diversity-exploration

This paper presents SPADER, a reinforcement learning framework for multi-answer QA that uses step-wise peer advantage for credit assignment and diversity-aware exploration rewards to improve recall of long-tail entities, achieving better performance on several benchmarks.