FlipGuard: Defending Preference Alignment against Update Regression with Constrained Optimization

TL;DR

FlipGuard is a constrained optimization method designed to prevent preference regression in large language models during updates, ensuring they retain previous alignments while improving overall performance.

Contribution

The paper introduces FlipGuard, a novel approach that detects and mitigates update regression in preference alignment through constrained optimization and focal attention.

Findings

Effectively reduces update regression in preference alignment.

Maintains high overall alignment performance.

Preserves knowledge during model updates.

Abstract

Recent breakthroughs in preference alignment have significantly improved Large Language Models' ability to generate texts that align with human preferences and values. However, current alignment metrics typically emphasize the post-hoc overall improvement, while overlooking a critical aspect: regression, which refers to the backsliding on previously correctly-handled data after updates. This potential pitfall may arise from excessive fine-tuning on already well-aligned data, which subsequently leads to over-alignment and degeneration. To address this challenge, we propose FlipGuard, a constrained optimization approach to detect and mitigate update regression with focal attention. Specifically, FlipGuard identifies performance degradation using a customized reward characterization and strategically enforces a constraint to encourage conditional congruence with the pre-aligned model during training. Comprehensive experiments demonstrate that FlipGuard effectively alleviates update regression while demonstrating excellent overall performance, with the added benefit of knowledge preservation while aligning preferences.