Preference as Reward, Maximum Preference Optimization with Importance Sampling (2312.16430v5)
Abstract: Preference learning is a key technology for aligning LLMs with human values. Reinforcement Learning from Human Feedback (RLHF) is a model-based algorithm to optimize preference learning, which first fits a reward model for preference scores and then optimizes the generating policy with an on-policy PPO algorithm to maximize the reward. The processing of RLHF is complex, time-consuming, and unstable. The Direct Preference Optimization (DPO) algorithm uses an off-policy algorithm to directly optimize the generating policy and eliminates the need for a reward model. DPO is more data-efficient and stable. However, DPO has a drawback of overfitting to the preference data and ignoring the KL-regularization term when the preference is deterministic. Identity mapping Preference Optimization(IPO) uses a root-finding MSE loss to incorporate KL-regularization. However, both DPO and IPO fail to properly address the KL-regularization term because the support of the preference distribution is not equal to the reference distribution. In this paper, we propose a simple and intuitive off-policy preference optimization algorithm from an importance sampling view, which we call Maximum Preference Optimization (MPO). MPO incorporates the off-policy KL-regularization term, making regularization truly effective. MPO achieves the best of both worlds by combining the objectives of RLHF and IPO while being an off-policy algorithm. Furthermore, MPO eliminates the need for a reward model and reference policy, simplifying the learning process and reducing memory usage.
- Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.
- Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1–113, 2023.
- Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023.
- Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
- Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021.
- Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.
- Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019.
- Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022.
- Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017.
- Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008–3021, 2020.
- Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022.
- Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
- Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324–345, 1952.
- Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023.
- A general theoretical paradigm to understand learning from human preferences. arXiv preprint arXiv:2310.12036, 2023.
- Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928–1937. PMLR, 2016.
Collections
Sign up for free to add this paper to one or more collections.
Paper Prompts
Sign up for free to create and run prompts on this paper using GPT-5.