Skip to content

Sakana AI Scientist (v1)

external · status: dormant · focus: end-to-end · discipline: computer-science · started: 2024

Project page: https://github.com/SakanaAI/AI-Scientist

Licence: unclassified (licence file present; see repository)

Source: projects/landscape/sakana-ai-scientist-v1.yml

Positioning

The original AI Scientist release (arXiv:2408.06292): an end-to-end agentic pipeline that ideates, runs experiments, and writes a paper with self-review on a fixed set of CS templates (NanoGPT, 2D Diffusion, Grokking). Foundational reference for the AI-scientist line and direct predecessor to v2.

Distinctive contribution

First widely-noticed system to demonstrate a complete machine-learning paper produced autonomously by an LLM pipeline, including in-pipeline reviewer simulation. Released with example papers and per-template scaffolds, making the line of work concrete and inspectable.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Seven stages from ideation through self-review; no literature-synthesis or external dissemination stage.
Autonomy level 3 Runs end-to-end without per-step human approval; user supplies a template and topic.
Architectural transparency 3 Code, prompts, templates, and example outputs publicly released; arXiv paper documents the design.
Inputs supported 1 Narrow input form (area + curated template); limited external data integration.
Outputs / reproducibility 2 Papers + code + logs persisted; reproducibility tied to the seeded template.
Internal evaluation 3 External validation: the underlying research (Lu, C. et al.) was peer-reviewed and published in Nature 651, 914-919 (26 Mar 2026), reporting that an AI Scientist-v2-generated manuscript passed workshop-level human peer review (ICLR 2025 ICBINB, avg. score 6.33).
Openness 1 Relicensed Dec 2025 from permissive terms to the custom 'AI Scientist Source Code License v1.0' (based on the Responsible AI Source Code License) — a non-permissive, restriction-bearing license (mandatory AI-disclosure clause, use restrictions, termination-on-violation); requires Linux + NVIDIA GPU.
Maturity / traction 3 14.5k+ stars; peer attention; basis for the v2 release; capstone research published in Nature (Mar 2026). Repo itself quiet since 2025-12-19 (~11 months, no tagged releases); Sakana folded this lineage into a new internal 'RSI Lab' (recursive self-improvement, reported to involve Jürgen Schmidhuber, 2026-09) rather than continuing standalone commits — reflected in flipping status to dormant.
Cross-family policy 0 Same as v2 — self-refinement within one model family.
Runtime assurance 1 Self-review pass + per-template scaffold checks.
Cross-platform portability 0 Linux + NVIDIA GPU + texlive locked.

Scored on 2026-11-01. See the evaluation rubric.

Tags

Pipeline stages: hypothesis-generation research-design data-analysis code-generation paper-drafting revision-editing referee-simulation

Architectural features: multi-agent tool-use iterative-loop artifact-versioning

Inputs: research-area template

Outputs: paper-pdf code experiment-logs llm-generated-review

Data sources: template-datasets

Knowledge sources: semantic-scholar

Limitations

  • Locked to three CS templates; portability to other domains community-maintained.
  • Self-review correlates weakly with external peer-review judgments.
  • Non-permissive license (AI Scientist Source Code License v1.0, Dec 2025); requires Linux + NVIDIA GPU + texlive-full.
  • Executes LLM-written code; security/containment is the user's responsibility.
  • An ICLR 2025 ICBINB workshop paper generated by this lineage (AI Scientist-v2) was passed peer review in March 2025 and was then withdrawn by Sakana under an agreement made with the workshop organisers before submission; it is separate from the peer-reviewed Nature paper.
  • Repo dormant since 2025-12-19; the lineage continued as a separate project under a new name — see SakanaAI/AI-Scientist-v3 ('The Agentic Scientist', released 2026-10-15, cataloged as sakana-ai-scientist-v3) — rather than as a commit history on this repo.

Papers describing this project

  • The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery — Lu, C., Lu, C., Lange, R. T., Foerster, J., Clune, J., Ha, D. (2024). arXiv. arXiv:2408.06292
  • Towards end-to-end automation of AI research — Lu, C., Lange, R. T., Yamada, Y., Hu, S., Foerster, J., Ha, D., Clune, J. (2026). Nature. arXiv:2606.15497 · doi

Also compared in

  • ARIS Table 4 (yang2026aris) — Cited as exhibiting 'recurring limitations' of same-model self-refinement; scored: no cross-family, partial adversarial review, no composable skills, ✓ E2E, partial assurance, no portability.
  • A Survey of AI Scientists (tie2025aiscientistsurvey) — Covered as a foundational AI-scientist system.