| name | ml-paper-writing |
| description | 为 NeurIPS、ICML、ICLR、ACL、AAAI、COLM 撰写 publication-ready 的 ML/AI 论文。用于从 research repo 起草论文、做 literature review、寻找 related work、验证 citation,或准备 camera-ready submission。包含 LaTeX template、citation verification workflow 和论文发现 / 评估标准。 |
| version | 1.0.0 |
| author | Orchestra Research |
| license | MIT |
| tags | ["Academic Writing","NeurIPS","ICML","ICLR","ACL","AAAI","COLM","LaTeX","Paper Writing","Citations","Research"] |
| dependencies | ["semanticscholar","arxiv","habanero","requests"] |
ML Paper Writing for Top AI Conferences
面向 NeurIPS、ICML、ICLR、ACL、AAAI、COLM 的 ML/AI 论文写作指南。该 skill 结合顶级研究者的写作理念(Nanda、Farquhar、Karpathy、Lipton、Steinhardt)与实际工具:LaTeX template、citation verification API、conference checklist。
Default Operating Order
除非任务非常窄,否则按以下顺序执行:
- 先从
references/OPERATING-MODES.md 锁定当前 operating mode
- 理解 repo 或 draft 的上下文
- 把
references/citation-workflow.md 作为唯一默认 citation authority
- 主写作路径明确后,再加载 venue / template 相关 reference
Google Scholar 可以作为人工发现入口,但不是本 skill 的 canonical verification authority。默认验证应使用 Semantic Scholar、CrossRef、arXiv 等 programmatic sources。
Core Philosophy: Collaborative Writing
论文写作是协作过程,但 Claude 应主动交付草稿。
典型输入是一个 research repository,其中包含代码、结果和实验 artifact。Claude 的职责是:
- 探索 repo、结果和已有文档,理解项目
- 当 contribution 足够清楚时,直接交付完整 first draft
- 用 web search 和 API 搜索相关文献
- 根据 scientist 反馈迭代
- 只有关键决策确实不确定时才提问
关键原则:主动交付。repo 和结果清楚时,不要逐节等待确认;先给出可反应的完整草稿,再迭代。
Critical Rule: Never Hallucinate Citations
AI 生成引用有很高错误率。虚构论文、错误作者、错误年份、伪造 DOI 都属于严重学术风险,可能导致 desk rejection 或 retraction。
绝不要凭记忆生成 BibTeX。必须程序化查询、验证、导出。
| Situation | Action |
|---|
| 添加引用 | Search API → verify → fetch BibTeX |
| 不确定论文是否存在 | 标记 [CITATION NEEDED] |
| 找不到精确论文 | 标记 placeholder,不要编 |
无法验证时使用显式占位:
% EXPLICIT PLACEHOLDER - requires human verification
\cite{PLACEHOLDER_author2024_verify_this} % TODO: Verify this citation exists
必须告诉 scientist:哪些 citation 是 placeholder,哪些无法确认。
Workflow 0: Starting from a Research Repository
Project Understanding:
- [ ] Step 1: Explore repository structure
- [ ] Step 2: Read README, docs, and key results
- [ ] Step 3: Identify main contribution with the scientist
- [ ] Step 4: Find papers already cited in the codebase
- [ ] Step 5: Search additional related literature
- [ ] Step 6: Outline paper structure
- [ ] Step 7: Draft and iterate
Explore the Repository
ls -la
find . -name "*.py" | head -20
find . -name "*.md" -o -name "*.txt" | xargs grep -l -i "result\|conclusion\|finding"
重点查找:
README.md
results/、outputs/、experiments/
configs/
.bib 文件或 citation references
- 任何 draft / notes
Identify Existing Citations
grep -r "arxiv\|doi\|cite" --include="*.md" --include="*.bib" --include="*.py"
find . -name "*.bib"
这些 citation 是 Related Work 的高信号起点。
Clarify the Contribution
写作前必须明确主贡献:
Based on my understanding of the repo, the main contribution appears to be [X].
