Self-Review Reinforcement Learning (SRRL) with Cross-Episode Memory and Policy Distillation

Muhammad Zain Amin, Kibele Sebnem Yildirim

Reinforcement Learning is commonly used to train large language models using environmental feedback. In applied settings, the environment usually provides sparse or delayed feedbac…