Curiosity · AI Model
AI Scientist v2
The AI Scientist v2 is Sakana AI's 2025 autonomous research agent — a framework that uses frontier LLMs to generate machine-learning research ideas, implement experiments in code, run them, analyse results, and draft full LaTeX papers without human intervention. It made headlines when a v2-generated paper was accepted at an ICLR 2025 workshop, marking a milestone in automated science.
Model specs
- Vendor
- Sakana AI
- Family
- AI Scientist
- Released
- 2025-04
- Context window
- 200,000 tokens
- Modalities
- text, code
Strengths
- End-to-end pipeline from idea to PDF paper
- Uses frontier LLMs (Claude/GPT) as reasoning backbone
- Code and framework released openly on GitHub
Limitations
- Best on small-scale ML experiments; weak on novel paradigms
- Quality of output papers varies widely; peer-review selection is still rare
- Can 'hallucinate' citations — human review essential before publication
Use cases
- Automated machine-learning ablation studies
- Research on LLM-driven scientific discovery
- Generating baseline experiments for human researchers
- Teaching agentic workflows in graduate ML courses
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| ICLR 2025 workshop acceptance | 1 paper | 2026-04 |
| Experiments executed per run | ~10+ | 2026-04 |
| Cost per paper (USD) | ~$15 | 2026-04 |
Frequently asked questions
What is the AI Scientist v2?
The AI Scientist v2 is Sakana AI's open-source autonomous research agent framework. It orchestrates an LLM to propose research ideas, write experiment code, run it, analyse results, and draft full LaTeX papers.
Did the AI Scientist actually publish a peer-reviewed paper?
Yes — in 2025, a paper authored end-to-end by AI Scientist v2 was accepted at an ICLR 2025 workshop (with peer-reviewer awareness and controlled conditions). It remains an early, carefully-scoped milestone rather than a routine outcome.
Sources
- Sakana AI — AI Scientist v2 blog — accessed 2026-04-20
- AI Scientist GitHub — accessed 2026-04-20