DPO-Shift: Shifting the Distribution of Direct Preference Optimization
This work addresses a significant problem for researchers and practitioners using DPO for aligning language models with human preferences, providing an incremental yet effective solution.
The authors tackled the problem of likelihood displacement in Direct Preference Optimization (DPO) and achieved a solution with DPO-Shift, which exhibits a trade-off between improving the chosen probability and sacrificing the reward margin, with demonstrated superiority over DPO on downstream tasks. Experimental validation showed the effectiveness of DPO-Shift.
Direct Preference Optimization (DPO) and its variants have become increasingly popular for aligning language models with human preferences. These methods aim to teach models to better distinguish between chosen (or preferred) and rejected (or dispreferred) responses. However, prior research has identified that the probability of chosen responses often decreases during training, and this phenomenon is known as likelihood displacement. To tackle this challenge, in this work we introduce DPO-Shift to controllably shift the distribution of the chosen probability. Then, we show that DPO-Shift exhibits a fundamental trade-off between improving the chosen probability and sacrificing the reward margin, as supported by both theoretical analysis and experimental validation. Furthermore, we demonstrate the superiority of DPO-Shift over DPO on downstream tasks such as MT-Bench and a designed win rate experiment. We believe this study shows that the likelihood displacement issue of DPO can be effectively mitigated with a simple, theoretically grounded solution. Our code is available at https://github.com/Meaquadddd/DPO-Shift.