DeMTS: Denoising Trajectories as Multivariate Time Series for Hallucination Detection in Diffusion Language Models

DeMTS treats diffusion language model denoising trajectories as multivariate time series to capture two-dimensional token-step dependencies. By preserving the full structure instead of compressing dimensions, the framework identifies hallucination signals like inconsistent convergence and cross-token fault propagation that traditional temporal or token-wise methods overlook.