Reproduce “Optimal Design for Reward Modeling in RLHF” 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
Antoine Scheid, Etienne Boursier, Alain Durmus, Michael I. Jordan, Pierre Ménard, Eric Moulines, Michal Valko
ODPO selects preference comparisons from a pool of prompt-completion candidates before labels arrive. It uses bounded embedding differences and an optimal-design criterion to choose pairs that cover informative directions, then fits a reward model from Bradley–Terry preferences. Under linear rewards in a known embedding space, its worst-case simple regret scales as d/√T, matching a lower bound up to logarithmic factors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.