Oral Poster

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Rafael Rafailov · Archit Sharma · Eric Mitchell · Christopher D Manning · Stefano Ermon · Chelsea Finn

Great Hall & Hall B1+B2 (level 1) #625
award Outstanding Paper Runner-up
[ ]
Thu 14 Dec 3 p.m. PST — 5 p.m. PST
Oral presentation: Oral 6B RL
Thu 14 Dec 1:20 p.m. PST — 2:20 p.m. PST


While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper, we leverage a mapping between reward functions and optimal policies to show that this constrained reward maximization problem can be optimized exactly with a single stage of policy training, essentially solving a classification problem on the human preference data. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight, eliminating the need for fitting a reward model, sampling from the LM during fine-tuning, or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds RLHF's ability to control sentiment of generations and improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.

Chat is not available.