Divergence-Based methods for Diffusion LLMs: Semantic Prompt Shifts, Unmasking, and Robustness
A. N. Vidyashankar${}^A$
${}^a$Department of Statistics, George Mason University, USA

Diffusion LLMs replace next-token generation with iterative denoising, creating new opportunities for refinement and control but also new challenges for robustness and uncertainty quantification. We develop a divergence-based approach to dLLMs that uses KL, Hellinger, total variation, and related divergences to track discrepancies across denoising steps and induced text distributions. The framework provides a mathematically clean way to study sensitivity, stability, and distributional shift in diffusion-based language generation.