AutoResearchClaw

AutoResearchClaw is an open-source research automation pipeline from Aiming Lab. Given a research topic, it runs a 23-stage workflow that collects literature, designs and runs experiments, analyzes results, and writes a conference-style paper. The latest listed release is v0.5.0, published 2026-05-20.

What It Is

AutoResearchClaw is a command-line, Python-based system for autonomous scientific paper generation. A user supplies a topic via researchclaw run, and the pipeline moves through eight phases: scoping, literature discovery, knowledge synthesis, experiment design, execution, analysis and decision, paper writing, and finalization. It can run fully autonomously or with human-in-the-loop guidance.

Outputs include a markdown paper draft, conference-template LaTeX (NeurIPS, ICLR, ICML), a BibTeX file, a citation verification report, experiment code with results, generated charts, peer-review notes, and a deliverables folder intended to be compile-ready for Overleaf.

How the Workflow Operates

Literature is gathered from OpenAlex, Semantic Scholar, and arXiv, and citations pass a four-layer verification (arXiv ID, CrossRef/DataCite DOI, title match, LLM relevance scoring). Experiments run in a hardware-aware sandbox that auto-detects GPU, Apple MPS, or CPU, with options for Docker and SSH remote execution. When experiments fail the pipeline attempts self-repair, and a decision stage can proceed, refine, or pivot the research direction. Version 0.5.0 adds domain-specialist execution agents for high-energy physics, biology, and statistics, plus the ARC-Bench benchmark of 55 topics.

Human-in-the-Loop and Integrations

Since v0.4.0 the project offers several intervention modes, from full-auto to step-by-step, including a co-pilot mode with an Idea Workshop, Baseline Navigator, and Paper Co-Writer. CLI commands such as attach, status, approve, reject, and guide let users steer a paused run. It works with OpenClaw, can use ACP-compatible coding agents (Claude Code, Codex CLI, Copilot CLI, Gemini CLI, OpenCode, Kimi CLI) as the LLM backend, and optionally integrates MetaClaw for cross-run learning that converts lessons into reusable skills.

Skills and Extensibility

The project ships 20 built-in skills and loads custom skills from SKILL.md files or configured directories. Configuration is handled through a YAML file covering LLM providers, experiment mode, export target conference, and bridge options.