Disentangling Length Bias In Preference Learning Via Response-Conditioned Modeling

TL;DR

This paper introduces a response-conditioned framework to mitigate length bias in preference learning for LLMs, improving alignment with human preferences and length instructions.

Contribution

The paper proposes the Rc-BT model and associated algorithms to explicitly differentiate semantic preferences from length, addressing length bias in reward modeling and policy optimization.

Findings

Effective reduction of length bias across models

Improved adherence to explicit length instructions

Enhanced alignment with human preferences

Abstract

Reinforcement Learning from Human Feedback (RLHF) has achieved considerable success in aligning large language models (LLMs) by modeling human preferences with a learnable reward model and employing a reinforcement learning algorithm to maximize the reward model's scores. However, these reward models are susceptible to exploitation through various superficial confounding factors, with length bias emerging as a particularly significant concern. Moreover, while the pronounced impact of length bias on preference modeling suggests that LLMs possess an inherent sensitivity to length perception, our preliminary investigations reveal that fine-tuned LLMs consistently struggle to adhere to explicit length instructions. To address these two limitations, we propose a novel framework wherein the reward model explicitly differentiates between human semantic preferences and response length requirements. Specifically, we introduce a $\textbf{R}$esponse-$\textbf{c}$onditioned $\textbf{B}$radley-$\textbf{T}$erry (Rc-BT) model that enhances the model's capability in length bias mitigating and length instruction following, through training on our augmented dataset. Furthermore, we propose the Rc-RM and Rc-DPO algorithm to leverage the Rc-BT model for reward modeling and direct policy optimization (DPO) of LLMs, simultaneously mitigating length bias and promoting adherence to length instructions. Extensive experiments across various foundational models and datasets demonstrate the effectiveness and generalizability of our approach.