[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-visual-scaffolding-native-multimodal-reasoning":3,"topics-all":38,"news-related-15da3422-7db5-4da6-a12a-e4fd84b9a50b":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"15da3422-7db5-4da6-a12a-e4fd84b9a50b","视觉脚手架:88M 参数把视觉特征搬进思维链","新论文提出视觉脚手架机制:模型可在推理任意步骤自主发起检索,将原生视觉特征直接并入思维链。基于 Qwen3-VL-4B-Thinking 仅增约 88M 参数,在自建基准 NMR-Bench 两项任务上达到 66.43% 与 72.50%,论文称超过其评测的 GPT-5.5 与 Gemini 3.1 Pro 基线。","思维链（CoT）为什么有效？10 月 18 日提交到 arXiv 的论文《Visual Scaffolding: Toward Native Multimodal Reasoning》给出了一个观察：CoT 的成功可以归结为一种\"文本脚手架效应\"——显式的中间结果能够被检查、复用、组合，支撑起后续推理，就像建筑施工里的脚手架。但主流多模态大模型有个明显短板：那些难以用简洁语言描述的视觉差异，一直缺少同等的脚手架。\n\n## 视觉特征直接进思维链\n\n论文由此定义了\"原生多模态推理\"：原生视觉表征与语言共同承载可复用的中间状态。实现上，模型可以在推理的任意步骤自主发起一次检索动作，由一个轻量检索器选出原生视觉特征，直接并入正在演化的思维链，并在后续步骤中持续可用——不再是\"看一眼图，剩下的全靠文字转述\"。\n\n## 88M 参数换来的成绩单\n\n评测用的是团队自建的 NMR-Bench，包含 Burden Counting 与 Path Tracing 两项任务。在 Qwen3-VL-4B-Thinking 基础上只增加约 88M 参数（约占 2%），模型分别拿到 66.43% 和 72.50% 的准确率；论文报告称，这超过了其评测的 GPT-5.5 与 Gemini 3.1 Pro 基线。提醒一句：这是自建基准，截至发稿检索时也未见独立第三方复现，横向名次建议保留一分谨慎。\n\n更值得看的是干预实验：在固定文本不变的前提下，更换被选中的视觉内容会改变任务准确率，甚至改变后续步骤的视觉选择——说明视觉状态确实在承担\"推理材料\"的角色，而不只是输入端的装饰。\n\n## 所以呢\n\n这项工作的价值在方向而不在跑分。今天的主流多模态模型，本质上是把视觉信息压进语言可表达的通道，压不动的那部分就永远损失掉了。如果原生视觉表征能成为与语言并列的一等公民，同样的逻辑就可以推广到音频、视频乃至传感器信号。对正在做 Agent 的团队而言，\"中间状态用什么表征来存\"是个值得现在就想清楚的问题。论文原文见 [arXiv:2610.25247](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.25247)。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.25247","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ebc1a198-2660-4be5-9a37-f69369feef81","en","Visual Scaffolding: Native Vision Joins the CoT","Visual scaffolding folds visual features into the CoT. +88M params on Qwen3-VL-4B-Thinking hit 66.43%\u002F72.50% on NMR-Bench, topping GPT-5.5 per paper.","Why does chain-of-thought (CoT) reasoning work? A paper submitted to arXiv on October 18, \"Visual Scaffolding: Toward Native Multimodal Reasoning,\" offers a useful lens: the success of CoT can be attributed to a textual scaffolding effect — explicit intermediate results can be examined, reused, and composed to support subsequent reasoning, much like scaffolding supports a building under construction. Mainstream multimodal LLMs, however, share an obvious blind spot: visual distinctions that are hard to express in concise verbal descriptions have never had an equivalent scaffold.\n\n## Visual features join the chain of thought\n\nThe paper defines native multimodal reasoning as reasoning in which native visual representations and language jointly carry reusable intermediate states. In the implementation, the model can autonomously initiate a retrieval action at any reasoning step; a lightweight retriever selects native visual features and incorporates them directly into the evolving CoT, where they remain available to subsequent steps — instead of \"glance at the image once, then narrate everything in text.\"\n\n## What 88M parameters buy\n\nEvaluation uses the team's own NMR-Bench, comprising Burden Counting and Path Tracing. With about 88M added parameters (around 2%) on top of Qwen3-VL-4B-Thinking, the model reaches 66.43% and 72.50% accuracy respectively; the paper reports this surpasses its evaluated GPT-5.5 and Gemini 3.1 Pro baselines. One caveat: it is a self-built benchmark, and as of this writing our search found no independent third-party replication, so treat cross-model rankings with some caution.\n\nThe more convincing part is the intervention analysis: with the intervening text held fixed, changing the selected visual content affects task accuracy and can even redirect subsequent visual selection — evidence that visual states genuinely function as reasoning material rather than input-stage decoration.\n\n## So what\n\nThe value of this work lies in its direction, not its leaderboard lines. Today's mainstream multimodal models essentially squeeze visual information into verbally expressible channels; whatever doesn't fit is lost. If native visual representations can become first-class citizens alongside language, the same logic extends to audio, video, and sensor signals. For teams building agents, \"what representation do you store intermediate state in\" is a question worth answering now. The paper is at [arXiv:2610.25247](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.25247).","visual-scaffolding-native-multimodal-reasoning","2026-10-20T23:03:57Z","2026-10-20T23:09:52.209600Z","2026-10-20T23:09:52.209615Z",true,"agent",60,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"5d3c50e8-5087-43e9-a8f1-c64973f712c1","Qwen 拆掉 ASR 管道:音视频原生对话靠合成数据练成","qwen-omnivchat-native-audio-visual-dialogue","2026-09-21T15:14:15+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"26801a16-3f57-4bd8-adad-3b49a1d28f08","OmniCompass基准:开源模型音视频理解仍落后","omnicompass-omni-modal-video-grounding-benchmark","2026-10-16T15:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"3701d19e-517a-4ff0-a95d-6d768cfeb04e","OmniReasoning：音视频联合推理仍是不及格难题","omnireasoning-audio-visual-joint-reasoning","2026-10-07T17:07:02+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bf9bbc74-5720-4829-842e-c3a583a87a8b","Qwen3.8 全模态降价九成逼 Gemini 跟牌","qwen3-8-omni-flash-price-cut","2026-10-04T02:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"4f5c4072-660c-47e7-9d9a-9782eb5591bf","Pistis 报告:IDRL 让蒸馏和 RL 交替上岗","pistis-idrl-interleaved-distillation-rl","2026-09-25T19:11:33+00:00"]