← Reference notes

License: CC BY 4.0. Source: https://arxiv.org/abs/2408.06292

Summary

Chris Lu, Cong Lu, Lange, Foerster, Clune, and Ha describe a pipeline that asks a frontier model to carry a small machine-learning study from idea to writeup. The system proposes an idea, writes code, runs experiments, plots results, drafts a paper, and then scores that paper with an automated reviewer. They try it on diffusion modeling, transformer language modeling, and learning dynamics, and they say a full paper from one idea cost under 15 dollars. The automated reviewer, on their check, scored papers about as people do, and some generated papers cleared a conference-style bar according to that reviewer, not according to a human program committee. They suggest the loop can repeat. Code is linked from the paper. This note follows the abstract page only.

Why it is here

The auto-research note and the IdeaScientist reference already name "The AI Scientist" as a related system. This is that paper, on a license the library can keep. It shows an end-to-end agent loop for a narrow kind of ML paper. It does not show that the papers are correct, or that the loop compounds into stronger science.

@misc{iil_note_fpm72eee,
  author = {{iinventedlight Library}},
  title = {{The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-10-18},
  url = {https://iinventedlight.com/library/reference/arxiv/2408-06292-the-ai-scientist/},
  note = {Permanent ID iil:note:fpm72eee. License: CC-BY-4.0. Summary of https://arxiv.org/abs/2408.06292.},
}

For language models and agents: start at /llms.txt. This page in Markdown: /library/reference/arxiv/2408-06292-the-ai-scientist.md.