KTO：KTO: Model Alignment as Prospect Theoretic Optimization

ICML · Stanford University, Contextual AI

借鉴前景理论提出KTO，仅需输出的二元好坏标签即可直接最大化生成效用，达到或超越DPO的效果。