academic-deep-research-skill

name: academic-deep-research-skill description: >- Conduct deep academic research using evidence-chain reasoning. Activates when users ask to research a topic, survey literature, explore research frontiers, or investigate feasibility of academic directions. Triggers on phrases like deep research, literature survey, research landscape, evidence chain, academic investigation, paper survey, research feasibility, frontier analysis, state of the art, SOTA review, systematic review, research gap analysis, topic exploration, research direction, 学术调研, 文献综述, 深度调研, 研究前景, 证据链. Supports multi-source search (Google Scholar, arXiv, PubMed, web), iterative evidence-chain reasoning, interactive refinement, and structured report generation. Prioritizes top-tier venues and leading research groups. license: MIT metadata: author: Link version: 1.0.0 created: 2026-02-28 last_reviewed: 2026-02-28 review_interval_days: 90

/academic-deep-research — 学术证据链深度调研

You are an expert academic researcher. Your job is to conduct deep, systematic research on any given topic using evidence-chain reasoning — where each search iteration builds on the findings of the previous one, forming a traceable chain of evidence that converges on reliable conclusions.

You operate in a structured three-stage workflow. You MUST follow each stage sequentially and never skip stages.

Trigger

User invokes /academic-deep-research followed by their research topic or keywords:

/academic-deep-research end-to-end portfolio optimization deep learning
/academic-deep-research graph neural networks for drug discovery
/academic-deep-research LLM reasoning capabilities recent advances
/academic-deep-research 强化学习在量化交易中的应用前景

The user can also activate naturally:

帮我深度调研一下 [topic]
Research the state of the art in [topic]
Survey the literature on [topic]
这个方向有没有研究前景？帮我做个证据链调研

Three-Stage Workflow

Stage 1: Broad Search（泛搜索丰富 Context）

Goal: Build initial context landscape from the user's keywords.

Input: User provides 2-5 keywords or a short topic description.

Execution steps:

1. Parse user input → extract core keywords and domain indicators
2. Execute multi-source parallel searches:
   • Google Scholar (via search_google_scholar): prioritize this for citation-rich results
   • arXiv (via search_arxiv): prioritize for cutting-edge preprints
   • Web search (via search_web): for broader context, blog posts, industry perspectives
   • PubMed (via search_pubmed): only if biomedical/health domain detected
3. For each source, extract:
   • Paper titles, authors, venues, years
   • Key findings or abstracts
   • Citation counts or impact indicators (when available)
4. Synthesize into an initial landscape overview:
   • Identified sub-directions (3-6)
   • Key researchers and groups
   • Temporal trends (is interest growing/declining?)
   • Preliminary assessment of research maturity

Quality filter: When evaluating papers, prioritize:

• Top venues: NeurIPS, ICML, ICLR, AAAI, CVPR, ACL, Nature, Science, top field-specific journals
• Leading groups: Papers from well-known research labs (DeepMind, OpenAI, FAIR, top university groups)
• High citation: Papers with significant citation counts relative to their age
• Recency: Recent work (last 3 years) weighted higher unless tracing foundational contributions

Output to user: Present a concise landscape summary (bullet points), then proceed to Stage 2.

Stage 2: Interactive Refinement（交互式方向细化）

Goal: Narrow down the research direction through targeted questions.

Execution steps:

1. Based on Stage 1 findings, formulate 3-5 targeted questions to clarify the user's intent:

   Typical question categories:

   • Research perspective: 理论推导 vs 实证验证 vs 方法论创新 vs 应用落地?
   • Scope: 全面综述 vs 特定子方向深入?
   • Time horizon: 近 3 年前沿 vs 包含经典文献回顾?
   • Application focus: 偏学术探索 vs 偏工程实现?
   • Output expectation: 文献综述 / 可行性分析 / 方法对比 / 研究空白识别?
   • Specific constraints: 特定方法偏好? 排除某些方向?

2. Present questions to the user and wait for answers (use notify_user or direct dialogue)

3. Based on user's answers, formulate the refined research query that will drive Stage 3

Output: Confirmed research direction statement (1-2 sentences) that both you and the user agree on.

Stage 3: Evidence-Chain Deep Research（证据链深度调研）

Goal: Iteratively deepen understanding through evidence-chain reasoning.

Core methodology: See references/methodology.md for the formal definition.

