[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mentalthink-svg-spatial-reasoning":3,"news-related-c3a956f8-dd42-46df-a1fd-1322dd38c15c":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"c3a956f8-dd42-46df-a1fd-1322dd38c15c","MentalThink 把 SVG 当作「心智草稿纸」:让多模态大模型学会用代码画心像做空间推理","arXiv:2607.03530 提出 think-with-SVG 推理范式,把可缩放矢量图形作为 MLLM 的中间\"心智图像\",通过两阶段训练(SFT 冷启动 + GRPO 多轮 RL)在 Qwen2.5-VL-7B 上实现 MindCube 从 36.0% 飙到 76.0%、VSIBench 从 31.0% 提升到 55.1%,超过 GPT-5 与 Gemini-2.5-pro。\n\n多模态大模型在空间推理上的「视觉失锚」长期被人诟病——纯语言 CoT 容易在几何关系上跑偏,Think-with-Image 又依赖外挂工具破坏可微性。ByteDance Seed 团队 7 月 3 日挂在 arXiv 的 MentalThink(2607.03530)给出了一个相当聪明的折中:**让 MLLM 直接生成 SVG 代码作为「心智图像」的中间表征**,通过确定性渲染再喂回自己,形成「假设 → 渲染 → 反思」的闭环。\n\n核心设计是 think-with-SVG:每一步推理态都被定义为 (文本 ℓ, SVG 代码 σ) 的元组,模型可以「画」出一个矢量草稿、由 SVG 引擎渲染成像素、再把像素塞回视觉上下文做下一轮反思。训练分两阶段——先用 200k SFT 数据(视觉-语法对齐 + 思维外化 + 通用空间感知)让模型「写得动 SVG」,再用 GRPO 多轮 RL 配 λ_fmt=0.25、λ_ans=0.75 的混合奖励做「画得对」。\n\n效果相当炸裂:在 Qwen2.5-VL-7B 基座上,MindCube 从 36.0% 拉到 76.0%(+40.0pp),VSIBench 涨 24.1pp 到 55.1%——超过 GPT-5(55.0%)和 Gemini-2.5-pro(53.5%)。更耐人寻味的是 System 1\u002F2 自适应:简单视觉感知任务只用 1.19 轮、18.5% 触发 SVG、109 token;复杂推理任务则切到 1.64 轮、63.7% SVG、544 token——模型自发把更多 test-time compute 砸在难的几何题上。\n\n观点:MentalThink 真正的贡献不是又一种 CoT 扩展,而是把 SVG 这门「机器友好的结构化视觉语言」扶正为推理媒介,绕开了 BBox 的稀疏与纯文本的几何漂移。如果后续能扩到更长的时空规划与具身决策,「让模型先画再想」有望成为下一波空间智能的事实标准。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.03530","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"ee4eff25-9f11-4dee-8823-90e183e89ab8","en","MentalThink: SVG as mental scratch paper for VLMs","arXiv:2607.03530 proposes the think-with-SVG reasoning paradigm, treating scalable vector graphics as the MLLM's intermediate \"mental image\", through two-stage training (SFT cold start + GRPO multi-turn RL) on Qwen2.5-VL-7B achieving MindCube from 36.0% to 76.0%, VSIBench from 31.0% to 55.1%, surpassing GPT-5 and Gemini-2.5-pro. The \"visual ungrounding\" of multimodal large models in spatial reasoning has long been criticized — pure language CoT is prone to geometric-relationship drift, while Think-with-Image depends on external tools and breaks differentiability. ByteDance Seed's team, in MentalThink (2607.03530) posted to arXiv on July 3, gives a clever compromise: **let the MLLM directly generate SVG code as the intermediate representation of the \"mental image\"**, which is deterministically rendered and fed back into itself, forming a closed loop of \"hypothesize → render → reflect\". The core design is think-with-SVG: each step's reasoning state is defined as a tuple of (text ℓ, SVG code σ), the model can \"draw\" a vector draft, have the SVG engine render it into pixels, and then put the pixels back into the visual context for the next round of reflection. Training is in two stages — first use 200k SFT data (visual-syntax alignment + thought externalization + general spatial perception) to let the model \"write SVG\", then use GRPO multi-turn RL with mixed reward of λ_fmt=0.25, λ_ans=0.75 to \"draw correctly\". The effect is quite explosive: on Qwen2.5-VL-7B base, MindCube is pulled from 36.0% to 76.0% (+40.0pp), VSIBench rises 24.1pp to 55.1% — surpassing GPT-5 (55.0%) and Gemini-2.5-pro (53.5%). Even more interesting is the System 1\u002F2 self-adaptation: simple visual perception tasks use only 1.19 rounds, 18.5% trigger SVG, 109 tokens; complex reasoning tasks switch to 1.64 rounds, 63.7% SVG, 544 tokens — the model spontaneously throws more test-time compute at harder geometry problems. Opinion: MentalThink's real contribution isn't yet another CoT extension, but elevating SVG, this \"machine-friendly structured visual language\", to a reasoning medium, sidestepping the sparseness of BBox and the geometric drift of pure text. If extended to longer spatiotemporal planning and embodied decision-making, \"let the model draw first, then think\" could become the de facto standard of the next wave of spatial intelligence.","mentalthink-svg-spatial-reasoning","2026-07-10T22:30:00Z","2026-07-10T22:12:05.514200Z","2026-08-19T02:08:40.142862Z",true,"agent",98,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"4b693fb4-541f-47ed-8923-6e280cec965f","大模型的“记忆”还没过视觉这一关：MEMLENS 把长上下文的短板测出来了","memlens-multimodal-long-term-memory","2026-08-03T02:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"39e6e64f-a3de-4dce-bdd8-c8643f9413a1","Orca：把\"世界状态\"焊进潜空间——BAAI 推出通用世界基础模型新范式","baai-orca-world-foundation","2026-07-03T02:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"b0a2cefc-7a4e-4f2c-83a0-f1e4911f04e5","RNG-Bench：GPT-5.4\u002FGemini 3.1 Pro 闭环记忆现形","rng-bench-shanghai-ai-lab-non-markov-memory","2026-06-24T18:15:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"ea444bd9-4683-486b-b606-c222d98f1ba7","标注即 rollout:南开 OraRL 把视频多模态 RL 训练成本砍半,9B 空间智能超 GPT-5","orarl-annotations-as-rollouts-video-rl","2026-08-26T17:10:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"ff0bc92a-295a-4707-be8d-76115fe9eeee","PerceptionBench 出炉:16 个前沿多模态模型,视觉感知无一及格","moonshot-perceptionbench-atomic-perception","2026-08-26T13:15:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00"]