Reproduce “Do LLM Agents Still Need to Be Told How to Self-Improve?” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shobhita Sundaram, Cansu Sancaktar, Phillip Isola
The authors test whether language models can design their own learning procedures for puzzle play, using a frozen teacher to create games, observe a player, and revise its text context. Newer models improve on held-out PuzzleScript-t20, with Claude Opus 5 gaining +20%, and its context transfers to longer PuzzleScript, ARC-AGI-3, and MazeBench. The evidence is limited to puzzle settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.