Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models
This paper measures dLLM safety behavior under harmful prompts by tracing intermediate token distributions and commitment decisions throughout denoising, and proposes Refusal-Aware Early Commitment (RAEC), a simple training-free decoding method that commits persistent refusal signals from early steps.
Guoli Wang, Haonan Shi, Tu Ouyang et al.
· 0 citations