This research addresses a critical gap in LLM safety by examining the reliability and calibration of guard models designed to moderate harmful content.

• Evaluates the performance of LLM-based guard models that filter inappropriate content
• Identifies calibration issues that affect real-world reliability of content moderation
• Proposes methods to improve guard model confidence calibration
• Demonstrates how proper calibration enhances moderation precision and effectiveness

Security Implications: Properly calibrated guard models provide more dependable protection against harmful content generation, reducing false positives/negatives in content filtering systems while maintaining robustness against evasion attempts.

On Calibration of LLM-based Guard Models for Reliable Content Moderation