citation-deep-research

name

citation-deep-research

description

Deep literature research by tracing a paper's citation graph. Starting from a seed paper, follow forward citations (papers that cite it) using the Semantic Scholar Graph API plus web search, VERIFY every reported paper against arXiv/OpenReview/PMLR/DBLP, and produce a structured progress report. Use when the user wants to survey research progress, trace how a field evolved from a seed paper, do a citation-based literature review, or focus on specific venues (e.g. NeurIPS/ICML/ICLR) or sub-topics. Triggers: "follow this paper's citations", "deep research on X", "what's the progress since this paper", "调研一下…现在到什么地步了", "沿着引用往下看".

Citation Deep Research

A reproducible method for tracing the research frontier downstream of a seed paper. The whole value of this skill is citation-graph coverage + ruthless source verification — most of the cost of a bad literature survey is hallucinated or mis-attributed papers, so the verification discipline below is not optional.

0. Do NOT scrape Google Scholar

Google Scholar blocks automated access (CAPTCHA + IP bans) and returns unstable results. Even when the user says "use Google Scholar", use these sources instead and say so:

• Semantic Scholar Graph API — the primary citation-graph engine.
• OpenAlex / arXiv / OpenReview / PMLR / DBLP — metadata + verification.
• WebSearch / WebFetch — for very recent work and reading specific pages.

1. Pin down the request (ask only what you can't infer)

Gather, then proceed with sensible defaults rather than over-asking:

1. Seed paper — exact title, arXiv ID, or DOI. Disambiguate (many papers share titles).
2. Direction — forward citations (papers citing the seed → "what came after / current progress", the usual ask) vs backward references (what it built on). Default: forward.
3. Scope — breadth (map all major branches, survey-style) vs depth (track the few most important SOTA lines to the latest). Default: breadth first, then depth on the strongest lines.
4. Focus filters — specific venues (e.g. the big three: NeurIPS, ICML, ICLR), sub-topic, year range. Honor these aggressively when filtering.

If a seed PDF is local and unreadable, install a text extractor on the fly: python3 -m pip install --quiet --user pymupdf then extract with fitz.

2. Pull the citation graph (Semantic Scholar)

Fetch citing papers via WebFetch on the Graph API. Template:

https://api.semanticscholar.org/graph/v1/paper/arXiv:<ID>/citations?fields=title,year,venue,publicationVenue,citationCount,abstract,externalIds,authors&limit=100&offset=0

• Replace arXiv:<ID> with the seed (e.g. arXiv:2212.13345). For a DOI use DOI:<doi>; for a Semantic Scholar ID use it directly.
• Paginate: re-fetch with offset=100, offset=200, … until coverage is enough (3–4 pages of 100 usually covers the substantive work; the long tail is low-quality application papers).
• If it rate-limits (HTTP 429): wait and retry, or reduce fields. A Semantic Scholar API key (env S2_API_KEY, header x-api-key) raises limits if available.
• citationCount is a relative impact signal only — it undercounts very recent venue acceptances. Use it to rank, not to judge.

Use citationCount + year to prioritize which citing papers to read.

3. Targeted web search to fill gaps

The citation graph misses the newest and the venue-specific. Run WebSearch for:

• "<topic>" <venue> <year> (e.g. "Forward-Forward" ICLR 2024)
• "<seed concept>" <sub-topic> (e.g. forward-forward catastrophic forgetting)
• <topic> OpenReview / <topic> site:openreview.net

4. VERIFY everything (the part that matters)

For every paper you intend to report:

• Confirm title + venue + year on a real landing page (arXiv abstract page, OpenReview forum, PMLR proceedings, DBLP, or publisher DOI).
• Never fabricate or guess titles, authors, venues, or IDs.
• Watch for DB artifacts: Semantic Scholar sometimes returns implausible future-dated arXiv IDs or phantom entries. If an ID looks wrong or you can't open it, do not report it in the body — quarantine it in the verification notes.
• Prefer fewer verified papers over many unverified ones.

Track three tiers explicitly: ✅ opened & confirmed · ⚠️ venue per Semantic Scholar but not re-opened (likely correct, flag for a 30-sec DBLP/OpenReview check) · ❌ could not verify / suspect (exclude from body).

5. Parallelize with a background agent (optional but recommended)

For a wide fan-out, launch a background general-purpose agent to gather candidates while you do other work (e.g. reading the seed paper or writing diagrams), then verify its key claims yourself before finalizing — don't trust a sub-agent's citations blindly. Give the agent: the seed ID, the exact Semantic Scholar URL template above, the focus filters, the verification rules, and the required output structure. Spawn with run_in_background: true; you'll be notified on completion.

6. Deliverable

Write a single markdown file. Follow the project's file-naming convention if one exists (e.g. zero-padded number prefix + English kebab-case name like 002-<topic>.md). Structure:

# <Topic> 研究进展调研（侧重 <focus>）

> 种子论文 · 调研方法（S2 引用图谱 N 页 + WebSearch + 逐篇核实）· 引用数来源 · 调研日期

## TL;DR            # 4-6 bullet honest bottom-line, including what is NOT true
## 一、整体格局      # momentum, consensus limitations, where the field's center of gravity is
## 二、<重点维度>    # e.g. a table organized by venue (NeurIPS/ICML/ICLR) if venues were the focus
## 三、<焦点专题>    # deep synthesis of the user's specific angle: what's tried / what works / open problems
## 四、按主题分类的完整列表   # grouped by theme; each: **title** — authors, venue+year, 1-2 sentence what+result, URL, (~cites)
## 五、信源可信度 & 注意事项  # the ✅/⚠️/❌ tiers from step 4 — REQUIRED, never omit
## 六、判断 + 下一步可追什么   # honest take + 2-3 candidate next seed papers to go deeper

Each paper entry must carry: title, authors (first-author et al. ok), venue + year, one or two sentences on what it does and its result, a URL, and citation count if known. Group by theme (e.g. variants/objectives, scaling, architecture, theory, hardware, applications, and the user's focus topic).

7. Honesty rules

• Lead with the bottom line even when it contradicts the framing of the request (e.g. "X is not actually the story; Y is"). Don't pad a thin result.
• Distinguish the strict topic (e.g. the seed method itself) from the broader framing (e.g. the general paradigm) when the evidence splits between them.
• The verification-notes section is mandatory — the user needs to know what to trust.

Pitfalls cheat-sheet

• Google Scholar → blocked; use Semantic Scholar.
• Semantic Scholar 429 → retry / fewer fields / API key.
• Future-dated or odd arXiv IDs from the API → likely phantom; verify or drop.
• Sub-agent output → verify before citing.
• Local PDF won't open → pip install --user pymupdf, read with fitz.