Reproduce “Decoupling Exploration from Optimization in RLVR” 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
Saif Punjwani, Micah Goldblum
ExpDis separates exploratory math reasoning from training the student. A novelty-rewarded explorer generates solutions, and the system transfers filtered, correct trajectories for student training with correctness rewards. On three small models, ExpDis outperforms DAPO and DAPO trained for four times as many steps on math measures. These results concern math tasks and do not guarantee preservation of other abilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.