Reproduce “Playing Dumb: How a Chess Transformer Simulates Users Across Skill Levels” 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
Difan Jiao, George Eilender, Zhenwei Tang, Ashton Anderson
Maia-2 tests whether lower-skill chess play reflects reduced access to concepts or difficulty using them. Linear probes recover 172 board concepts with similar accuracy across skill settings, while targeted policy-head fine-tuning changes move predictions without altering the backbone. On held-out positions for 71 concepts, lowest-band best-move accuracy rises from 14.4% to 82.8%; this applies to selected positions and measured concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.