Reproduce “Baselines and Score Centering in Policy Gradient Estimation” 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
Shuo Han, Ziqing Hu
The authors analyze reward baselines and score centering in off-policy policy-gradient estimators with regularized importance weights. They compare sequence-level and token-level sampling through reward–score covariance. With a group-mean reward baseline, score centering adds nothing at sequence level; at token level, omitting it generally adds bias under regularized ratios, but not under unregularized ratios. These are estimator results, not training gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.