paper-research-agent（论文研究 Agent）

Paper Research Agent - Autonomous Multi-Agent Research System

When to Use

Use this skill when the user wants to:

• Research papers on a specific topic
• Find related literature for a research area
• Analyze academic papers in depth
• Build a literature survey
• Identify research gaps
• Compare methods across papers

Core Workflow

The system autonomously executes the full research pipeline:

User Query → Research Probe → PDF Download → Parallel Agent Analysis → Integrated Report

Phase 1: Research Probe (Automated)

• Parse user's research intent from natural language
• Execute vertical deep search or iterative exploration
• Generate research graph with papers at different levels

Phase 2: PDF Download (Automated)

• Download PDFs from arxiv
• Deduplicate and version management
• Standard naming: {paper_title}-{arxiv_id}.pdf

Phase 3: Parallel Agent Analysis (Automated - Key)

• Spawn multiple sub-agents (one per paper)
• Each agent reads full PDF using paper-reader
• Generate 6-section detailed analysis
• Agents run in parallel for speed

Phase 4: Report Integration (Automated)

• Collect all agent analyses
• Generate comparison tables
• Identify research gaps
• Output comprehensive survey

Agent Analysis Requirements

Each sub-agent MUST generate a 6-section report following the detailed standards in: references/analysis_standards.md

SubAgent MUST read this reference file before starting analysis to understand:

• Detailed requirements for each of the 6 sections
• Possible sub-sections to consider (as hints, not rigid requirements)
• Quality checklists
• How to use paper-reader tool
• Report format template

Summary of 6 Required Sections

Section 1: Research Background

• Domain context and research lineage
• Key prior works cited (3-5 papers)
• Technical state when this paper was published
• Goal: Help user understand the research landscape

Section 2: Research Problem

• Specific problem being solved
• Limitations of existing methods (cite original text)
• Core assumptions and insights
• Goal: Clarify what problem the author identified

Section 3: Core Innovation

• Detailed method/system architecture
• Technical details (network structure, dimensions)
• Key formulas in LaTeX format
• Comparison table with prior methods
• Goal: Understand exactly what the author did

Section 4: Experimental Design

• Dataset details (name, scale, characteristics)
• Baseline methods used
• Evaluation metrics
• REAL experimental data tables (copy from paper)
• Ablation study results
• Goal: Extract real data for comparison

Section 5: Key Insights

• Core findings from experiments
• Domain insights (what works/doesn't work)
• Practical recommendations
• Goal: Learn actionable lessons

Section 6: Future Work

• Limitations acknowledged by authors
• Unsolved problems
• Potential research directions (at least 3)
• Goal: Identify research gaps for user's innovation

For full details, sub-section hints, and quality standards - READ references/analysis_standards.md

Quality Enforcement

Agents MUST:

• ✅ Read EVERY section of the PDF (not just abstract)
• ✅ Extract REAL tables with actual data
• ✅ Cite sources with exact locations [Section X.Y]
• ✅ Use paper-reader tool for extraction
• ❌ NEVER fabricate data
• ❌ NEVER skip sections

Usage

Agent Execution (When User Requests Research)

Trigger phrases:

• "帮我调研一下XXX的相关论文"
• "Research papers on X"
• "Find related literature about X"
• "分析XXX领域的论文"

Agent Action:

Step 1: Execute main pipeline

import subprocess
result = subprocess.run([
    "python3", 
    "~/.openclaw/workspace/skills/paper-research-agent/scripts/research_pipeline.py",
    "--query", "{user_topic}",
    "--mode", "vertical",
    "--max-papers", "10",
    "--output", "./research_{topic}"
], capture_output=True, text=True)

print(result.stdout)

Step 2: Read generated agent tasks

import json
with open("./research_{topic}/_agent_tasks.json") as f:
    tasks = json.load(f)

Step 3: Spawn parallel sub-agents for analysis (CRITICAL)

# Spawn multiple agents in parallel for each paper
for task_info in tasks:
    sessions_spawn(
        agentId="main",
        mode="run", 
        runtime="subagent",
        task=task_info['task'],
        timeoutSeconds=600  # 10 minutes per paper
    )

Important: Launch as many agents in parallel as possible for speed.

