Tag
This paper introduces a reinforcement learning-based distillation framework for training compact instruction-following rerankers, using off-policy GRPO for teacher enhancement and on-policy distillation for student learning, demonstrating superior performance under distribution shift.