#artificial intelligence
Feb 2026
F-GRPO: Don't Let Your Policy Learn the Obvious and Forget the Rare
A difficulty-aware scaling coefficient is proposed, inspired by Focal loss, that down-weights updates on high-success sampled groups and improves average math pass@256 in all three cases, without increasing group size or computational cost.
Daniil Plyusov, Alexey Gorbatovski, B. Shaposhnikov et al.
· arXiv.org · 9 citations
· ⚡1