The AI Scientist-v2
The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree SearchAn end-to-end agent system that forms ML research hypotheses, runs experiments with agentic tree search, analyzes the data and writes full papers, without the human-written code templates v1 relied on.
LLM agents propose machine-learning research ideas, write and run the experiment code under an experiment-manager agent using progressive tree search, and write up the results, with a vision-language model critiquing the figures. Its output is new knowledge about training and evaluating models; feeding such findings back into how models (including its own) are built is the step not yet shown.
- LLM agents propose ML research ideas
- Agents run experiments with tree search
- Agents write up results and papers
- Output is new knowledge about models
- LLM agents propose ML research ideas
- Agents run experiments with tree search
- Agents write up results and papers
- Output is new knowledge about models
Why it is a road to recursion
An agent that can carry out and write up ML research end to end is the component that, pointed at its own training and scaffolding, would make AI research recursive.
Evidence
Of three fully AI-generated papers submitted to an ICLR 2025 workshop (ICBINB), one received reviewer scores of 6, 7 and 6 (average 6.33), above the average human acceptance threshold, and was withdrawn before publication as planned.
Related loops
More research automation →- Coding agent edits LLM training script
- Runs a 5-minute training job
- Keeps change if validation improves
- Changes transferred to larger models
- repeat
- Qwen3-14B edits GPT training script
- Edit is trained and scored
- Score rewards Qwen3-14B weights update
- Trained checkpoint re-run in autoresearch loop
- repeat
- LLM agents propose hypotheses on AI tasks
- Agents implement and test them
- Findings Memory steers later proposals
- Validated finding directly improves AI system
- repeat