Robust Preference Optimization through Reward Model Distillation

• URL: http://arxiv.org/abs/2405.19316v1
• Date: Wed, 29 May 2024 17:39:48 GMT
• Title: Robust Preference Optimization through Reward Model Distillation
• Authors: Adam Fisch, Jacob Eisenstein, Vicky Zayats, Alekh Agarwal, Ahmad Beirami, Chirag Nagpal, Pete Shaw, Jonathan Berant,
• Abstract summary: Language model (LM) post-training involves maximizing a reward function that is derived from preference annotations. DPO is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. We analyze this phenomenon and propose distillation to get a better proxy for the true preference distribution over generation pairs.
• Score: 68.65844394615702
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Language model (LM) post-training (or alignment) involves maximizing a reward function that is derived from preference annotations. Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. However, typical preference datasets have only a single, or at most a few, annotation per preference pair, which causes DPO to overconfidently assign rewards that trend towards infinite magnitude. This frequently leads to degenerate policies, sometimes causing even the probabilities of the preferred generations to go to zero. In this work, we analyze this phenomenon and propose distillation to get a better proxy for the true preference distribution over generation pairs: we train the LM to produce probabilities that match the distribution induced by a reward model trained on the preference data. Moreover, to account for uncertainty in the reward model we are distilling from, we optimize against a family of reward models that, as a whole, is likely to include at least one reasonable proxy for the preference distribution. Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations, while preserving the simple supervised nature of DPO.

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)
• Simultaneous Reward Distillation and Preference Learning: Get You a Language Model Who Can Do Both [6.102274021710727]
  Direct Reward Distillation and policy-Optimization (DRDO) is a supervised knowledge distillation-based preference alignment method. DRDO directly mimics rewards assigned by an oracle while learning human preferences from a novel preference likelihood formulation. Our experimental results on the Ultrafeedback and TL;DR datasets demonstrate that policies trained using DRDO surpass previous methods.
  arXiv  Detail & Related papers  (2024-10-11T02:19:11Z)
• General Preference Modeling with Preference Representations for Aligning Language Models [51.14207112118503]
  We introduce preference representation learning, an approach that embeds responses into a latent space to capture intricate preference structures efficiently. We also propose preference score-based General Preference Optimization (GPO), which generalizes reward-based reinforcement learning from human feedback. Our method may enhance the alignment of foundation models with nuanced human values.
  arXiv  Detail & Related papers  (2024-10-03T04:22:55Z)
• Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models [54.132297393662654]
  We introduce a hybrid method that fine-tunes cutting-edge diffusion models by optimizing reward models through RL. We demonstrate the capability of our approach to outperform the best designs in offline data, leveraging the extrapolation capabilities of reward models.
  arXiv  Detail & Related papers  (2024-05-30T03:57:29Z)
• Rejection via Learning Density Ratios [50.91522897152437]
  Classification with rejection emerges as a learning paradigm which allows models to abstain from making predictions. We propose a different distributional perspective, where we seek to find an idealized data distribution which maximizes a pretrained model's performance. Our framework is tested empirically over clean and noisy datasets.
  arXiv  Detail & Related papers  (2024-05-29T01:32:17Z)
• Direct Preference Optimization With Unobserved Preference Heterogeneity [16.91835461818937]
  This paper presents a new method to align generative models with varied human preferences. We propose an Expectation-Maximization adaptation to DPO, generating a mixture of models based on latent preference types of the annotators. Our algorithms leverage the simplicity of DPO while accommodating diverse preferences.
  arXiv  Detail & Related papers  (2024-05-23T21:25:20Z)
• Soft Preference Optimization: Aligning Language Models to Expert Distributions [40.84391304598521]
  SPO is a method for aligning generative models, such as Large Language Models (LLMs), with human preferences. SPO integrates preference loss with a regularization term across the model's entire output distribution. We showcase SPO's methodology, its theoretical foundation, and its comparative advantages in simplicity, computational efficiency, and alignment precision.
  arXiv  Detail & Related papers  (2024-04-30T19:48:55Z)
• Aligning Crowd Feedback via Distributional Preference Reward Modeling [28.754532173765686]
  We propose the Distributional Preference Reward Model (DPRM) to align large language models with diverse human preferences. Our experiments show that DPRM significantly enhances the alignment of LLMs with population preference, yielding more accurate, unbiased, and contextually appropriate responses.
  arXiv  Detail & Related papers  (2024-02-15T07:29:43Z)
• 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)
• Models of human preference for learning reward functions [80.39289349661364]
  We learn the reward function from human-generated preferences between pairs of trajectory segments. We find this assumption to be flawed and propose modeling human preferences as informed by each segment's regret. Our proposed regret preference model better predicts real human preferences and also learns reward functions from these preferences that lead to policies that are better human-aligned.
  arXiv  Detail & Related papers  (2022-06-05T17:58:02Z)

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.