Preprint
Aug 2026
SoftmaxGRPO: Learning to Reason using Softmax Advantage Group Estimation
Softmax Advantage Group Estimation (SoftmaxGRPO), a drop-in alternative that replaces z-score-normalized group advantages with temperature-scaled softmax advantages, keeping weights bounded regardless of prompt difficulty, is introduced.
Jefferson Hernandez, Jaywon Koo, Zi-Lin Xiao et al.
· 0 citations