Online Self-Preferring Language Models

TL;DR

This paper introduces Online Self-Preferring (OSP) language models that learn from self-generated response pairs and preference strengths, improving alignment, robustness, and self-improvement without external supervision.

Contribution

The paper proposes a novel OSP method that explicitly models preference strength, leading to better alignment and robustness compared to existing offline and online methods.

Findings

OSP achieves state-of-the-art alignment performance.

Leveraging preference strength prevents overfitting.

OSP models can self-improve without external supervision.

Abstract

Aligning with human preference datasets has been critical to the success of large language models (LLMs). Reinforcement learning from human feedback (RLHF) employs a costly reward model to provide feedback for on-policy sampling responses. Recently, offline methods that directly fit responses with binary preferences in the dataset have emerged as alternatives. However, existing methods do not explicitly model preference strength information, which is crucial for distinguishing different response pairs. To overcome this limitation, we propose Online Self-Preferring (OSP) language models to learn from self-generated response pairs and self-judged preference strengths. For each prompt and corresponding self-generated responses, we introduce a ranked pairing method to construct multiple response pairs with preference strength information. We then propose the soft-preference cross-entropy loss to leverage such information. Empirically, we demonstrate that leveraging preference strength is crucial for avoiding overfitting and enhancing alignment performance. OSP achieves state-of-the-art alignment performance across various metrics in two widely used human preference datasets. OSP is parameter-efficient and more robust than the dominant online method, RLHF when limited offline data are available and generalizing to out-of-domain tasks. Moreover, OSP language models established by LLMs with proficiency in self-preferring can efficiently self-improve without external supervision.