Tag
This paper investigates why reasoning models lose coverage during supervised fine-tuning, linking the phenomenon to decision points in training data where multiple valid paths exist, and proposes data synthesis and diversity-aware decoding as mitigations.