Bei Li

Papers on Lattice

Total citations

Topics

h-index

Research focus

Natural Language Processing (1)RLHF & Preference Learning (1)Training Efficiency & Optimization (1)

Frequent co-authors

Ruichen Shao (1)Gangao Liu (1)Yang Chen (1)Xiang Zhou (1)

Papers (1)

Feb 20, 2025

Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective

By recognizing that not all tokens are created equal, D2PO offers a simple temporal weighting fix that boosts DPO alignment scores by up to 9.7 points.

Ruichen Shao, Bei Li, Gangao Liu +57

Natural Language Processing RLHF & Preference Learning Training Efficiency & Optimization

Search

Bei Li

Research focus

Frequent co-authors

Papers (1)