Regularized Score Centering Boosts Agentic RL Performance

A UIUC and CMU paper proposes Regularized Score Centering (RSC) to improve advantage estimation in agentic RL, addressing async RL stability issues, with gains up to 9.9x on BabyAI and 14.3% across six tasks.

2026-11-18 ~ 2026-11-20 · 2 related posts