GDPO
A policy optimization method for multi-reward reinforcement learning that employs per-prompt normalization, serving as a comparison baseline for WAS-PO.
Timeline (last 90 days)
WAS-PO replaces GDPO's per-prompt normalization with batch-level normalization, but this change and the whitening operation were not separated in the experiments.
Under three rewards, WAS-PO's held-out accuracy is slightly lower than GDPO's (not significant); under four rewards it is absolutely lower by 0.03 (only two seeds were evaluated).
Under the two-reward setting, WAS-PO's training reward is lower than GDPO's.
All Facts (3)
WAS-PO replaces GDPO's per-prompt normalization with batch-level normalization, but this change and the whitening operation were not separated in the experiments.
50%UnverifiedUnder three rewards, WAS-PO's held-out accuracy is slightly lower than GDPO's (not significant); under four rewards it is absolutely lower by 0.03 (only two seeds were evaluated).
50%UnverifiedUnder the two-reward setting, WAS-PO's training reward is lower than GDPO's.
50%