AutoResearchClaw is a production-ready, open-source, 23-stage autonomous research pipeline that transforms a single idea into a conference-ready paper. It outperforms AI Scientist v2 by 54.7% on ARC-Bench and is the first system to demonstrate that strategic human oversight at decision points — not full autonomy — produces the best research results.
From literature discovery through hypothesis generation, sandboxed experiments, multi-agent review, and LaTeX export
Outperforms AI Scientist v2 on ARC-Bench; 100% pipeline completion rate; 94.3% citation integrity
Co-Pilot mode with SmartPause — only interrupts when AI is uncertain. Beats full auto and micro-management.
Specialized agents for ML, physics, biology, quantum computing, and statistics; MIT license, 12K+ stars
| Phase | Stages | What Happens |
|---|---|---|
| Scoping & Discovery | 1–5 | Problem decomposition, domain detection, literature discovery via OpenAlex/Semantic Scholar/arXiv |
| Hypothesis Generation | 6–9 | Multi-agent debate: Investigator, Innovator, Pragmatist, Contrarian |
| Experimentation | 10–14 | Code generation, sandboxed Docker execution, self-healing repair |
| Analysis | 15–18 | Multi-agent interpretation: Optimist, Skeptic, Methodologist |
| Writing & Export | 19–23 | Paper drafting, 4-layer citation verification, LaTeX export |
v0.5.0: Domain-specialized agents — ColliderAgent (physics), COBRApy (biology), Monte Carlo (statistics), generic Docker (chemistry/materials)
100% — all runs completed all 23 stages
94.3% — 4-layer verification catches most hallucinations
93–96% across ML tasks
44% → 52% result analysis accuracy with human co-pilot
The 8-point quality delta between full-auto and co-pilot is the paper's most consequential finding. Human judgment at critical decision points is the difference between acceptable and failing.
| Mode | Intervention Points | Best For |
|---|---|---|
| Full Auto | Never | Quick exploration, low stakes |
| Gate Only | 3 gate stages | Light oversight |
| Checkpoint | 8 phase endpoints | Phase-level review |
| Co-Pilot | Critical stages + SmartPause | Production quality |
| Step-by-Step | All 23 stages | Learning the pipeline |
| Express | 3 most critical gates | Experienced users |
SmartPause: Confidence-driven dynamic intervention — only pauses when the AI is uncertain. Precision collaboration beats both full autonomy and burnout-inducing micromanagement.
| System | Developer | Key Differentiator | ARC Gap |
|---|---|---|---|
| AutoResearchClaw | AIMING Lab (UNC) | 23-stage + HITL + multi-domain | — |
| AI Scientist v2 | Sakana AI | Agentic tree search, first peer-reviewed AI paper | -54.7% |
| AI-Researcher | HKU DS | Full end-to-end, Scientist-Bench | Less mature |
| n-autoresearch | Karpathy | Multi-GPU parallelism | ML-only, no HITL |
| GPT Researcher | assafelovic | Deep web research, planner agents | Synthesis only |
| ZeroPaper | alejandroll10 | Adversarial verification | Smaller community |
AutoResearchClaw's moat: 23-stage depth × multi-domain breadth × HITL × MetaClaw cross-run learning — unmatched in the current landscape.
Sources 6, 7, 8, 9: Multiple GitHub repos + arXiv:2605.19156Install and test in Co-Pilot mode — Deploy v0.5.0 on sandboxed Docker. Test with a workforce development topic and evaluate output quality.
Conduct a citation audit — Manually verify all references from a trial paper to establish baseline hallucination rate for your LLM backbone.
Develop workforce analytics domain skill — Create a SKILL.md for NCWorks, NCcareers.gov, workforce fluidity metrics, and supply-demand landscape analysis.
Run comparative quality study — Generate papers with varying HITL modes and blind-review with CRTP faculty to quantify quality delta.
Integrate into CRTP methods curriculum — Position as a research methods teaching tool for critiquing AI-generated designs.
Tier guide: Tier 1 = Primary institutional source (official website or announcement) · Tier 2 = Company-owned media or trade publication · Tier 3 = Breaking news / single-source — verify independently before citing
Generated May 24, 2026 · DeepSeek V4 Pro · research-deep v1.0 · Gambit / brief-to-slides v2.0