"/typst/templates/paper-slides.typ": body-s = 14pt head-s = 17pt : paper-slides.with( title: "The AI Scientist v1 → v2: 全自动科研框架演进", subtitle: "从线性流水线到 Agentic Tree Search", author: "Biscuit", date: "2026-07-09", ) -slide[ (center + horizon)[ (size: 28pt, weight: "bold", fill: accent)[The AI Scientist] (0.3em) (size: 20pt, fill: accent)[v1 → v2: 全自动科研框架的演进] (2em) (size: 14pt, fill: secondary-text)[Sakana AI · Oxford · UBC · Vector Institute] (0.3em) (size: 13pt, fill: secondary-text)[arXiv 2408.06292 (v1) · arXiv 2504.08066 (v2)] (0.5em) (size: 13pt, fill: secondary-text)[github.com/SakanaAI/AI-Scientist] (1em) -block[ (size: 13pt, weight: "bold", fill: accent)[里程碑：首篇 AI 全自动论文通过 Workshop 同行评审（ICLR 2025 ICBINB, 评分 6.33, top 45%）] ] ] ] 目录 ( columns: (1fr, 1fr), gutter: 2em, align: top, [ ( columns: (0.5fr, 4fr), inset: 6pt, stroke: none, align: (center, left), [ (size: 22pt, weight: "bold", fill: accent)[1]], [ (size: body-s)[v1 → v2 对比总览 \ (fill: secondary-text, size: 12pt)[10 维度演进一览]]], [ (size: 22pt, weight: "bold", fill: accent)[2]], [ (size: body-s)[v1 流水线 + v2 改进 \ (fill: secondary-text, size: 12pt)[四阶段优缺点分析]]], [ (size: 22pt, weight: "bold", fill: accent)[3]], [ (size: body-s)[Tree Search 原理 \ (fill: secondary-text, size: 12pt)[节点模型、搜索策略、参数]]], [ (size: 22pt, weight: "bold", fill: accent)[4]], [ (size: body-s)[实验进度管理器 \ (fill: secondary-text, size: 12pt)[四阶段分配与传播]]], [ (size: 22pt, weight: "bold", fill: accent)[5]], [ (size: body-s)[Workshop 验证 \ (fill: secondary-text, size: 12pt)[ICLR 2025 盲审实验]]], [ (size: 22pt, weight: "bold", fill: accent)[6]], [ (size: body-s)[整体架构设计 \ (fill: secondary-text, size: 12pt)[树图 + 四大组件]]], [ (size: 22pt, weight: "bold", fill: accent)[7]], [ (size: body-s)[AI4Sci 框架启示 \ (fill: secondary-text, size: 12pt)[复用模块、规避陷阱、局限与展望]]], ) ], [ [ (size: body-s, weight: "bold", fill: accent)[核心命题] (0.3em) (size: body-s, fill: secondary-text)[能否构建一个全自动 AI 科学家，从 Idea 到论文不依赖人类介入？] (0.5em) (size: body-s, weight: "bold", fill: accent)[两个版本的回答] (0.2em) ( columns: (1fr, 1fr), inset: 4pt, stroke: 0.4pt, align: (center, center), [ (size: 12pt, weight: "bold")[v1]], [ (size: 12pt, weight: "bold")[v2]], [ (size: 12pt)[需人工模板]], [ (size: 12pt)[零模板通用]], [ (size: 12pt)[线性浅层]], [ (size: 12pt)[树搜索并行]], [ (size: 12pt)[未提交审稿]], [ (size: 12pt)[Workshop 接收]], ) ] (0.3em) (size: 12pt, fill: secondary-text)[v1: arXiv 2408.06292 · v2: arXiv 2504.08066 · github.com/SakanaAI/AI-Scientist] ], ) v1 → v2 对比总览 (size: head-s, weight: "bold", fill: accent)[10 个关键维度的演进] ( columns: (2.2fr, 2.2fr, 2.6fr), inset: 4pt, stroke: 0.4pt, align: (left, center, center), table.header( [ (size: body-s, weight: "bold")[维度]], [ (size: body-s, weight: "bold")[v1]], [ (size: body-s, weight: "bold")[v2]], ), [ (size: 13pt)[代码模板]], [ (size: 13pt, fill: accent)[需人工编写]], [ (size: 13pt, fill: accent)[完全自动生成]], [ (size: 13pt)[实验范式]], [ (size: 13pt)[线性编辑 → 运行]], [ (size: 13pt)[Tree Search + 并行节点]], [ (size: 13pt)[进度管理]], [ (size: 13pt)[MAX ITERS=4, RUNS=5]], [ (size: 13pt)[4 阶段（预研→调参→议程→消融）]], [ (size: 13pt)[视觉反馈]], [ (size: 13pt)[无]], [ (size: 13pt)[VLM 审图（GPT-4o JSON）]], [ (size: 13pt)[论文写作]], [ (size: 13pt)[Aider 逐节+chktex]], [ (size: 13pt)[单次+o1 反思+VLM 图注]], [ (size: 13pt)[审稿机制]], [ (size: 13pt)[单一 GPT-4o]], [ (size: 13pt)[5 集成+AC 聚合]], [ (size: 13pt)[并行能力]], [ (size: 13pt)[无]], [ (size: 13pt)[节点并行执行]], [ (size: 13pt)[LLM 支持]], [ (size: 13pt)[ 10 个 API]], [ (size: 13pt)[ 30 API（o1, Gemini 2.5）]], [ (size: 13pt)[人体验证]], [ (size: 13pt)[未提交审稿]], [ (size: 13pt)[Workshop 接收 6.33/10]], [ (size: 13pt)[耗时/成本]], [ (size: 13pt)[1-2h, 15-250 USD]], [ (size: 13pt)[15h左右, API 费用更高]], ) v1 流水线 + v2 改进 ( columns: (1fr, 1fr), gutter: 0.6em, align: top, [ (title: "Idea 生成")[ (size: body-s)[ (weight: "bold")[v1]：模板变异，reflection 3 轮，Semantic Scholar 10 轮 \ (weight: "bold")[局限]：Idea 相似度高 \ (weight: "bold", fill: accent)[v2]：广义 Idea 生成，开放探索 ] ] (title: "实验迭代")[ (size: body-s)[ (weight: "bold")[v1]：Aider 编辑 → 运行，MAX ITERS=4 \ (weight: "bold")[局限]：线性无并行，深度浅 \ (weight: "bold", fill: accent)[v2]：Tree Search + 4 阶段 + 并行 ] ] ], [ (title: "论文写作")[ (size: body-s)[ (weight: "bold")[v1]：Aider 逐节，引用 20 轮，chktex 5 轮 \ (weight: "bold")[局限]：无视觉验证 \ (weight: "bold", fill: accent)[v2]：+ o1 反思 + VLM 图注审查 ] ] (title: "自动审稿")[ (size: body-s)[ (weight: "bold")[v1]：GPT-4o, 5 reflection, 5 集成 \ (weight: "bold")[结果]：F1 0.57 (vs Human 0.49) \ (weight: "bold", fill: accent)[v2]：+ AC 聚合, Workshop 验证 ] ] ], ) v2 核心: Agentic Tree Search -col( ratio: (1.2fr, 1fr), [ -img("/typst/presentations/figures/experiment tree.pdf", width: 100%) ], [ (title: "搜索策略")[ (size: body-s)[ (weight: "bold")[节点选择]：概率 p 选 Buggy 修复，否则 Non-buggy。Best-First Search。 \ (0.15em) (weight: "bold")[6 种节点]：Buggy / Non-buggy / Debug(≤3) / Refine / Hyperparameter / Ablation / Replication / Aggregation \ (0.15em) (weight: "bold")[执行循环]：Plan → Code → Execute → Plot → VLM → Buggy/Non-buggy ] ] (0.25em) (title: "关键参数")[ (size: body-s)[ Claude 3.5 Sonnet (v2) 代码 · GPT-4o 反馈 · T=0.5 · 单节点超时 1h · 总耗时数h 15h ] ] ], ) v2 实验进度管理器 (size: head-s, weight: "bold", fill: accent)[四阶段 + 节点分配 + 阶段间传播] ( columns: (1.8fr, 1fr, 1.8fr, 2.2fr), inset: 4pt, stroke: 0.4pt, align: (left, center, center, left), table.header( [ (size: body-s, weight: "bold")[阶段]], [ (size: body-s, weight: "bold")[节点数]], [ (size: body-s, weight: "bold")[输入]], [ (size: body-s, weight: "bold")[终止条件]], ), [ (size: 13pt)[1. 预研]], [ (size: 13pt)[21]], [ (size: 13pt)[Idea]], [ (size: 13pt)[最小原型成功执行]], [ (size: 13pt)[2. 调参]], [ (size: 13pt)[12]], [ (size: 13pt)[最佳原型]], [ (size: 13pt)[曲线收敛 + ≥2 数据集]], [ (size: 13pt)[3. 议程]], [ (size: 13pt)[12]], [ (size: 13pt)[调优基线]], [ (size: 13pt)[预算耗尽/自动增复杂度]], [ (size: 13pt)[4. 