Hybrid Alignment Training for Large Language Models

TL;DR

This paper introduces Hybrid Alignment Training (Hbat), a novel method that alternates between instruction-following and human-preference alignment to improve large language model alignment, demonstrating significant performance gains.

Contribution

The paper proposes a new hybrid training approach that alternates alignment objectives, addressing conflicts in traditional sequential alignment methods for large language models.

Findings

Hbat outperforms all baseline methods in summarization and dialogue tasks.

Hbat achieves consistent improvements over traditional two-stage alignment training.

Experimental results validate the effectiveness of alternating alignment objectives.

Abstract

Alignment training is crucial for enabling large language models (LLMs) to cater to human intentions and preferences. It is typically performed based on two stages with different objectives: instruction-following alignment and human-preference alignment. However, aligning LLMs with these objectives in sequence suffers from an inherent problem: the objectives may conflict, and the LLMs cannot guarantee to simultaneously align with the instructions and human preferences well. To response to these, in this work, we propose a Hybrid Alignment Training (Hbat) approach, based on alternating alignment and modified elastic weight consolidation methods. The basic idea is to alternate between different objectives during alignment training, so that better collaboration can be achieved between the two alignment tasks.We experiment with Hbat on summarization and dialogue tasks. Experimental results show that the proposed \textsc{Hbat} can significantly outperform all baselines. Notably, Hbat yields consistent performance gains over the traditional two-stage alignment training when using both proximal policy optimization and direct preference optimization.