.png){kind=icon}
Return
Reward-guided direct preference optimization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130295.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Analyzed the likelihood changes of chosen and rejected responses in SFT and DPO. • Proposed a dynamic switching method for balancing likelihood and ranking objectives. • RG-DPO outperforms DPO for preference learning. • Proposed algorithm achieves competitive performance on HH-RLHF and TLDR datasets.
Keywords:
Alignment
Large language model
Preference optimization
Natural language processing
Journal
IF:
7.5
Papers:
3.0W
Citations:
10.2W
.png){kind=icon}
