SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models

TL;DR

SafeRoute is an adaptive system that intelligently routes safety guard tasks between small and large language models, optimizing for both efficiency and accuracy in detecting harmful prompts.

Contribution

We introduce SafeRoute, a binary router that selectively applies large models only to hard examples, reducing computational costs while maintaining safety performance.

Findings

Significantly reduces computational cost compared to using only large models.

Maintains high safety detection accuracy on benchmark datasets.

Outperforms baseline methods in efficiency-accuracy trade-offs.

Abstract

Deploying large language models (LLMs) in real-world applications requires robust safety guard models to detect and block harmful user prompts. While large safety guard models achieve strong performance, their computational cost is substantial. To mitigate this, smaller distilled models are used, but they often underperform on "hard" examples where the larger model provides accurate predictions. We observe that many inputs can be reliably handled by the smaller model, while only a small fraction require the larger model's capacity. Motivated by this, we propose SafeRoute, a binary router that distinguishes hard examples from easy ones. Our method selectively applies the larger safety guard model to the data that the router considers hard, improving efficiency while maintaining accuracy compared to solely using the larger safety guard model. Experimental results on multiple benchmark datasets demonstrate that our adaptive model selection significantly enhances the trade-off between computational cost and safety performance, outperforming relevant baselines.