MASPO Improves RL Rewards Optimization For LLMs

Jiaye Lin (arXiv v1, Feb 19, 2026) proposes Mass-Adaptive Soft Policy Optimization (MASPO), a unified RL with verifiable rewards (RLVR) framework addressing three limitations in methods like GRPO. MASPO integrates a differentiable soft Gaussian gating, a mass-adaptive limiter, and an asymmetric risk controller, and the paper reports MASPO significantly outperforms strong baselines; code is available.