Training AI to generate diverse, natural responses without costly reward models

Original: Pairwise Preference Reward and Group-Based Diversity Enhancement for Superior Open-Ended Generation

Source: arxiv.org ↗