Step 4: After all agents complete, integrate results

# Collect all analysis reports
# Generate integrated survey
# Present to user

Output Structure

research_output/
├── _research_summary.json              # Research metadata
├── probe/
│   ├── _probe_results.json            # Search results
│   └── _probe_report.md               # Human-readable probe report
├── papers/
│   ├── {title}-{arxiv_id}.pdf         # Downloaded PDFs
│   └── ...
├── analysis/
│   ├── {title}-{arxiv_id}_analysis.md # 6-section agent reports
│   └── ...
└── _integrated_survey.md              # Final integrated survey

Key Scripts

• scripts/research_pipeline.py: Main orchestration script
• scripts/research_probe.py: Intelligent search module
• scripts/paper_downloader.py: PDF download module
• scripts/agent_task_generator.py: Sub-agent task generator

Report Format Standards

Each sub-agent analysis report MUST follow this exact 6-section structure:

# 📄 {Paper Title}

> **ArXiv ID**: {id}  
> **Authors**: {authors}  
> **Published**: {date}

---

## Section 1: Research Background
- Domain context
- Key prior works (3-5 papers with citations)
- Technical state at publication time
- Citations: [Section X.Y]

## Section 2: Research Problem
- SPECIFIC problem being solved
- SPECIFIC limitations of existing methods (quote original)
- Core assumptions
- Citations: [Section X.Y, "exact quote"]

## Section 3: Core Innovation
- Method/system architecture (detailed)
- Technical details (network structure, dimensions)
- Key formulas in LaTeX: $...$
- Comparison table:
  | Aspect | Prior Work | This Paper | Advantage |
  |--------|-----------|------------|-----------|
- What is genuinely new

## Section 4: Experimental Design
- Dataset: Name, size, characteristics
- Baseline methods: Specific names
- Metrics: Formulas, units
- Results table (REAL data):
  | Method | Metric1 | Metric2 |
  |--------|---------|---------|
  | This | X.XX | X.XX |
  | Baseline | X.XX | X.XX |
- Ablation study results

## Section 5: Key Insights
- Core findings from experiments
- What works/doesn't work
- Design choices and impact
- Practical recommendations

## Section 6: Future Work
- Limitations acknowledged by authors
- Unsolved problems
- Future directions (3+)

---

*Analysis by Paper Research Agent*  
*Date: {timestamp}*

Quality Requirements:

• Minimum 3000 words
• At least 3 data tables
• At least 10 citations to original text
• All citations must include exact location [Section X.Y] or [Table N]
• No fabricated data - all numbers must come from the actual paper

Error Handling

If paper download fails:

• Skip and continue with available papers
• Log error in summary

If agent analysis fails:

• Retry once
• If still failing, mark as "analysis_failed" in summary
• Continue with other papers

Best Practices

1. For deep research: Use --mode vertical (searches 4 levels deep)
2. For exploration: Use --mode iterative (progressive discovery)
3. For specific paper: Use --mode horizontal (find related work)
4. Parallel agents: System auto-spawns optimal number based on paper count
5. Quality check: Always verify a few random citations manually

Example Session

User: "帮我调研扩散策略在机器人操作中的应用"

Agent:

1. Executes research probe with query "扩散策略 机器人操作"
2. Finds 30 related papers across 4 levels
3. Downloads PDFs for top 10 papers
4. Spawns 10 sub-agents in parallel
5. Each agent analyzes one paper with 6-section format
6. Collects all analyses
7. Generates integrated survey with comparison tables
8. Presents final report to user

Output: Complete research package with all papers analyzed and integrated survey.

Dependencies

Required Python packages (auto-installed):

• arxiv
• requests
• pdfplumber (for paper-reader)

Notes

• Each paper analysis takes 5-10 minutes
• Parallel execution significantly speeds up research
• Always verify critical data points manually
• The system respects arxiv rate limits (3s delay between downloads)