Lit Review

Use MULTIPLE search strategies in parallel:

1a. arXiv API (primary)

curl -s "http://export.arxiv.org/api/query?search_query=all:{query}&start=0&max_results={N}&sortBy=submittedDate&sortOrder=descending" | head -500

• Build query: all:keyword1+AND+all:keyword2 (max 2-3 terms per query, AND logic)
• Run 2-3 variant queries to cover different angles
• Fetch 30-80 papers total
• Parse XML: extract <entry> blocks with title, summary, id, published, authors

1b. OpenAlex API (backup, broader coverage)

When arXiv results are insufficient (e.g., non-CS fields, interdisciplinary topics, or need citation counts):

curl -s "https://api.openalex.org/works?filter=title_and_abstract.search:{query},from_publication_date:{year_min}-01-01&sort=relevance_score:desc&per_page=50&select=id,title,authorships,publication_year,primary_location,abstract_inverted_index,cited_by_count,concepts" | python3 -c "import sys,json; print(json.dumps(json.load(sys.stdin),indent=2,ensure_ascii=False))" | head -500

• Free, no API key needed (polite pool: add [email protected] param for faster rate)
• Covers journals, conferences, preprints — much broader than arXiv alone
• Provides cited_by_count for importance ranking
• abstract_inverted_index needs reconstruction: " ".join(w for w,_ in sorted(((w,p) for w,pos in idx.items() for p in pos), key=lambda x:x[1]))

1c. OpenReview API (backup, for ML venues)

When the topic targets specific ML venues (NeurIPS, ICLR, ICML) and you need review scores or acceptance status:

curl -s "https://api.openreview.net/notes/search?query={query}&limit=50&source=forum" | python3 -c "import sys,json; print(json.dumps(json.load(sys.stdin),indent=2,ensure_ascii=False))" | head -500

• Best for: NeurIPS, ICLR, ICML, EMNLP papers with peer review data
• Can filter by venue: add &venue=ICLR.cc/2025/Conference

1d. grok-search (supplementary)

Use grok-search skill to find:

• Seminal/highly-cited works that API searches might miss
• Very recent preprints or blog posts about the topic
• Survey papers that already exist on this topic

1e. ACE tool (if codebase-related)

If the topic relates to code in the current project, use mcp__ace-tool__search_context to find relevant implementations, then trace back to their source papers.

Search strategy priority

1. arXiv API — always first, best for CS/ML/AI preprints
2. OpenAlex — when arXiv coverage is thin, or need citation data
3. OpenReview — when targeting specific ML venue papers
4. grok-search — fill gaps, find seminal works, latest news

Output of Step 1: A deduplicated list of papers with title, authors, year, abstract, source_url.

Step 2: Screen & Rank (1-2min)

Read ALL abstracts yourself. For each paper, assign:

• High: Directly addresses the user's focus questions
• Medium: Related methodology or adjacent problem
• Low: Tangential or irrelevant

Selection criteria:

• Keep all High papers (target: 5-15)
• Keep Medium papers only if total High < 8
• Drop all Low papers

Sort selected papers by relevance (not date).

Step 3: Deep Read Key Papers (3-5min)

For the top 5-10 High papers:

3a. Fetch extended content

Use WebFetch on https://ar5iv.labs.arxiv.org/html/{arxiv_id} (HTML version of arXiv papers, much easier to parse than PDF).

If ar5iv is unavailable, fall back to the abstract page: https://arxiv.org/abs/{arxiv_id}

3b. Extract key information

For each paper, note:

• Problem: What problem does it solve?
• Method: Core technical approach (1-3 sentences)
• Results: Key quantitative results or claims
• Limitations: Acknowledged or apparent weaknesses
• Relevance: How it connects to the user's questions

3c. Identify connections

• Which papers build on each other?
• What are the competing approaches?
• Where do results conflict?

Step 4: Generate Report (2-3min)

Write a structured Markdown report in the user's preferred language (default: Chinese).

Report Template

# {Topic} 文献调研报告

> 生成时间: {date} | 检索范围: arXiv {year_min}-{year_max} | 筛选: {total_found} 篇中选出 {selected} 篇

## 1. 研究现状概览
<!-- 2-3 段，勾勒该领域的整体脉络和发展趋势 -->

## 2. 关键方法分类
<!-- 按方法/思路分类，不是按时间 -->

### 2.1 {Method Category A}
- 核心思路
- 代表工作: [Paper1], [Paper2]
- 优劣势

### 2.2 {Method Category B}
...

## 3. 关键论文详评

| # | 论文 | 年份 | 核心贡献 | 方法要点 |
|---|------|------|----------|----------|
| 1 | [Title](arxiv_url) | 2025 | ... | ... |
| ... |

### 3.1 {Paper Title}
- **问题**: ...
- **方法**: ...
- **结果**: ...
- **局限**: ...

<!-- repeat for top papers -->

## 4. 方法对比

| 方法 | 优势 | 劣势 | 适用场景 |
|------|------|------|----------|
| ... | ... | ... | ... |

## 5. Research Gaps 与未来方向
<!-- 基于以上分析，指出尚未解决的问题和值得探索的方向 -->

## 参考文献
<!-- 所有引用的论文，按出现顺序编号 -->
[1] Author et al. "Title." arXiv:XXXX.XXXXX, Year.

Report output

• Save to: ~/Desktop/lit-review-{topic-slug}-{date}.md
• Also display a concise summary (5-8 sentences) in the conversation

Step 5: Persist to Memory Palace

Save key findings to Memory Palace for future reference:

mcp__memory-palace__create_memory:
  title: "Lit Review: {topic}"
  body: |
    ## Summary
    {3-5 sentence overview}
    ## Key Papers
    {top 5 papers with one-line descriptions}
    ## Key Findings
    {main takeaways relevant to user's work}
  domain: "research"
  path: "lit-review/{topic-slug}"
  priority: 1

This ensures future conversations can recall what was surveyed without re-running the review.

Key Principles

1. You are the reviewer — don't delegate to external tools. Read, think, synthesize.
2. Abstracts are enough for screening — only fetch full text for papers you're actually analyzing.
3. arXiv keyword tips (from FWMA lessons): max 2-3 keywords per query, AND logic, too many = 0 results.
4. Prefer breadth first — it's better to survey 30 papers shallowly than 5 papers deeply, then zoom in on what matters.
5. Be honest about limitations — you can't read PDFs with complex math/figures as well as a human. Flag when a paper needs human attention.
6. Chinese by default — unless user specifies otherwise, report in Chinese.