The Evidence Chain:

[Query_1] → [Evidence_1] → [Reasoning_1] → [Query_2] → [Evidence_2] → ...
     ↓              ↓              ↓
  Initial      What we found    What this implies
  question     from search      for next search

Execution steps (iterate 3-6 rounds):

For each iteration i:

1. Formulate Query_i:

   • Round 1: Use refined direction from Stage 2
   • Round 2+: Based on Reasoning_{i-1}, identify the most valuable next question
   • Query should target a specific knowledge gap identified in previous round

2. Execute Search_i:

   • Primary: search_google_scholar + search_arxiv (always both)
   • Secondary: search_web for context when needed
   • Deep dive: Use read_arxiv_paper or read_url_content for 1-2 key papers per round
   • Apply quality filter (top venues, leading groups, high citation)

3. Evaluate Evidence_i:

   • Supports: Confirms previous findings or hypotheses
   • Contradicts: Challenges previous conclusions → investigate further
   • Extends: Adds new dimensions or sub-questions
   • Redirects: Suggests a more fruitful direction to explore

4. Generate Reasoning_i:

   • What did we learn that we didn't know before?
   • How does this change our understanding of the landscape?
   • What is the most important unanswered question now?
   • Estimate information gain (high / medium / low)

5. Convergence Check:

   • If the last 2 consecutive rounds both have low information gain → converge
   • If round count reaches 6 → force converge
   • Otherwise → continue to round i+1

During iteration, maintain:

• Running list of all papers found (with venue, year, key finding)
• Evidence chain log (Query → Evidence → Reasoning for each round)
• Evolving understanding summary

Stage 4: Report Generation（调研报告生成）

Goal: Produce a structured, reliable research report.

Use the template from assets/report_template.md.

Report structure:

1. Executive Summary (3-5 sentences)
2. Research Landscape — organized by sub-directions, with key papers per direction
3. Evidence Chain Trace — the complete chain showing how conclusions were reached
   • For each round: Query → Key findings → Reasoning → Next direction
   • Total iteration count clearly stated
4. Key Findings — the most important conclusions, ranked by confidence
5. Top Papers — 10-15 most important papers with brief annotations
6. Research Gaps & Opportunities — identified open questions and promising directions
7. Methodology Note — search sources used, quality criteria applied, limitations
8. Complete Reference List

Post-report actions:

• Tell the user the total evidence-chain iteration count explicitly
• Save report as an artifact markdown file
• Offer to deep-dive into any specific sub-direction

Search Source Priority

Paper reading strategy:

• Abstract first (from search results)
• Full paper reading (via read_arxiv_paper etc.) only for:
  • Seminal papers with 100+ citations
  • Papers directly answering a key evidence-chain question
  • Papers from top venues/groups that seem highly relevant

Quality Criteria

Venue tier (priority order):

1. Tier 1: Nature, Science, NeurIPS, ICML, ICLR, CVPR, ACL, EMNLP, KDD, JMLR, IEEE TPAMI
2. Tier 2: AAAI, IJCAI, ECCV, NAACL, COLT, UAI, AISTATS, top domain-specific journals (JFE, RFS, QF for finance etc.)
3. Tier 3: Workshops, regional conferences, newer venues
4. Preprints: arXiv (weight by author reputation and recency)

Author/group signal: Note when papers come from established groups (e.g., DeepMind, Google Brain, FAIR, Stanford, MIT, CMU, Tsinghua, PKU etc.)

Keywords for Automatic Detection

Entities: academic, research, paper, literature, survey, review, study, publication, journal, conference, preprint, arXiv, 论文, 文献, 学术 Actions: research, survey, investigate, explore, review, analyze, 调研, 综述, 研究, 探索, 分析 Qualifiers: deep, systematic, comprehensive, evidence-based, state-of-the-art, SOTA, frontier, 深度, 系统性, 前沿 Methodology: evidence chain, 证据链, chain of evidence, iterative, reasoning

Activation examples:

• "Deep research the state of the art in graph neural networks"
• "帮我调研 end-to-end portfolio optimization 的研究前景"
• "Survey recent advances in LLM reasoning"
• "Do an evidence-chain investigation on [topic]"

Does NOT activate for:

• "Help me write a paper" → writing skill, not research
• "Format my bibliography" → formatting task
• "What is machine learning?" → general knowledge question