RetireOPD

A new method called RetireOPD proposes self-retiring on-policy distillation to give multi-turn reinforcement learning agents dense token-level supervision from a self-teacher with privileged task skil…