SAD framework steers text diffusion models toward provably safe outputs

Text diffusion models are gaining traction as an alternative to autoregressive generation, offering parallelism and flexibility. However, their safety control remains underexplored—existing approaches rely on post-hoc filtering or inference-time interventions designed for autoregressive models, which are inadequate for diffusion models. Researchers introduced the Safety-Aware Denoiser (SAD), which modifies the iterative denoising process to steer final outputs toward provably safe regions of text space. This inference-time method integrates safety constraints directly into the denoiser, avoiding expensive retraining of the underlying diffusion model and enabling lightweight, flexible safety guidance.

SAD was evaluated across three dimensions: hazard taxonomy, memorization of unsafe content, and jailbreak resistance. Experimental results show that SAD substantially reduces unsafe generations while preserving generation quality, diversity, and fluency—outperforming existing safety methods. The approach scales well and requires no architectural changes. With 27 pages and 12 figures, the paper provides a rigorous theoretical foundation and practical code release, making it a significant step toward safe deployment of text diffusion models in real-world applications.