AI Scientist v2 Generates ICLR Workshop Paper Reviewed and Accepted

The article reports that a paper fully generated by Sakana AI’s AI Scientist‑v2 system was submitted to an ICLR workshop, received scores of 6/7/6, and surpassed the average human acceptance threshold.

The AI Scientist‑v2 is an end‑to‑end agentic system that autonomously formulates scientific hypotheses, designs and runs experiments, analyzes results, creates visualizations, and writes the complete manuscript, including titles, references, and figure placement.

The system employs a progressive agentic tree search guided by an experiment‑management agent, allowing it to explore multiple research hypotheses and select the most promising code checkpoints for further testing.

Three AI‑generated papers were submitted; two fell short of acceptance, while the third achieved an average score of 6.33, ranking around the 45th percentile among all submissions, above the typical human acceptance threshold.

Although the paper was later withdrawn from the public OpenReview forum for transparency reasons, the code and data remain available in the project's GitHub repository.

The article also notes occasional citation errors made by the AI, such as misattributing LSTM work to Goodfellow (2016) instead of Hochreiter and Schmidhuber (1997).

Behind the technology is Sakana AI, a startup founded by Llion Jones, one of the original Transformer authors, and the project involves collaborators from UBC and Oxford, including researchers Cong Lu, Jeff Clune, Chris Lu, and Jakob Foerster.

The authors emphasize that the ultimate goal is not to compare AI‑generated science with human science, but to produce discoveries that benefit humanity, potentially leading to publications in top journals like Nature or Science.