Autoresearchclaw Autonomous Research

1---2name: autoresearchclaw-autonomous-research3description: Fully autonomous research pipeline that turns a topic idea into a complete academic paper with real citations, experiments, and conference-ready LaTeX.4triggers:5  - "research this topic automatically"6  - "generate a paper from an idea"7  - "run autonomous research"8  - "use AutoResearchClaw to write a paper"9  - "chat an idea get a paper"10  - "run the research pipeline"11  - "autonomous paper generation"12  - "set up AutoResearchClaw"13---14 15# AutoResearchClaw — Autonomous Research Pipeline16 17> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.18 19AutoResearchClaw is a fully autonomous 23-stage research pipeline that takes a natural language topic and produces a complete academic paper: real arXiv/Semantic Scholar citations, sandboxed experiments, statistical analysis, multi-agent peer review, and conference-ready LaTeX (NeurIPS/ICML/ICLR). No hallucinated references. No human babysitting.20 21---22 23## Installation24 25```bash26# Clone and install27git clone https://github.com/aiming-lab/AutoResearchClaw.git28cd AutoResearchClaw29python3 -m venv .venv && source .venv/bin/activate30pip install -e .31 32# Verify CLI is available33researchclaw --help34```35 36**Requirements:** Python 3.11+37 38---39 40## Configuration41 42```bash43cp config.researchclaw.example.yaml config.arc.yaml44```45 46### Minimum config (`config.arc.yaml`)47 48```yaml49project:50  name: "my-research"51 52research:53  topic: "Your research topic here"54 55llm:56  provider: "openai"57  base_url: "https://api.openai.com/v1"58  api_key_env: "OPENAI_API_KEY"59  primary_model: "gpt-4o"60  fallback_models: ["gpt-4o-mini"]61 62experiment:63  mode: "sandbox"64  sandbox:65    python_path: ".venv/bin/python"66```67 68```bash69export OPENAI_API_KEY="$YOUR_OPENAI_KEY"70```71 72### OpenRouter config (200+ models)73 74```yaml75llm:76  provider: "openrouter"77  api_key_env: "OPENROUTER_API_KEY"78  primary_model: "anthropic/claude-3.5-sonnet"79  fallback_models:80    - "google/gemini-pro-1.5"81    - "meta-llama/llama-3.1-70b-instruct"82```83 84```bash85export OPENROUTER_API_KEY="$YOUR_OPENROUTER_KEY"86```87 88### ACP (Agent Client Protocol) — no API key needed89 90```yaml91llm:92  provider: "acp"93  acp:94    agent: "claude"   # or: codex, gemini, opencode, kimi95    cwd: "."96```97 98The agent CLI (e.g. `claude`) handles its own authentication.99 100### OpenClaw bridge (optional advanced capabilities)101 102```yaml103openclaw_bridge:104  use_cron: true              # Scheduled research runs105  use_message: true           # Progress notifications106  use_memory: true            # Cross-session knowledge persistence107  use_sessions_spawn: true    # Parallel sub-sessions108  use_web_fetch: true         # Live web search in literature review109  use_browser: false          # Browser-based paper collection110```111 112---113 114## Key CLI Commands115 116```bash117# Basic run — fully autonomous, no prompts118researchclaw run --topic "Your research idea" --auto-approve119 120# Run with explicit config file121researchclaw run --config config.arc.yaml --topic "Mixture-of-experts routing efficiency" --auto-approve122 123# Run with topic defined in config (omit --topic flag)124researchclaw run --config config.arc.yaml --auto-approve125 126# Interactive mode — pauses at gate stages for approval127researchclaw run --config config.arc.yaml --topic "Your topic"128 129# Check pipeline status / resume a run130researchclaw status --run-id rc-20260315-120000-abc123131 132# List past runs133researchclaw list134```135 136**Gate stages** (5, 9, 20) pause for human approval in interactive mode. Pass `--auto-approve` to skip all gates.137 138---139 140## Python API141 142```python143from researchclaw.pipeline import Runner144from researchclaw.config import load_config145 146# Load config and run147config = load_config("config.arc.yaml")148config.research.topic = "Efficient attention mechanisms for long-context LLMs"149config.auto_approve = True150 151runner = Runner(config)152result = runner.run()153 154# Access outputs155print(result.artifact_dir)          # artifacts/rc-YYYYMMDD-HHMMSS-<hash>/156print(result.deliverables_dir)      # .../deliverables/157print(result.paper_draft_path)      # .../deliverables/paper_draft.md158print(result.latex_path)            # .../deliverables/paper.tex159print(result.bibtex_path)           # .../deliverables/references.bib160print(result.verification_report)  # .../deliverables/verification_report.json161```162 163```python164# Run specific stages only165from researchclaw.pipeline import Runner, StageRange166 167runner = Runner(config)168result = runner.run(stages=StageRange(start="LITERATURE_COLLECT", end="KNOWLEDGE_EXTRACT"))169```170 171```python172# Access knowledge base after a run173from researchclaw.knowledge import KnowledgeBase174 175kb = KnowledgeBase.load(result.artifact_dir)176findings = kb.get("findings")177literature = kb.get("literature")178decisions = kb.get("decisions")179```180 181---182 183## Output Structure184 185After a run, all outputs land in `artifacts/rc-YYYYMMDD-HHMMSS-<hash>/`:186 187```188artifacts/rc-20260315-120000-abc123/189├── deliverables/190│   ├── paper_draft.md          # Full academic paper (Markdown)191│   ├── paper.tex               # Conference-ready LaTeX192│   ├── references.bib          # Real BibTeX — auto-pruned to inline citations193│   ├── verification_report.json # 4-layer citation integrity report194│   └── reviews.md              # Multi-agent peer review195├── experiment_runs/196│   ├── run_001/197│   │   ├── code/               # Generated experiment code198│   │   ├── results.json        # Structured metrics199│   │   └── sandbox_output.txt  # Execution logs200├── charts/201│   └── *.png                   # Auto-generated comparison charts202├── evolution/203│   └── lessons.json            # Self-learning lessons for future runs204└── knowledge_base/205    ├── decisions.json206    ├── experiments.json207    ├── findings.json208    ├── literature.json209    ├── questions.json210    └── reviews.json211```212 213---214 215## Pipeline Stages Reference216 217| Phase | Stage # | Name | Notes |218|-------|---------|------|-------|219| A | 1 | TOPIC_INIT | Parse and scope research topic |220| A | 2 | PROBLEM_DECOMPOSE | Break into sub-problems |221| B | 3 | SEARCH_STRATEGY | Build search queries |222| B | 4 | LITERATURE_COLLECT | Real API calls to arXiv + Semantic Scholar |223| B | 5 | LITERATURE_SCREEN | **Gate** — approve/reject literature |224| B | 6 | KNOWLEDGE_EXTRACT | Extract structured knowledge |225| C | 7 | SYNTHESIS | Synthesize findings |226| C | 8 | HYPOTHESIS_GEN | Multi-agent debate to form hypotheses |227| D | 9 | EXPERIMENT_DESIGN | **Gate** — approve/reject design |228| D | 10 | CODE_GENERATION | Generate experiment code |229| D | 11 | RESOURCE_PLANNING | GPU/MPS/CPU auto-detection |230| E | 12 | EXPERIMENT_RUN | Sandboxed execution |231| E | 13 | ITERATIVE_REFINE | Self-healing on failure |232| F | 14 | RESULT_ANALYSIS | Multi-agent analysis |233| F | 15 | RESEARCH_DECISION | PROCEED / REFINE / PIVOT |234| G | 16 | PAPER_OUTLINE | Structure paper |235| G | 17 | PAPER_DRAFT | Write full paper |236| G | 18 | PEER_REVIEW | Evidence-consistency check |237| G | 19 | PAPER_REVISION | Incorporate review feedback |238| H | 20 | QUALITY_GATE | **Gate** — final approval |239| H | 21 | KNOWLEDGE_ARCHIVE | Save lessons to KB |240| H | 22 | EXPORT_PUBLISH | Emit LaTeX + BibTeX |241| H | 23 | CITATION_VERIFY | 4-layer anti-hallucination check |242 243---244 245## Common Patterns246 247### Pattern: Quick paper on a topic248 249```bash250export OPENAI_API_KEY="$OPENAI_API_KEY"251researchclaw run \252  --topic "Self-supervised learning for protein structure prediction" \253  --auto-approve254```255 256### Pattern: Reproducible run with full config257 258```yaml259# config.arc.yaml260project:261  name: "protein-ssl-research"262 263research:264  topic: "Self-supervised learning for protein structure prediction"265 266llm:267  provider: "openai"268  api_key_env: "OPENAI_API_KEY"269  primary_model: "gpt-4o"270  fallback_models: ["gpt-4o-mini"]271 272experiment:273  mode: "sandbox"274  sandbox:275    python_path: ".venv/bin/python"276  max_iterations: 3277  timeout_seconds: 300278```279 280```bash281researchclaw run --config config.arc.yaml --auto-approve282```283 284### Pattern: Use Claude via OpenRouter for best reasoning285 