Meta paper: GRPO post-training unevenly boosts subgroups, proposes fix
A Meta paper shows GRPO-style RL post-training unevenly boosts capabilities across demographic subgroups, amplifying counterfactual bias, and proposes a group-level reward transformation exploiting GRPO's multi-rollout structure to correct it.
2026-10-20 ~ 2026-10-20 · 3 related posts
- RL alignment amplifies counterfactual bias, group-level reward fix shows new paper — nagpalchirag · 2026-10-20
- Meta paper quantifies how GRPO post-training gains skew across demographics — nagpalchirag · 2026-10-20
- Meta proposes reward transformation exploiting GRPO rollouts to equalize subgroups — nagpalchirag · 2026-10-20