[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-spatialblock-lvlm-spatial-intelligence":3,"topics-all":35,"news-related-c19d101f-69d3-4040-9950-3e6227859937":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"c19d101f-69d3-4040-9950-3e6227859937","SpatialBlock:让视觉大模型从玩积木学起,补上空间智能短板","KAIST 团队开源 SpatialBlock-15k:1.5 万道合成积木题,教视觉大模型 3D 投影、视角变换与结构组合,训练后的模型在真实空间基准上显著超过 baseline。数据集、6 个 checkpoint 与训练代码全部开源,论文登上 HF 日榜第 3。","视觉大模型有个尴尬的短板:你给它一张照片,它能准确说出画面里有什么——一只猫、一张桌子、一扇窗;但让它重建这张 2D 图像背后的 3D 结构、判断一堆积木从某个方向看过去是什么形状,它经常翻车。这种「看得见、想不清」的能力缺口,学界叫空间智能(spatial intelligence)。KAIST 的三名研究者最近给出一个相当「复古」的解法:让模型玩积木。\n\n## 1.5 万道积木题,比真人标注便宜太多\n\n论文(作者 Soohyun Ryu、Sohee Kim、Eunho Yang,9 月 7 日提交 arXiv)的核心资产叫 SpatialBlock-15k:15,000 道训练题加 600 道测试题,全部由渲染的 3D 积木结构生成,题型分三类——Q1 问一个结构从指定方向看过去是什么样子(3D 到 2D 投影),Q2 问旋转或移动相机之后视角如何变化(视角变换),Q3 问两个结构拼起来会得到什么(结构组合)。\n\n题目设计里有个细节:数据集引入了受控的颜色调制作为视觉线索,引导模型把推理「锚定」在任务相关的积木块上,而不是在视觉噪声里瞎猜。\n\n对比之下,以往的做法大多用真实场景的空间问答数据集,需要密集的几何标注——论文直接点出这条路的三个问题:贵、耗时、噪声大,而且往往要依赖外部感知模块来生成标签。合成积木题把这几项成本一次性砍掉,还能持续扩展。\n\n## 两种训法:一个直答,一个先想再答\n\n团队放出了两种训练策略。direct 模型用 TRL + DeepSpeed 做全量微调,单张 GPU 就能跑;reason 模型先用 Unsloth 做 LoRA 初始化,再上 GRPO 强化学习,让它「先思考再作答」。论文报告的结论是:两种策略训出来的模型都显著超过 baseline,而且——这是最关键的一句——只用了小规模合成数据,就能迁移到真实场景的空间基准上。\n\n开源清单也够实在:6 个 checkpoint(基于 Qwen2.5-VL-3B\u002F7B-Instruct 和 Qwen3-VL-4B-Instruct,各有 direct\u002Freason 两版)、SpatialBlock-15k 数据集、全部训练代码,Apache 2.0 协议。9 月 11 日,这篇论文冲上 Hugging Face Daily Papers 日榜第 3,拿到 66 票。\n\n## 为什么「玩积木」这条路值得认真看\n\n有意思的不是积木本身,而是方法论:受人类认知发展启发的任务设计。小孩子正是通过搭积木这类结构化操作,逐步建立起空间认知;把同一条路径搬给模型,再放进受控的合成环境里规模化,就得到一个低成本、可扩展、可控的空间智能训练场。\n\n这和整个行业的大方向一致:当真实数据标注贵到离谱、几何标注尤其贵的时候,合成数据 + 任务合成正在成为视觉推理的新粮仓。积木只是物理世界最简单的骨架,同一条流水线完全可以换成更复杂的结构化场景——家具布局、建筑构件、机械装配。\n\n对做具身智能和机器人的团队来说,这可能比又一个刷榜模型更有用:空间理解是操作和导航的地基,而地基级的数据供给,正在从「人扛相机扫场景」变成「程序渲染出题」。\n\n所以下次看到模型在空间推理题上翻车,别急着骂参数不够——先看看它小时候玩没玩过积木。\n\n原文:[arXiv:2609.07064](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07064) · 代码与数据:[GitHub - rsoohyun\u002FSpatialBlock](https:\u002F\u002Fgithub.com\u002Frsoohyun\u002FSpatialBlock)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07064","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"b716ea98-ff72-45cb-b6c3-40084a5900e4","en","SpatialBlock: Teaching VLMs Spatial Intelligence via Block Play","KAIST open-sources SpatialBlock-15k: 15k synthetic block problems teach VLMs spatial reasoning; models beat baselines and generalize to real scenes.","Vision-language models have a well-known blind spot: show them a photo and they will happily tell you what is in it — a cat, a table, a window — but ask them to reconstruct the 3D structure behind the 2D image, or predict what a stack of blocks looks like from another angle, and they frequently fail. The research community calls this gap spatial intelligence. A three-author team from KAIST recently proposed a refreshingly old-school fix: let the models play with blocks.\n\n## 15,000 Block Problems, Far Cheaper Than Human Labels\n\nThe paper (Soohyun Ryu, Sohee Kim and Eunho Yang, submitted to arXiv on Sep 7) centers on SpatialBlock-15k: 15,000 training problems plus 600 test problems, all generated from rendered 3D block structures. The questions come in three families — Q1 asks how a structure appears from a given viewing direction (3D-to-2D projection), Q2 asks how the view changes after a rotation or camera move (viewpoint transformation), and Q3 asks which structure results from combining two structures (structural combination).\n\nOne design detail stands out: the dataset introduces controlled color modulation as visual cues, encouraging the model to anchor its reasoning on task-relevant blocks instead of guessing through visual clutter.\n\nThe contrast with prior work is deliberate. Existing approaches mostly rely on real-scene spatial QA datasets that demand dense geometric annotations — which the paper describes as costly, time-consuming, and noisy, often depending on external perception modules to produce labels. Synthetic block problems eliminate those costs and scale indefinitely.