D^2-Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation-Aware Routing

Researchers introduced D^2-Monitor, a two-stage safety monitor for diffusion LLMs that watches intermediate denoising states.

It uses hesitation signals to detect when a model is near a decision boundary and only then routes cases to a heavier probe.

Across three safety datasets and four diffusion LLMs, it achieved state-of-the-art detection with a small parameter budget.

The approach improves text safety monitoring by spending compute more efficiently while preserving strong performance.