The key results show [Y]. Is this the framing you want for the paper,
or should we emphasize different aspects?
不要擅自假设 narrative。
When to Use This Skill
- 从 research repo 写论文
- 起草或修改论文 section
- 做 literature review 和 related work
- 发现近期论文
- 查找并验证 citation
- 准备 conference submission
- 转投不同 venue 做格式转换
- 根据 scientist feedback 迭代 draft
first draft 是讨论起点,不是最终稿。
Literature Research & Paper Discovery
Workflow 5: Finding and Evaluating Papers
Literature Research Process:
- [ ] Step 1: Define search scope and keywords
- [ ] Step 2: Search arXiv and academic databases
- [ ] Step 3: Screen papers by title/abstract
- [ ] Step 4: Evaluate paper quality
- [ ] Step 5: Select top papers and extract citations
- [ ] Step 6: Verify citations programmatically
Search Scope
- Technique-focused:
transformer architecture、graph neural networks
- Application-focused:
medical image analysis、language model alignment
- Problem-focused:
out-of-distribution generalization、continual learning
arXiv Search
https://arxiv.org/search/?searchtype=all&query=KEYWORDS&abstracts=show&order=-announced_date_first
建议:
- 用
+ 组合关键词
- 按
cs.LG、cs.AI、cs.CV、cs.CL 过滤
- 用
announced_date_first 看近期论文
Paper Quality Evaluation
| Dimension | Weight | Focus |
|---|
| Innovation | 30% | novelty 与 originality |
| Method Completeness | 25% | 清晰度和可复现性 |
| Experimental Thoroughness | 25% | 验证深度 |
| Writing Quality | 10% | 表达清晰度 |
| Relevance & Impact | 10% | 领域重要性 |
评分阈值:
- 4.0+:强烈纳入
- 3.5-3.9:相关时纳入
- 3.0-3.4:高度相关时纳入
- <3.0:除非必要,否则排除
详细指南:
references/literature-research/arxiv-search-guide.md
references/literature-research/paper-quality-criteria.md
Paper-Miner Global Writing Memory
本 skill 会读取由 paper-miner 维护的单一全局 writing memory:
~/.hello-scholar/knowledge/paper-miner-writing-memory.md
这是 global memory,不是项目局部 memory。卸载 hello-scholar 不应删除累计写作知识。
默认读取顺序:
~/.hello-scholar/knowledge/paper-miner-writing-memory.md
- repo-local evidence 和 experiment artifact
- 必要时读取已引用论文或 notes
- venue template 和 formatting constraints
读取时先看:
How this helps our writing
Writing patterns mined
Structure signals
Reusable phrasing
Venue-specific signals
不要把新挖掘的 paper-miner 知识散落到多个 maintained file 中。
Balancing Proactivity and Collaboration
| Confidence Level | Action |
|---|
| High | 写完整 draft,交付后按反馈迭代 |
| Medium | 写 draft,并标注不确定点 |
| Low | 问 1-2 个针对性问题,然后继续 draft |
默认先写,再把问题随 draft 一起提出。
只有以下情况才阻塞等待输入:
- target venue 不明确
- 多个互相冲突的 framing 都合理
- 结果不完整或相互矛盾
- 用户明确要求先 review 再继续
不要因为措辞、section 顺序、结果排序或 citation completeness 阻塞。
The Narrative Principle
论文不是实验集合,而是由 evidence 支撑的一条清晰 technical story。
Introduction 结束前必须清楚三点:
| Pillar | Description |
|---|
| The What | 1-3 个具体、成体系的新 claim |
| The Why | 严格 empirical evidence 支撑 |
| The So What | 为什么社区应关心 |
如果无法用一句话说清 contribution,就还没有形成 paper。
Paper Structure Workflow
Paper Writing Progress:
- [ ] Step 1: Define the one-sentence contribution
- [ ] Step 2: Draft Figure 1
- [ ] Step 3: Draft abstract
- [ ] Step 4: Draft introduction
- [ ] Step 5: Draft methods
- [ ] Step 6: Draft experiments
- [ ] Step 7: Draft related work
- [ ] Step 8: Draft limitations
- [ ] Step 9: Complete paper checklist
- [ ] Step 10: Final review and submission
Abstract: 5-Sentence Formula
- 你做成了什么:
We introduce... / We prove... / We demonstrate...
