English Β· νκ΅μ΄
π 1st Place at Ralphthon @ICML "Auto Research" β won $10,000 OpenAI credits
Co-authored by Byungjun Yoon, Woomin Song & our AI Scientist π
π¬ Fully AI-written, with only light high-level human guidance, then judged by human experts. The AI scientist wrote the paper end to end; humans stepped in only for occasional high-level direction, never to revise the text. It was then evaluated in ICML format by a panel of 11 expert judges: professors and researchers across ML, AI safety, robotics, and biotech.
So we're fully open-sourcing the git log β every step of how our AI scientist reasoned, revised, and drove Overleaf to write the LaTeX.
# track whole AI Scientist's writing history
git clone https://github.com/happyhappy-jun/depth-ar.git
git logThe AI kept setting itself hard rules and quoting them back β banning its own weasel-words, refusing to ship a figure it hadn't looked at, and publishing a result that argued against its own paper. Then it got bitten by the exact bug one of those rules was written to prevent. All verbatim from the commit messages:
Depth-AR β instead of zeroing a skipped Transformer layer's update, predict it from the layers before it. π paper.pdf
This repo is the paper β the harness that wrote it lives in happyhappy-jun/writing-driven-autoresearch: the three agent personas, the decision ledger (136 timestamped decisions), and the integrity tooling that kept the paper honest. The full story of how and why we built it: the blog post.
Built at Ralphthon @ICML β "Auto Research": once the "Ralph Loop" starts, you're not allowed to touch your agent. (To touch your laptop at all, you wear a lobster costume β which is why every commit is authored by lobster. π¦)
How it was judged. In the AI Scientist track, each team builds an agent that writes a research paper (this repo). The paper was peer-reviewed in ICML format by a panel of 11 expert human judges β university professors and researchers, startup founders and engineers - who scored both the paper and the agent workflow behind it. We took 1st place.