消融]], [ (size: 13pt)[12]], [ (size: 13pt)[最终节点]], [ (size: 13pt)[复制取统计→聚合出图]], ) (0.3em) ( columns: (1fr, 1fr), gutter: 0.6em, align: top, [ (title: "阶段间传播")[ (size: body-s)[ LLM 评估器选择最佳节点 → 传给下阶段为根节点。每阶段末尾启动复制节点（多 seeds）提供统计量用于最终图表。 ] ] ], [ (title: "节点生命周期")[ (size: body-s)[ 节点 = Plan + Code + Error + Runtime + Metrics + LLM 反馈 + VLM 反馈 + Status。Aggregation 节点不执行实验，仅汇总复制结果。 ] ] ], ) v2 Workshop 验证 (size: head-s, weight: "bold", fill: accent)[ICLR 2025 ICBINB Workshop · 43 篇投稿 · 盲审] ( columns: (2.5fr, 1fr, 1fr, 2.2fr), inset: 4pt, stroke: 0.4pt, align: (left, center, center, left), table.header( [ (size: body-s, weight: "bold")[论文]], [ (size: body-s, weight: "bold")[结果]], [ (size: body-s, weight: "bold")[评分]], [ (size: body-s, weight: "bold")[内部审查发现问题]], ), [ (size: 12.5pt)[Compositional Regularization (LSTM)]], [ (size: 12.5pt, fill: accent)[Accepted]], [ (size: 12.5pt)[6.33]], [ (size: 12pt)[57% 数据重叠, 图注错, 漏引用]], [ (size: 12.5pt)[Label Noise on Calibration]], [ (size: 12.5pt)[Rejected]], [ (size: 12.5pt)[N/A]], [ (size: 12pt)[温度缩放写未跑, 文字图不符]], [ (size: 12.5pt)[Pest Detection Failures]], [ (size: 12.5pt)[Rejected]], [ (size: 12.5pt)[N/A]], [ (size: 12pt)[DA 未跑通, mislead 描述]], ) (0.15em) -col( ratio: (1.3fr, 1fr), [ (title: "被接收论文分析")[ (size: body-s)[ (weight: "bold")[假设]：LSTM embedding 正则化 → 组合泛化 \ (weight: "bold")[发现]：未改进，有时损害。展示负面结果获认可 \ (weight: "bold")[生成流程]： 40 Idea → 选 3 → 多 seeds 全自动运行 → 选最优 → 撤回 ] ] ], [ -img("/typst/presentations/figures/workshop paper.png", width: 90%) ], ) 架构设计 (size: head-s, weight: "bold", fill: accent)[Agentic Tree Search 驱动的全自动科研框架] ( columns: (1fr, 1fr), gutter: 0.6em, align: top, [ -img("/typst/presentations/figures/experiment tree.pdf", width: 100%) ], [ (title: "框架组件")[ (size: body-s)[ (weight: "bold")[Exp. Manager]：4 阶段进度管理 \ (weight: "bold")[Template Contract]： + JSON 解耦 \ (weight: "bold")[VLM Reviewer]：GPT-4o 审图 → Debug/Refine \ (weight: "bold")[LLM Adapter]：Claude / GPT-4o / o1 ] ] (title: "核心机制")[ (size: body-s)[ 并行节点 → Buggy/Non-buggy → Debug/Refine → Best-First。 \ (fill: accent)[v1]：线性，重试 ≤4，5轮 \ (fill: accent)[v2]：并行探索，最佳传播 ] ] ], ) AI4Sci 框架启示 ( columns: (1fr, 1fr), gutter: 0.5em, align: top, [ (title: "可复用 + 需规避")[ (size: body-s)[ (weight: "bold")[可复用]：模板契约（ 解耦）、Tree Search（并行+4阶段）、VLM→物理检查（分子构型/光谱峰）、开放进化（存档+变异+分数反馈） \ (weight: "bold")[陷阱]：数据泄露(57%)→隔离检查点 · 未执行代码→log签名 · 假阳性→baseline+复杂度 · 沙箱逃逸→Docker+cgroups ] ] ], [ (title: "局限与展望")[ (size: body-s)[ (weight: "bold")[v1]：Idea 多样性不足 · 深度浅 · 无法看图 · 幻觉 · 沙箱逃逸 \ (weight: "bold")[v2]：Workshop 非主会 · 1/3 通过 · 数据泄露 · 未执行代码 \ (weight: "bold", fill: accent)[未来]：VLM 集成 · 自驱动实验室 · 开放进化 · 伦理规范 ] ] ], ) (0.3em) (size: body-s, weight: "bold", fill: accent)[v1: AI 可自动完成科研闭环 → v2: AI 论文通过同行评审。每一步都是 AI4Sci 框架的设计参考。] Q&A (3em) (center, text(size: 18pt, weight: "bold", fill: accent)[谢谢！Q & A]) (0.5em) (center, text(size: 11pt, fill: secondary-text)[v1: arXiv 2408.06292 / v2: arXiv 2504.08066 · Code: github.com/SakanaAI/AI-Scientist])