Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training

TL;DR

This paper introduces Decoupled Refusal Training (DeRTa), a novel method to improve LLM safety by enabling models to refuse unsafe content at any response point, addressing bias issues in safety tuning data.

Contribution

The paper presents DeRTa, a new training approach with two components that significantly enhances LLM safety and refusal capabilities without sacrificing performance.

Findings

DeRTa improves safety across multiple attack scenarios.

Models trained with DeRTa outperform baseline safety methods.

DeRTa maintains model performance while increasing safety.

Abstract

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses baseline methods in defending against attacks.