GuardReasoner: Towards Reasoning-based LLM Safeguards

TL;DR

GuardReasoner is a reasoning-based safeguard for LLMs that improves safety, explainability, and performance by training on a large dataset with detailed reasoning steps and employing specialized training techniques.

Contribution

The paper introduces GuardReasoner, a novel reasoning-guided safeguard for LLMs, with new training datasets and methods that enhance reasoning, safety, and generalization.

Findings

Outperforms GPT-4o+CoT by 5.74% F1 score

Surpasses LLaMA Guard 3 8B by 20.84% F1 score

Demonstrates superior performance on 13 benchmark tasks

Abstract

As LLMs increasingly impact safety-critical applications, ensuring their safety using guardrails remains a key challenge. This paper proposes GuardReasoner, a new safeguard for LLMs, by guiding the guard model to learn to reason. Concretely, we first create the GuardReasonerTrain dataset, which consists of 127K samples with 460K detailed reasoning steps. Then, we introduce reasoning SFT to unlock the reasoning capability of guard models. In addition, we present hard sample DPO to further strengthen their reasoning ability. In this manner, GuardReasoner achieves better performance, explainability, and generalizability. Extensive experiments and analyses on 13 benchmarks of 3 guardrail tasks demonstrate its superiority. Remarkably, GuardReasoner 8B surpasses GPT-4o+CoT by 5.74% and LLaMA Guard 3 8B by 20.84% F1 score on average. We release the training data, code, and models with different scales (1B, 3B, 8B) of GuardReasoner : https://github.com/yueliu1999/GuardReasoner/.