2025
DPL: Diverse Preference Learning Without A Reference Model
Abhijnan Nath, Andrey Volozin, Saumajit Saha, Albert Aristotle Nanda, Galina Grunin, Rahul Bhotika +1
NAACL 2025long
In direct preference alignment in LLMs, most existing methods seek to retrieve the reward function directly from preference data. However, real-world preference data often contains diversity in preference annotations reflective of true human preferences. Existing algorithms, including KTO, do not di…