AutoResearch: Turn Research Ideas into Reviewable Evidence

Overview

AutoResearch coordinates AI and machine learning research from idea discovery or a researcher-provided hypothesis through experiment planning, implementation, execution, analysis, and independent review. Its aim is to make the work behind research conclusions inspectable, rather than relying on an agent's unsupported claims.

The project is most relevant to researchers and technical teams willing to configure multiple model roles and manage experiments in their own environment. It records plans, code, logs, results, and reviews so work can be resumed or examined by a human. It is a workflow for producing evidence, not a guarantee that a conclusion is correct or a substitute for researcher judgment.

Key Features

• Discover candidate research directions from recent papers, community discussions, and open-source trends, with optional local domain knowledge.
• Generate and cross-review ideas using at least three distinct models in the Idea Forge stage.
• Run stateful experiments with persisted plans, queues, logs, results, and decisions to support recovery after interruption.
• Use pilot experiments to assess feasibility before scaling up or stopping.
• Preserve negative results and failure causes instead of requiring experiments to report success.
• Add critic reports and blind review to challenge conclusions independently.
• Keep model roles and provider routes in a unified local configuration.

Use Cases

• A research group with no settled question can use idea generation to combine current research signals with its own domain knowledge.
• A researcher with a hypothesis can use the execution workflow to plan, run, analyze, and review an experiment while keeping a durable project record.
• A team testing a computationally expensive idea can run a smaller pilot first and decide whether to revise, scale, or stop.
• Researchers who need to inspect how an agent reached a conclusion can review the saved sources, experiment artifacts, logs, and independent evaluations.

Getting Started

git clone https://github.com/EvoMap/AutoResearch.git
cd AutoResearch
bash scripts/bringup.sh

Configure model credentials and routes before running model-backed stages. See the README for setup details and the provider configuration guide.

Alternatives

• AI-Scientist-v2: AI-Scientist-v2 focuses on open-ended ML research with tree-search agents that generate experiments and draft papers, while AutoResearch emphasizes a traceable, resumable workflow.
• autoresearch: autoresearch iterates on language-model training in a small, single-GPU setup, while AutoResearch supports a broader research workflow and configurable experiment resources.

Considerations

• Model-backed workflows require API credentials and may incur provider charges. Idea generation requires at least three distinct models for its multi-model stage.
• The execution workflow requires Bun, Node.js, an experiment environment manager such as Conda, and the compute, data, and storage resources the experiment needs. GPU use depends on the experiment.
• The README warns that the current execution entrypoint grants Claude Code broad tool permissions. It recommends running it in an isolated, disposable environment without access to unrelated files or credentials.
• Automated reviews and evidence records support human assessment, but do not ensure that findings are correct. External research sources may also be unavailable in some environments.
• The repository describes the execution workflow as Alpha, so users should expect to evaluate its suitability for their own experiments.