286```bash287export OPENROUTER_API_KEY="$OPENROUTER_API_KEY"288 289cat > config.arc.yaml << 'EOF'290project:291  name: "my-research"292llm:293  provider: "openrouter"294  api_key_env: "OPENROUTER_API_KEY"295  primary_model: "anthropic/claude-3.5-sonnet"296  fallback_models: ["google/gemini-pro-1.5"]297experiment:298  mode: "sandbox"299  sandbox:300    python_path: ".venv/bin/python"301EOF302 303researchclaw run --config config.arc.yaml \304  --topic "Efficient KV cache compression for transformer inference" \305  --auto-approve306```307 308### Pattern: Resume after a failed run309 310```bash311# List runs to find the run ID312researchclaw list313 314# Resume from last completed stage315researchclaw run --resume rc-20260315-120000-abc123316```317 318### Pattern: Programmatic batch research319 320```python321import asyncio322from researchclaw.pipeline import Runner323from researchclaw.config import load_config324 325topics = [326    "LoRA fine-tuning on limited hardware",327    "Speculative decoding for LLM inference",328    "Flash attention variants comparison",329]330 331config = load_config("config.arc.yaml")332config.auto_approve = True333 334for topic in topics:335    config.research.topic = topic336    runner = Runner(config)337    result = runner.run()338    print(f"[{topic}] → {result.deliverables_dir}")339```340 341### Pattern: OpenClaw one-liner (if using OpenClaw agent)342 343```344Share the repo URL with OpenClaw, then say:345"Research mixture-of-experts routing efficiency"346```347 348OpenClaw auto-reads `RESEARCHCLAW_AGENTS.md`, clones, installs, configures, and runs the full pipeline.349 350---351 352## Compile the LaTeX Output353 354```bash355# Navigate to deliverables356cd artifacts/rc-*/deliverables/357 358# Compile (requires a LaTeX distribution)359pdflatex paper.tex360bibtex paper361pdflatex paper.tex362pdflatex paper.tex363 364# Or upload paper.tex + references.bib directly to Overleaf365```366 367---368 369## Troubleshooting370 371### `researchclaw: command not found`372```bash373# Make sure the venv is active and package is installed374source .venv/bin/activate375pip install -e .376which researchclaw377```378 379### API key errors380```bash381# Verify env var is set382echo $OPENAI_API_KEY383# Should print your key (not empty)384 385# Set it explicitly for the session386export OPENAI_API_KEY="sk-..."387```388 389### Experiment sandbox failures390The pipeline self-heals at Stage 13 (ITERATIVE_REFINE). If it keeps failing:391```yaml392# Increase timeout and iterations in config393experiment:394  max_iterations: 5395  timeout_seconds: 600396  sandbox:397    python_path: ".venv/bin/python"398```399 400### Citation hallucination warnings401Stage 23 (CITATION_VERIFY) runs a 4-layer check. If references are pruned:402- This is **expected behaviour** — fake citations are removed automatically403- Check `verification_report.json` for details on which citations were rejected and why404 405### PIVOT loop running indefinitely406Stage 15 (RESEARCH_DECISION) may pivot multiple times. To cap iterations:407```yaml408research:409  max_pivots: 2410  max_refines: 3411```412 413### LaTeX compilation errors414```bash415# Check for missing packages416pdflatex paper.tex 2>&1 | grep "File.*not found"417 418# Install missing packages (TeX Live)419tlmgr install <package-name>420```421 422### Out of memory during experiments423```yaml424# Force CPU mode in config425experiment:426  sandbox:427    device: "cpu"428    max_memory_gb: 4429```430 431---432 433## Key Concepts434 435- **PIVOT/REFINE Loop**: Stage 15 autonomously decides PROCEED, REFINE (tweak params), or PIVOT (new hypothesis direction). All artifacts are versioned.436- **Multi-Agent Debate**: Stages 8, 14, 18 use structured multi-perspective debate — not a single LLM pass.437- **Self-Learning**: Each run extracts lessons with 30-day time decay. Future runs on similar topics benefit from past mistakes.438- **Sentinel Watchdog**: Background monitor detects NaN/Inf in results, checks paper-evidence consistency, scores citation relevance, and guards against fabrication throughout the run.439- **4-Layer Citation Verification**: arXiv lookup → CrossRef lookup → DataCite lookup → LLM relevance scoring. A citation must pass all layers to survive.