standard decoding discards earlier state…

Quasi-Newton (QN) methods have long been among the most effective methods for large-scale unconstrained convex optimization. Two obstacles have limited their use in deep learning: non-convexity and enormous parameter sizes. We introduce So…

Current controllable video generation systems often rely on 2D motion trajectories or sparse drag signals for object motion. These controls are ambiguous because the same 2D trajectory can correspond to different 3D motions, especially whe…

Multi-teacher on-policy distillation (MOPD) aims to combine the strengths of RL-trained teachers in a single student, but how teacher signals affect parameter changes remains underexplored. We study Qwen3-1.7B with four domain teachers tra…

这 WiFi 真是绝了 （引用推文 @OpenAIDevs：dots 演示。 现在……WiFi 更好了。）

Databricks 为 Lakebase Postgres 推出基于分支的恢复机制，将恢复时间从数小时降至秒级，即使数据库规模达 100 TB。该架构将计算与存储解耦，数据库历史已存于对象存储中可直接引用，恢复只需在指定时间戳创建分支这一元数据操作，无需拷贝数据或重放 WAL。

NVIDIA 新论文提出，与其给 AI 智能体更多选项，不如给它一个更好的裁判：让小型智能体每步生成 8 个候选命令，再由裁判模型挑选执行。用前沿强模型当裁判时，成功率从 50% 提升到 68%，且无需重新训练；若由小模型自评，提升则小得多。

营销公司 Graphite 对前沿模型写作习惯的研究发现，共 13000 个短语在 AI 内容中出现频率至少是人类内容的两倍。

“Upgrade the authentication library safely” sounds like a task. For an agent, it is a bundle of...

Tavus 发布 Griffin，称其为首个通过视频图灵测试的人类交互模型（HIM），48% 与之实时对话的人认为它是真人，此前系统通过率低于 3%。Griffin 在 NVIDIA 全双工 AI 视频基准上得 3.83 分（真人为 3.92），次优系统为 2.80 分；它将听、想、说和面部动画整合在单一系统中，可在对话者仍在说话时读取表情、语气并实时反应。

今天我们推出 PROWL-2，智能体及其世界模型通过递归学习实现共同提升。 智能体暴露想象中存在的错误，修复这些错误又能促成进一步的学习。 我们相信，这种开放式学习是迈向超级智能的关键一步。

按需构建，简单直接。 你获得能更好满足自身需求的自定义工具，无需在市场上购买通用方案。 这就是自驱动公司的愿景。 看看我们在 Replit 是如何实现的：https://replit.com/blog/self-driving-company

Grok Bot 现在无需你开口，就能主动给出帮助建议。

你的主 Bot 会发现它可以帮你分担的工作，并主动提出接手。 它将在接下来几小时内逐步推送，而且建议不会计入你的用量。 下载 Grok Bot：https://x.ai/bot

Ideogram 发布图像编辑模型 Ideogram 4.5，声称只修改用户指定的局部区域，避免多次编辑后出现瑕疵。模型提供 0.8 到 22 美分每图的四个质量档位，均为原生 2K 分辨率，现已在 Ideogram 平台和 API 上线，合作方包括 Picsart、Runway、Pika 和 Leonardo AI，开放权重版本即将推出。

Cloudflare 发布两款决策模型 Clef 和 Clef-flash，托管在 Workers AI 上，并以 Apache 2.0 许可证在 Hugging Face 完全开源。

Figma 将 MCP 访问权限收紧至白名单客户端，Pi 不在其中。此举意味着非白名单客户端无法再通过 MCP 接入 Figma。

This is a submission for the Sanity Challenge, Path Two: Vibe-Code Something Strange What I...

dots demo。 现在……WiFi 更好了。

Cohere 发布 Embed 5 嵌入模型家族，分 Pro 和 Fast 两档，均支持文本、图像及图文融合输入、100+ 语言和 128K token 上下文，两档共享同一向量空间，可用 Pro 建索引、Fast 查询。

美国联邦法官 Amit Mehta 驳回了 Chegg 和 Rolling Stone 母公司 Penske Media 对 Google 的反垄断诉讼，两家出版商指控 Google 强迫出版商免费向 AI Overviews 供稿，否则从搜索结果中消失。

当 AI 生成内容变得如此逼真，我们如何验证哪些是人、机器或自然创造的？ 我们在播客中探讨溯源的关键作用——以及不可感知的水印如何同时保护数字媒体和生物设计。↓ 时间轴： 00:00 介绍 00:34 什么是水印？ 04:35 SynthID 15:16 图像与视频 19:28 SynthID Bio 28:00 韧性的未来

Simon Willison 的纯 Python WebAssembly 引擎 pwasm 发布 0.2a0，经 Claude Opus 5.5 完成 42 次提交后已支持几乎全部 WASM 规范。PyPI 上的 wheel 现已打包可用的 MicroPython、QuickJS 和 Micro QuickJS 的 WASM 构建，因稳定性未经验证仍标记为 alpha 版本。

自从有了好用的 AI，坐着听 PowerPoint 演示变得好多了。当然，有些人干脆让 AI 替他们做思考工作，但这很明显（而且他们以前就用默认模板）。对于真正用心的人来说，他们现在有了有趣的排版、好笑的视觉梗和不错的图表。

POV：你是 @1vnzh 的 10,000 名实习生之一

Why scalar RLHF fails at credit assignment, why barking direct code fixes turns you into an exhausted human compiler, and how Socratic questioning plus topological scars prune doomed search trees at depth d=1.

Boston Dynamics 给 Atlas 配备了新的 4 指机械手，目标是使用人类工具而非仅抓箱子，量产目标为每年 10 万只。新手有 13 个自由度（旧 3 指手为 7），4 自由度对掌拇指可沿每根手指滑动并对捏，手指可侧向张开以持工具柄同时操作扳机，覆盖电钻、砂轮机、射钉枪和焊枪；仍能搬运 100lb+ 的满载小冰箱。

《华尔街日报》用 AI 扫描 400 万条 TikTok，发现反对 AI 的声音正在把政治立场对立的创作者聚集到一起。愤怒的目标包括警方使用的 Flock Safety AI 车牌摄像头以及数据中心，报道截图显示这种反对正体现在中期选举中。