Optimizing Diversity and Quality through Base-Aligned Model Collaboration

Yichen Wang, Chenghao Yang, Tenghao Huang +3

Alignment has greatly improved large language models (LLMs)' output quality at the cost of diversity, yielding highly similar outputs across generations, especially in open-ended g…