Reproduce “Whitened Advantage Summation Policy Optimization: Correlation-Aware Aggregation for Multi-Reward Reinforcement Learning” 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
Olaitan Olaleye, Jayson Meribe, Temidayo Olaleye
WAS-PO estimates correlations among normalized rewards, whitens them, and combines them using intended weights. It also replaces GDPO’s per-prompt normalization with batch-level normalization, so the experiments do not isolate whitening. On a three-reward tool-calling task, it achieved higher mean training AUC than GDPO across seven seeds, but held-out accuracy was not higher, and the gain cannot be attributed to whitening alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.