SADDER: Safety-Aware Denoising Defense Method for Text Diffusion Models
Start here
An arXiv paper released on October 13, 2026 proposes SADDER, improving the training-free safety guidance framework Safety-Aware Denoiser (SAD): it extracts safety signals per token position in embedding space, gates the SAD update rule accordingly, prioritizes correcting unsafe tokens at inference time, and derives the per-position target distribution exactly corresponding to the corrected denoiser. The method is evaluated in unsafe continuation and jailbreak settings for Masked Diffusion Language Models and Uniform-State Diffusion Models.
Generated by AI from reports · Updated 29 days ago
Attention and developments
Attention, hourly
Now 23, 0% in 24 hours·Peak 23 (Nov 6 03:00)
Related research
This event's paper
- The Tokens Strike Back: Safety-Aware Denoising with Directional Embedding-based Revision
arXiv · 2026-10-13 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesSADDER: Safety-aware denoising and directional embedding correction for text diffusion models
SADDER improves the training-free safety guidance framework Safety-Aware Denoiser (SAD): it extracts a safety signal for each token position in embedding space, uses it to gate the SAD update rule, prioritizes correcting unsafe tokens during inference, and derives the position-wise target distribution that the corrected denoiser exactly corresponds to. In unsafe continuation and jailbreak settings for Masked Diffusion Language Models and Uniform-State Diffusion Models, this method achieves a better balance among safety, generation quality, and inference cost than existing methods.