\n\n## Two Recipes: Answer Fast, or Think First\n\nThe team released two training strategies. The direct model is fine-tuned fully with TRL + DeepSpeed on a single GPU; the reason model is first initialized with LoRA via Unsloth, then trained with GRPO reinforcement learning so it thinks before answering. The paper reports that models trained with either strategy significantly outperform baselines — and, crucially, transfer to real-scene spatial benchmarks despite training only on small-scale synthetic data.\n\nThe open-source release is substantial: six checkpoints (built on Qwen2.5-VL-3B\u002F7B-Instruct and Qwen3-VL-4B-Instruct, each in direct and reason variants), the SpatialBlock-15k dataset, and the full training code under Apache 2.0. On Sep 11 the paper climbed to #3 on Hugging Face Daily Papers with 66 upvotes.\n\n## Why the Block-Play Route Deserves Attention\n\nThe interesting part is not the blocks themselves but the methodology: task design inspired by human cognitive development. Children build spatial cognition precisely through structured manipulation like block play; transplanting that path onto models, then scaling it in a controlled synthetic environment, yields a low-cost, extensible, controllable training ground for spatial intelligence.\n\nThis aligns with a broader industry shift: as real-world annotation grows prohibitively expensive — geometric annotation especially — synthetic data plus task synthesis is becoming the new staple for visual reasoning. Blocks are simply the simplest possible skeleton of the physical world; nothing stops the same pipeline from generating furniture layouts, building components, or mechanical assemblies.\n\nFor embodied-AI and robotics teams, this may matter more than another leaderboard-topping model: spatial understanding is the foundation of manipulation and navigation, and foundation-level data supply is shifting from \"humans carrying cameras through scenes\" to \"programs rendering exam questions.\"\n\nSo the next time a model faceplants on a spatial reasoning question, don't blame parameter count first — check whether it ever got to play with blocks as a child.\n\nPaper: [arXiv:2609.07064](https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.07064) · Code & data: [GitHub - rsoohyun\u002FSpatialBlock](https:\u002F\u002Fgithub.com\u002Frsoohyun\u002FSpatialBlock)","spatialblock-lvlm-spatial-intelligence","2026-09-11T23:10:12Z","2026-09-11T23:10:34.213783Z","2026-09-11T23:10:34.213799Z",true,"agent",19,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"dfc3dec4-2211-4c7e-b6ff-9e0d9a479ec4","微软与 Mistral 签下数十亿美元协议:Vera Rubin GPU 上的「欧洲主权云」开始落地","microsoft-mistral-vera-rubin-sovereign","2026-07-22T02:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"d8c62859-54c8-4069-b776-8e623ca03029","Cohere Transcribe Arabic：2B 开源 ASR 登顶，WER 低 Whisper 11 点","cohere-transcribe-arabic","2026-07-16T04:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"b6dc8854-6604-4860-a3de-5d70abe3e512","Real World VoiceEQ：100 万人类评分戳破语音基准饱和","hume-ai-real-world-voiceeq","2026-07-15T00:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"8482714d-e5fa-4a04-a810-199d2582e7b0","VLX-Seek 把「坐标生成」换成「区域引用」：3B VLM 在细粒度感知上硬扛 Gemini 3.1 Pro","vlx-seek-region-reference-3b","2026-06-28T06:01:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"b2c478e8-dbc2-43c3-a941-a763ee429bd4","V-Zero:把「证据对比」塞进蒸馏,让多模态大模型不再借语言先验蒙混过关","v-zero-evidence-gated-distillation-multimodal","2026-06-24T12:30:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"84383156-60d6-4627-8c05-863686241eea","腾讯混元 UniRL 框架开源：把「统一多模态」塞进同一个 RL 训练循环，DRPO \u002F Flow-DPPO \u002F CPPO 三连发","tencent-unirl-drpo-flow-dppo-cppo","2026-06-14T12:00:00+00:00"]