Aligning Large Language Models with Implicit Preferences from User-Generated Content

TL;DR

This paper introduces PUGC, a scalable framework that leverages implicit preferences from unlabeled user-generated content to improve large language model alignment with human values, achieving state-of-the-art results.

Contribution

The paper presents PUGC, a novel method that uses implicit preferences from UGC for scalable LLM alignment, reducing reliance on costly curated data.

Findings

Models trained with PUGC outperform traditional methods by 9.37% on Alpaca Eval 2.

Achieves a 35.93% state-of-the-art length-controlled win rate.

Enhances reward quality, domain-specific alignment, and robustness.

Abstract

Learning from preference feedback is essential for aligning large language models (LLMs) with human values and improving the quality of generated responses. However, existing preference learning methods rely heavily on curated data from humans or advanced LLMs, which is costly and difficult to scale. In this work, we present PUGC, a novel framework that leverages implicit human Preferences in unlabeled User-Generated Content (UGC) to generate preference data. Although UGC is not explicitly created to guide LLMs in generating human-preferred responses, it often reflects valuable insights and implicit preferences from its creators that has the potential to address readers' questions. PUGC transforms UGC into user queries and generates responses from the policy model. The UGC is then leveraged as a reference text for response scoring, aligning the model with these implicit preferences. This approach improves the quality of preference data while enabling scalable, domain-specific alignment. Experimental results on Alpaca Eval 2 show that models trained with DPO and PUGC achieve a 9.37% performance improvement over traditional methods, setting a 35.93% state-of-the-art length-controlled win rate using Mistral-7B-Instruct. Further studies highlight gains in reward quality, domain-specific alignment effectiveness, robustness against UGC quality, and theory of mind capabilities. Our code and dataset are available at https://zhaoxuan.info/PUGC.github.io/