Online Safety Monitoring for LLMs
For practitioners deploying LLMs, this provides a simple and effective method for online safety monitoring.
The paper tackles the problem of unsafe outputs from LLMs at deployment time and proposes a simple real-time monitor using thresholding on a verifier signal calibrated via risk control, showing it is competitive with more advanced sequential hypothesis testing monitors on mathematical reasoning and red teaming datasets.
Despite alignment training, LLMs remain prone to generating unsafe outputs at deployment time. Monitoring outputs online and raising an alarm when safety can no longer be assumed is therefore critical. We study a simple real-time monitor that turns a verifier signal from an external model into an alarm decision by thresholding, with the threshold calibrated via risk control. In experiments on mathematical reasoning and red teaming datasets, we show that this simple design is competitive with more advanced monitors based on sequential hypothesis testing.