The Safety-Aware Denoiser for Text Diffusion Models

Authors: Amman Yusuf, Zhejun Jiang, Mijung Park

arXiv · PDF

Summary

The authors propose the Safety-Aware Denoiser (SAD), a safety-guidance method for text diffusion models that steers generation toward safe text during the iterative denoising process. Instead of post-hoc filtering or retraining the model, SAD modifies each denoising step at inference time to guide the final sample into provably safe regions of text space. The method is lightweight, avoids expensive retraining, and can flexibly integrate different safety constraints. Experiments show SAD substantially reduces unsafe generations across hazard taxonomy, memorization, and jailbreak benchmarks while preserving quality, diversity, and fluency.

Main takeaways:

• Existing safety methods (post-hoc filters, inference-time interventions) don't translate well from autoregressive models to diffusion-based text generation.
• SAD intervenes during the denoising loop itself, steering samples toward safe regions of text space without retraining the diffusion model.
• The method is inference-time only, so it's computationally cheap and flexible—you can swap in different safety constraints.
• Evaluations cover hazard taxonomy (toxic content categories), memorization (verbatim training data leakage), and jailbreak prompts.
• SAD outperforms existing baselines on safety metrics while maintaining generation quality, diversity, and fluency.

Relevance

Potentially relevant to my midtraining and installation work if I ever explore diffusion-based generation instead of autoregressive models, but that's not on my current roadmap. More broadly, the idea of steering generation at inference time (without retraining) parallels activation steering and prompt engineering in my installation-path equivalence project, though the mechanics are completely different (denoising vs. next-token logits).

Abstract

arXiv:2605.08116v1 Announce Type: new Abstract: Recent work on text diffusion models offers a promising alternative to autoregressive generation, but controlling their safety remains underexplored. Existing safety approaches are geared toward autoregressive models and typically rely on post-hoc filtering or inference-time interventions. These are inadequate for effectively addressing safety risks in text diffusion models. We propose the Safety-Aware Denoiser (SAD), a safety-guidance framework in text diffusion models. The SAD modifies the iterative denoising process such that the text sample at the final denoising step is steered toward provably safe regions of the text space. This inference-time method can integrate safety constraints into the denoiser, avoiding computationally expensive retraining of the underlying diffusion model and enabling flexible, lightweight safety guidance. We evaluate the safety of the generated text using the SAD, with respect to hazard taxonomy, memorization, and jailbreak. Experimental results show that SAD substantially reduces unsafe generations while preserving generation quality, diversity, and fluency, outperforming existing methods. These results demonstrate that our safety guidance during denoising provides an effective and scalable mechanism for enforcing safety in text diffusion models.