Reproduce “Beyond Token Positions: Safety Alignment Across Denoising Steps in Diffusion Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guoli Wang, Haonan Shi, Tu Ouyang, An Wang
Researchers at Case Western Reserve University developed a step-wise and position-wise measurement framework to analyze safety alignment in Diffusion Language Models (dLLMs), identifying a "shallow-step alignment" where early denoising commitments significantly influence final safety. A training-free method, Refusal-Aware Early Commitment (RAEC), was introduced to leverage these insights, reducing the Attack Success Rate (ASR) on harmful prompts by up to 100% (e.g., Dream-v0-Instruct-7B ASR on StrongREJECT from 1.92% to 0.00%) while preserving utility.
Key insights
- Refusal signals are predominantly predicted and emerge within the first 8% of denoising steps and the first 2-8% of response positions, indicating early manifestation of safety-related behavior.
- A "shallow-step alignment" exists in dLLMs, where commitments made in early denoising steps, particularly in leading response positions, profoundly determine the final safety of the output.
- Unsafe generations often contain transient refusal signals, but these signals lack the strength and persistence observed in safe generations, leading to their eventual overwrite by non-refusal content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.