🏆 1st Place at Ralphthon @ICML "Auto Research"

July 14, 2026 · View on GitHub

English · 한국어

🏆 1st Place at Ralphthon @ICML "Auto Research" — won $10,000 OpenAI credits

Depth-AR — a research paper written end-to-end by an AI scientist

Co-authored by Byungjun Yoon, Woomin Song & our AI Scientist 🙏

🔬 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.

✍️ We believe there's far more to learn from how an agent writes than from the paper itself.

So we're fully open-sourcing the git log — every step of how our AI scientist reasoned, revised, and drove Overleaf to write the LaTeX.

Full `git log` of Our AI-Scientist

# track whole AI Scientist's writing history
git clone https://github.com/happyhappy-jun/depth-ar.git
git log

Commit-history timeline: 45 commits in one ~4-hour session, 39 by the AI, with self-corrections highlighted — a pre-registered test, an honest non-replication, a figure fix flagged not faked, and a retraction.

Left alone, it wrote itself a rulebook

The AI kept setting itself hard rules and quoting them back — 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:

Rules the AI wrote for itself, verbatim from its commits: a number I cannot account for is a number I do not print; a figure I have not looked at does not go in the paper; a warn is a non-check; 'matches or beats' is a banned construction; and it goes in anyway.

The paper, in one line

Depth-AR — instead of zeroing a skipped Transformer layer's update, predict it from the layers before it. 📄 paper.pdf

How it was built

This repo is the paper — 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.

The event

Built at Ralphthon @ICML — "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 — which is why every commit is authored by lobster. 🦞)

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 — university professors and researchers, startup founders and engineers - who scored both the paper and the agent workflow behind it. We took 1st place.