HPS: Hard Preference Sampling for Human Preference Alignment

TL;DR

HPS is a new training framework that improves human preference alignment in large language models by focusing on hard dispreferred responses, reducing computational costs, and enhancing safety and control.

Contribution

HPS introduces a novel loss and sampling strategy that prioritizes hard dispreferred responses, improving efficiency and safety in preference alignment for LLMs.

Findings

HPS achieves comparable BLEU and reward scores to existing methods.

HPS significantly increases reward margins, reducing harmful content.

HPS reduces computational overhead compared to traditional PL methods.

Abstract

Aligning Large Language Model (LLM) responses with human preferences is vital for building safe and controllable AI systems. While preference optimization methods based on Plackett-Luce (PL) and Bradley-Terry (BT) models have shown promise, they face challenges such as poor handling of harmful content, inefficient use of dispreferred responses, and, specifically for PL, high computational costs. To address these issues, we propose Hard Preference Sampling (HPS), a novel framework for robust and efficient human preference alignment. HPS introduces a training loss that prioritizes the most preferred response while rejecting all dispreferred and harmful ones. It emphasizes "hard" dispreferred responses -- those closely resembling preferred ones -- to enhance the model's rejection capabilities. By leveraging a single-sample Monte Carlo sampling strategy, HPS reduces computational overhead while maintaining alignment quality. Theoretically, HPS improves sample efficiency over existing PL methods and maximizes the reward margin between preferred and dispreferred responses, ensuring clearer distinctions. Experiments on HH-RLHF and PKU-Safety datasets validate HPS's effectiveness, achieving comparable BLEU and reward scores while greatly improving reward margins and thus reducing harmful content generation.