Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-constraint

By Wei Xiong, Hanze Dong, Chen Ye, Han Zhong, Nan Jiang, Tong Zhang. International Conference on Machine Learning, 2023

383 citations recorded by the source corpus.

This paper studies the alignment process of generative models with Reinforcement Learning from Human Feedback (RLHF). We first identify the primary challenges of existing popular methods like offline PPO and offline DPO as lacking in strategical exploration of the environment. Then, to understand the mathematical principle of RLHF, we consider a standard mathematical formulation, the reverse-KL regularized contextual bandit for RLHF. Despite its widespread practical application, a rigorous theoretical analysis of this formulation remains open. We investigate its behavior in three distinct settings -- offline, online, and hybrid -- and propose efficient algorithms with finite-sample theoretical guarantees. Moving towards practical applications, our framework, with a robust approximation of the information-theoretical policy improvement oracle, naturally gives rise to several novel RLHF algorithms. This includes an iterative version of the Direct Preference Optimization (DPO) algorithm for online settings, and a multi-step rejection sampling strategy for offline scenarios. Our empirical evaluations on real-world alignment experiment of large language model demonstrate that these proposed methods significantly surpass existing strong baselines, such as DPO and Rejection Sampling Optimization (RSO), showcasing the connections between solid theoretical foundations and their potent practical implementations.

Subfields: Artificial Intelligence, Cultural Studies

View Paper: https://www.semanticscholar.org/paper/44a9d8b0314d34aff91ccff9207d38eed37216ed

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cultural Studies

Yejin Choi

Citations in the corpus

Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing (2024, positive): “1 [57] as our reward model.”
SimPO: Simple Preference Optimization with a Reference-Free Reward (2024, positive): “Further studies have extended this approach to an iterative training setup, by continuously updating the reference model with the most recent policy model or generating new preference pairs at each iteration [27, 42, 62, 82, 87].”
Towards Data-Centric RLHF: Simple Metrics for Preference Dataset Comparison (2024, background): “While a substantial number of works have focused on new algorithms for learning from preference data to better train reward models [6, 7, 8, 9], relatively few works have examined qualities of these datasets themselves.”
RLHF Workflow: From Reward Modeling to Online RLHF (2024, methodology): “…a proxy reward constructed from the same offline dataset (similar to semi-supervised learning), iterative rejection sampling fine-tuning (RAFT) (Dong et al., 2023) and DPO based on samples from the target distribution estimator have been shown to outperform the original offline counterparts…”