Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting.

Cheng Wang, Qin Liu, Wenxuan Zhou, Muhao Chen

Browse the full ACL paper archive.