Venue
Neural Information Processing Systems
Domain
Natural Language Processing
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 introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss.The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight, eliminating the need for 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 PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.
This paper presents Direct Preference Optimization (DPO), a novel approach for fine-tuning large language models (LMs) without the complexities associated with reinforcement learning from human feedback (RLHF). DPO directly optimizes an implicit reward model based on human preferences, enabling the training of LMs using a straightforward binary cross-entropy loss. It simplifies the training process by eliminating the need for sampling and extensive hyperparameter tuning. The authors demonstrate that DPO aligns LMs with human preferences effectively, matching or surpassing existing methods such as PPO-based RLHF in various tasks, including sentiment modulation and summarization. The experiments conducted validate DPO's efficiency and effectiveness, showing superior performance in win rates against several baselines, including human-designed responses.
This paper employs the following methods:
- Direct Preference Optimization (DPO)
- PPO
The following datasets were used in this research:
- IMDB
- Reddit TL;DR
- Anthropic Helpful and Harmless dialogue
- ROUGE
- win rate against baseline
- KL-divergence
- binary cross-entropy loss
- DPO is stable, performant, and computationally lightweight
- DPO exceeds PPO-based RLHF in ability to control sentiment of generations
- DPO matches or improves response quality in summarization and single-turn dialogue
The authors identified the following limitations:
- The generalization of DPO policies out of distribution needs more comprehensive study.
- Exploration of scaling DPO to larger models is needed.
- Number of GPUs: None specified
- GPU Type: None specified
Preference Optimization
Reward Model
Reinforcement Learning from Human Feedback
RLHF
Preference-based Reinforcement Learning
Language Models
Fine-tuning
Alignment