One-Shot Safety Alignment for Large Language Models via Optimal Dualization

TL;DR

This paper introduces a novel dualization approach to simplify and stabilize the process of aligning large language models with safety constraints, significantly reducing computational costs.

Contribution

It proposes a dualization-based method that converts constrained alignment into an unconstrained problem, enabling efficient and stable training of safe language models.

Findings

Reduces computational cost of safety alignment

Improves training stability for large language models

Demonstrates effectiveness through extensive experiments

Abstract

The growing safety concerns surrounding large language models raise an urgent need to align them with diverse human preferences to simultaneously enhance their helpfulness and safety. A promising approach is to enforce safety constraints through Reinforcement Learning from Human Feedback (RLHF). For such constrained RLHF, typical Lagrangian-based primal-dual policy optimization methods are computationally expensive and often unstable. This paper presents a perspective of dualization that reduces constrained alignment to an equivalent unconstrained alignment problem. We do so by pre-optimizing a smooth and convex dual function that has a closed form. This shortcut eliminates the need for cumbersome primal-dual policy iterations, greatly reducing the computational burden and improving training stability. Our strategy leads to two practical algorithms in model-based and preference-based settings (MoCAN and PeCAN, respectively). A broad range of experiments demonstrate the effectiveness and merits of our algorithms.