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

2025-02-18

AI Summary

To address the efficiency–accuracy trade-off in deploying safety guard models for large language models (LLMs), this paper proposes a dynamic routing mechanism. It introduces a lightweight binary router that automatically classifies input prompts as “easy” or “hard” based on confidence and uncertainty estimates; only “hard” instances trigger the heavyweight safety guard, while “easy” ones are processed by a distilled lightweight model. Our key contribution is the first model-level, fine-grained adaptive scheduling architecture, integrating multi-stage safety assessment and collaborative inference—overcoming limitations of conventional static distillation or uniform heavyweight guarding. Evaluated across multiple safety benchmarks, our approach reduces average latency by 42% compared to full heavyweight guarding, maintains >99.3% harmful content detection accuracy, and achieves significantly higher interception accuracy per unit computational cost.

Technology Category

• Natural Language Processing: Safety and Robustness
• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Security and Privacy: Large-scale security measurements
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

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.

Problem

• Efficient safety guardrails for LLMs
• Selective use of large guard models
• Balancing computational cost and accuracy

Innovation

• Adaptive model selection
• Binary router efficiency
• Selective large model application