Why AI Distillation Can Produce Repetitive, Overlong Output

Original: Gains and Collapse in On-Policy Distillation:A Reinforcement Learning Perspective

Source: arxiv.org ↗