liushunyu/awesome-direct-preference-optimization

Curates 250+ peer-reviewed papers organized by a novel taxonomy that decomposes DPO methodologies across four dimensions: data strategy, learning framework, constraint mechanisms, and model properties. Provides systematic categorization of DPO variations spanning data quality and preference feedback approaches, learning paradigms and objectives, reference model constraints and safety mechanisms, and generation/optimization properties. Bridges foundational DPO work with recent extensions including heterogeneous preference handling, dynamic weighting schemes, and robustness improvements.

No commits in the last 6 months.