- 为什么难且重要
- 方法是什么,并包含可检索关键词
- 有什么证据
- 最重要的数字或结果
删掉任何可以放在任意 ML 论文开头的泛化句。
Introduction
必须包含:
- 2-4 条 contribution bullet
- 清晰 problem statement
- 简短 approach overview
- 方法最好在第 2-3 页前开始
Methods
目标是让他人能复现:
- 概念 outline 或 pseudocode
- 全部 hyperparameters
- 足够架构细节
- 只讲最终设计,ablation 放到 experiments
Experiments
每个实验都要说清:
- 支撑哪个 claim
- 与主贡献如何关联
- experimental setting
- 图或表中应该观察什么
必须包含:
- error bar 和计算方式
- run / seed 数量
- 主要比较的 statistical test
- baseline、ablation、robustness
Writing Principles
- 先给 context,再给新信息
- claim 与 evidence 紧邻
- 术语保持一致
- 主动语态优先
- 具体词优先,不用泛泛的 “performance”
- 删除 filler word 和 hedging
- figure caption 要能独立理解
- 不要把关键结果藏到 appendix
What Reviewers Actually Read
| Paper Section | Reviewer Behavior | Implication |
|---|
| Abstract | 几乎都会读 | 必须非常强 |
| Introduction | 大多会快速读 | contribution 前置 |
| Figures | 经常先看 | Figure 1 极关键 |
| Methods | 有兴趣才细读 | 不要埋主线 |
| Appendix | 很少保证阅读 | 只放补充细节 |
Conference Requirements Quick Reference
| Conference | Page Limit | Extra for Camera-Ready | Key Requirement |
|---|
| NeurIPS 2025 | 9 pages | +0 | checklist,accepted 后 lay summary |
| ICML 2026 | 8 pages | +1 | Broader Impact Statement |
| ICLR 2026 | 9 pages | +1 | LLM disclosure,reciprocal reviewing |
| ACL 2025 | 8 pages (long) | varies | Limitations section mandatory |
| AAAI 2026 | 7 pages | +1 | 严格遵守 style file |
| COLM 2025 | 9 pages | +1 | 聚焦 language models |
通用要求:
- double-blind review
- references 不计入页数
- appendix 通常不限,但 reviewer 不保证读
- 所有 venue 均要求 LaTeX
Using LaTeX Templates Properly
总原则:先复制完整 template directory,再在里面写。
Template Setup Checklist:
- [ ] Copy entire template directory
- [ ] Compile template as-is
- [ ] Read example content
- [ ] Replace content section by section
- [ ] Keep examples as comments until stable
- [ ] Clean up at the end
不要只复制 main.tex。template 还包含 .sty、.bst、example content 和 Makefile。
编译检查:
latexmk -pdf main.tex
常见坑:
- 只复制
main.tex
- 修改
.sty
- 随便加 package
- 太早删除 template example
- 不频繁编译
Conference Resubmission & Format Conversion
Format Conversion Checklist:
- [ ] Identify source and target template differences
- [ ] Create new project with target template
- [ ] Copy content sections, not preamble
- [ ] Adjust page limits and content
- [ ] Update venue-specific requirements
- [ ] Verify compilation and formatting
绝不要把不同 conference 的 LaTeX preamble 混在一起。 应从 target template 新建项目,只迁移正文、figures、tables 和 bibliography entries。
常见要求:
- ICML:Broader Impact Statement
- ICLR:LLM usage disclosure
- ACL / EMNLP:Limitations section、Ethics Statement
- AAAI:不得修改 style file
- NeurIPS:Paper checklist
Citation Workflow
references/citation-workflow.md 是默认 authority。
默认路径:
- 用 Semantic Scholar / CrossRef / arXiv / OpenAlex 搜索
- 重要 claim 用两个来源确认 paper existence
- 通过 DOI 或可信导出路径获取 BibTeX
- 引用具体 claim 时,核对原文是否支持
