Skip to content

Sakana AI Scientist v2

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

Project page: https://github.com/SakanaAI/AI-Scientist-v2

Licence: unclassified (licence file present; see repository)

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

Positioning

An autonomous "AI scientist" pipeline that ideates, runs experiments (primarily ML), drafts a paper, and self-reviews. Targets the full RISE arc end-to-end with minimal human oversight per task.

Distinctive contribution

Among the earliest publicly released systems to demonstrate a single agentic pipeline producing complete machine-learning papers (idea → experiments → write-up → self-review) without per-step human intervention.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Covers ~7 stages; weak on literature discovery/synthesis and external dissemination.
Autonomy level 3 Runs end-to-end without per-task human approval.
Architectural transparency 3 Code, prompts, and configurations publicly released.
Inputs supported 1 Narrow input form (research area + template); limited external data integration.
Outputs / reproducibility 2 Persists papers + code + experiment logs; reproducibility depends on the seeded template.
Internal evaluation 3 External validation: a v2-generated manuscript passed workshop-level human peer review (ICLR 2025 ICBINB, avg. score 6.33); the underlying research was subsequently peer-reviewed and published in Nature 651, 914-919 (26 Mar 2026).
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, confirmed in this repo's LICENSE) — a non-permissive, restriction-bearing license (mandatory AI-disclosure clause, use restrictions, termination-on-violation).
Maturity / traction 3 7.1k+ stars; multiple iterations (v1 → v2); capstone research published in Nature (Mar 2026), satisfying the 'citations'/production-adoption band. 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-model self-refinement is the canonical failure mode ARIS Table 4 identifies in this lineage.
Runtime assurance 1 Agentic tree search + self-review pass provide light runtime gating.
Cross-platform portability 0 Linux + NVIDIA GPU locked; single-template runtime.

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 seed-template

Outputs: paper-draft code experiment-logs self-review

Data sources: benchmark-datasets

Knowledge sources: semantic-scholar openreview

Limitations

  • ML-paper focus; portability to non-CS domains unclear.
  • Quality of produced papers strongly dependent on seed templates.
  • Self-review is not a substitute for external peer review.
  • Non-permissive license (AI Scientist Source Code License v1.0, Dec 2025), unlike the permissive license this project shipped under previously.
  • An ICLR 2025 ICBINB workshop paper generated by this system 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-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search — Yamada, Y., Lange, R. T., Lu, C., Hu, S., Lu, C., Foerster, J., et al. (2025). arXiv. arXiv:2504.08066
  • 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) — Same scoring profile as v1 in Table 4; noted as workshop-level paper produced via agentic tree search.
  • A Survey of AI Scientists (tie2025aiscientistsurvey) — Covered as the v2 release of the foundational AI-scientist line.