Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Disentangling Length from Quality in Direct Preference Optimization

About

Reinforcement Learning from Human Feedback (RLHF) has been a crucial component in the recent success of Large Language Models. However, RLHF is know to exploit biases in human preferences, such as verbosity. A well-formatted and eloquent answer is often more highly rated by users, even when it is less helpful and objective. A number of approaches have been developed to control those biases in the classical RLHF literature, but the problem remains relatively under-explored for Direct Alignment Algorithms such as Direct Preference Optimization (DPO). Unlike classical RLHF, DPO does not train a separate reward model or use reinforcement learning directly, so previous approaches developed to control verbosity cannot be directly applied to this setting. Our work makes several contributions. For the first time, we study the length problem in the DPO setting, showing significant exploitation in DPO and linking it to out-of-distribution bootstrapping. We then develop a principled but simple regularization strategy that prevents length exploitation, while still maintaining improvements in model quality. We demonstrate these effects across datasets on summarization and dialogue, where we achieve up to 20\% improvement in win rates when controlling for length, despite the GPT4 judge's well-known verbosity bias.

Ryan Park, Rafael Rafailov, Stefano Ermon, Chelsea Finn• 2024

Related benchmarks

TaskDatasetResultRank
Instruction FollowingAlpacaEval 2.0
Win Rate66.9
722
Multi-turn Dialogue EvaluationMT-Bench--
532
Instruction FollowingAlpacaEval
Win Rate46.98
420
Instruction FollowingMT-Bench
MT-Bench Score7.05
287
Instruction FollowingArena Hard
Win Rate57.9
263
Instruction FollowingAlpacaEval 2
LC (%)68.3
137
Multi-turn Instruction FollowingMT-Bench
MT-Bench Score (GPT-4)6.79
129
Multi-turn dialogueMT-Bench
MT-Bench Score8.2
126
Multilingual Mathematical ReasoningMGSM (test)
Accuracy65.49
109
Multilingual Language UnderstandingMMMLU (test)
Accuracy65.23
52
Showing 10 of 19 rows

Other info

Code

Follow for update