AI RESEARCH
Safer by Diffusion, Broken by Context: Diffusion LLM's Safety Blessing and Its Failure Mode
arXiv CS.LG
•
ArXi:2602.00388v2 Announce Type: replace Diffusion large language models (D-LLMs) offer an alternative to autoregressive LLMs (AR-LLMs) and have nstrated advantages in generation efficiency. Beyond the utility benefits, we argue that D-LLMs exhibit a previously underexplored safety blessing: their diffusion-style generation confers intrinsic robustness against jailbreak attacks originally designed for AR-LLMs.