Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer

• URL: http://arxiv.org/abs/2405.16436v2
• Date: Mon, 04 Nov 2024 06:17:06 GMT
• Title: Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer
• Authors: Zhihan Liu, Miao Lu, Shenao Zhang, Boyi Liu, Hongyi Guo, Yingxiang Yang, Jose Blanchet, Zhaoran Wang,
• Abstract summary: We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences. To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model. Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
• Score: 52.09480867526656
• License:
• Abstract: Aligning generative models with human preference via RLHF typically suffers from overoptimization, where an imperfectly learned reward model can misguide the generative model to output undesired responses. We investigate this problem in a principled manner by identifying the source of the misalignment as a form of distributional shift and uncertainty in learning human preferences. To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model; one that simultaneously minimizes the maximum likelihood estimation of the loss and a reward penalty term. Here, the reward penalty term is introduced to prevent the policy from choosing actions with spurious high proxy rewards, resulting in provable sample efficiency of the algorithm under a partial coverage style condition. Moving from theory to practice, the proposed algorithm further enjoys an equivalent but surprisingly easy-to-implement reformulation. Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines: (i) a preference optimization loss that directly aligns the policy with human preference, and (ii) a supervised learning loss that explicitly imitates the policy with a (suitable) baseline distribution. In the context of aligning large language models (LLM), this objective fuses the direct preference optimization (DPO) loss with the supervised fine-tuning (SFT) loss to help mitigate the overoptimization towards undesired responses, for which we name the algorithm Regularized Preference Optimization (RPO). Experiments of aligning LLMs demonstrate the improved performance of RPO compared with DPO baselines. Our work sheds light on the interplay between preference optimization and SFT in tuning LLMs with both theoretical guarantees and empirical evidence.

Related papers

• Uncertainty-Penalized Direct Preference Optimization [52.387088396044206]
  We develop a pessimistic framework for DPO by introducing preference uncertainty penalization schemes. The penalization serves as a correction to the loss which attenuates the loss gradient for uncertain samples. We show improved overall performance compared to vanilla DPO, as well as better completions on prompts from high-uncertainty chosen/rejected responses.
  arXiv  Detail & Related papers  (2024-10-26T14:24:37Z)
• Accelerated Preference Optimization for Large Language Model Alignment [60.22606527763201]
  Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal tool for aligning large language models (LLMs) with human preferences. Direct Preference Optimization (DPO) formulates RLHF as a policy optimization problem without explicitly estimating the reward function. We propose a general Accelerated Preference Optimization (APO) framework, which unifies many existing preference optimization algorithms.
  arXiv  Detail & Related papers  (2024-10-08T18:51:01Z)
• Discovering Preference Optimization Algorithms with and for Large Language Models [50.843710797024805]
  offline preference optimization is a key method for enhancing and controlling the quality of Large Language Model (LLM) outputs. We perform objective discovery to automatically discover new state-of-the-art preference optimization algorithms without (expert) human intervention. Experiments demonstrate the state-of-the-art performance of DiscoPOP, a novel algorithm that adaptively blends logistic and exponential losses.
  arXiv  Detail & Related papers  (2024-06-12T16:58:41Z)
• $i$REPO: $i$mplicit Reward Pairwise Difference based Empirical Preference Optimization [12.266207199002604]
  Large Language Models (LLM) can sometimes produce outputs that deviate from human expectations. We propose a novel framework named $i$REPO, which utilizes implicit Reward pairwise difference regression for Empirical Preference Optimization. We show that $i$REPO effectively achieves self-alignment using soft-label, self-generated responses and the logit of empirical AI annotators.
  arXiv  Detail & Related papers  (2024-05-24T05:42:11Z)
• Overcoming Reward Overoptimization via Adversarial Policy Optimization with Lightweight Uncertainty Estimation [46.61909578101735]
  Adversarial Policy Optimization (AdvPO) is a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback. In this paper, we introduce a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model.
  arXiv  Detail & Related papers  (2024-03-08T09:20:12Z)
• Towards Efficient Exact Optimization of Language Model Alignment [93.39181634597877]
  Direct preference optimization (DPO) was proposed to directly optimize the policy from preference data. We show that DPO derived based on the optimal solution of problem leads to a compromised mean-seeking approximation of the optimal solution in practice. We propose efficient exact optimization (EXO) of the alignment objective.
  arXiv  Detail & Related papers  (2024-02-01T18:51:54Z)
• Preference as Reward, Maximum Preference Optimization with Importance Sampling [3.7040071165219595]
  We propose a simple and intuitive off-policy preference optimization algorithm from an importance sampling view, which we call Maximum Preference Optimization (MPO) MPO achieves the best of both worlds by combining the objectives of RLHF and IPO while being an off-policy algorithm.
  arXiv  Detail & Related papers  (2023-12-27T06:34:54Z)
• Statistical Rejection Sampling Improves Preference Optimization [42.57245965632205]
  We introduce a novel approach to source preference data from the target optimal policy using rejection sampling. We also propose a unified framework that enhances the loss functions used in both Sequence Likelihood (SLiC) and Direct Preference Optimization (DPO) from a preference modeling standpoint.
  arXiv  Detail & Related papers  (2023-09-13T01:07:25Z)
• Fine-Tuning Language Models with Advantage-Induced Policy Alignment [80.96507425217472]
  We propose a novel algorithm for aligning large language models to human preferences. We show that it consistently outperforms PPO in language tasks by a large margin. We also provide a theoretical justification supporting the design of our loss function.
  arXiv  Detail & Related papers  (2023-06-04T01:59:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.