一作ACL 2026Findings· Findings of ACL
Improved Policy Optimization for Mixture-of-Experts Models: Importance Sampling and Rewarding from an Expert-Centric Perspective
Yining Qian · Jinpeng Li · Fei Mi · Lifeng Shang · Xiang Zhang
摘要
Reinforcement learning (RL) has demonstrated considerable promise in enhancing large language models. However, its application to Mixture-of-Experts (MoE) architectures is frequently hindered by training instability, primarily stemming from token-level misalignment in expert assignments between c...