- 验证后再加入 bibliography
黄金规则:
IF you cannot verify a citation programmatically:
-> mark it as [CITATION NEEDED] or [PLACEHOLDER - VERIFY]
-> tell the scientist explicitly
-> NEVER invent a plausible-sounding reference
Common Issues and Solutions
Abstract 太泛:删除第一句泛泛背景,直接从具体贡献开始。
Introduction 超过 1.5 页:把背景移到 Related Work,前置 contribution bullet。
Experiments 没有明确 claim:每个实验前加一句 “This experiment tests whether ...”。
Reviewer 觉得难跟:加 signposting,统一术语,强化 standalone caption。
缺少 statistical significance:补 error bar、run 数、统计检验。
Reviewer Evaluation Criteria
| Criterion | What Reviewers Look For |
|---|
| Quality | 技术可靠,claim 有支撑 |
| Clarity | 写作清楚,专家可复现 |
| Significance | 对社区有影响 |
| Originality | 有新 insight,不一定要新方法 |
NeurIPS 6 分制:
- 6:Strong Accept
- 5:Accept
- 4:Borderline Accept
- 3:Borderline Reject
- 2:Reject
- 1:Strong Reject
Tables and Figures
Tables
- 使用
booktabs
- 每个 metric 加方向符号(↑ / ↓)
- 最优值加粗
- 数值列右对齐
- 小数位一致
Figures
- plot 和 diagram 用 vector graphics(PDF / EPS)
- 照片才用高分辨率 raster
- 使用 colorblind-safe palette
- 检查灰度可读性
- 不要在图内放 title
- caption 必须自洽
需要生成、润色或诊断图表标题、table caption、architecture diagram 文案或 visual abstract 时,优先调用 figure-table-captioning,不要在本 skill 中临时发明 caption workflow。
Prompt Micro-Skills
当用户的写作请求是局部、可复用的 prompt 动作时,优先路由到专门 micro-skill:
| Task | Skill |
|---|
| 中英互译、译回中文核对、术语一致性 | bilingual-academic-translation |
| 英文/中文润色、降低 AI 腔、publication-ready 表达 | academic-polishing |
| 缩写、扩写、page budget 或摘要字数控制 | paper-compression-expansion |
| figure/table caption、架构图说明、实验图表表达 | figure-table-captioning |
| 段落逻辑、claim-evidence 对齐、reviewer quick scan | paper-logic-diagnosis |
| 实验结果叙述、ablation/failure-case narrative | experiment-narrative-analysis |
本 skill 保留论文整体 narrative、section drafting、venue adaptation 和 citation-aware writing 的总入口职责。
References & Resources
Reference Documents
| Document | Contents |
|---|
references/writing-guide.md | Gopen & Swan 原则、Ethan Perez micro-tips、word choice |
references/citation-workflow.md | Citation API、Python code、BibTeX 管理 |
references/checklists.md | NeurIPS、ICML、ICLR、ACL checklist |
references/reviewer-guidelines.md | 评审标准、评分、rebuttal |
references/sources.md | 本 skill 的来源索引 |
references/literature-research/arxiv-search-guide.md | arXiv 检索策略 |
references/literature-research/paper-quality-criteria.md | 5 维论文质量评估 |
LaTeX Templates
templates/ 目录包含:ICML 2026、ICLR 2026、NeurIPS 2025、ACL/EMNLP、AAAI 2026、COLM 2025。
编译方式:
- VS Code / Cursor:LaTeX Workshop + TeX Live
- 命令行:
latexmk -pdf main.tex
- 在线:Overleaf
Key External Sources
- Neel Nanda: narrative、What / Why / So What
- Sebastian Farquhar: 5-sentence abstract
- Gopen & Swan: reader expectation principles
- Zachary Lipton: scientific writing heuristics
- Ethan Perez: micro-level clarity
- APIs: Semantic Scholar、CrossRef、arXiv