Advancing LLM Safe Alignment with Safety Representation Ranking

TL;DR

This paper introduces Safety Representation Ranking (SRR), a novel method that uses internal LLM states to better evaluate and select safe responses, enhancing robustness against harmful content.

Contribution

The paper presents a new ranking framework that leverages internal model representations for safety assessment, improving upon existing response-based safety evaluation methods.

Findings

SRR significantly improves safety robustness against adversarial prompts.

Utilizes internal transformer states for safety evaluation.

Achieves better safety performance across multiple benchmarks.

Abstract

The rapid advancement of large language models (LLMs) has demonstrated milestone success in a variety of tasks, yet their potential for generating harmful content has raised significant safety concerns. Existing safety evaluation approaches typically operate directly on textual responses, overlooking the rich information embedded in the model's internal representations. In this paper, we propose Safety Representation Ranking (SRR), a listwise ranking framework that selects safe responses using hidden states from the LLM itself. SRR encodes both instructions and candidate completions using intermediate transformer representations and ranks candidates via a lightweight similarity-based scorer. Our approach directly leverages internal model states and supervision at the list level to capture subtle safety signals. Experiments across multiple benchmarks show that SRR significantly improves robustness to adversarial prompts. Our code will be available upon publication.