A new detection method wants to catch AI hallucinations mid-generation by treating a diffusion model's denoising steps like a stock ticker.
Diffusion language models are an emerging alternative to autoregressive systems that write text token by token, instead starting from a noisy draft and refining the whole sequence over repeated denoising steps. Earlier detection tools tracked how uncertain the model looked across steps or across tokens, but not both together, missing patterns like inconsistent convergence and errors spreading sideways between tokens. Researchers built DeMTS, which converts the entire denoising trajectory, every token at every step, into a multivariate time series using learnable latent variables, then models how those variables relate to each other and change over time to predict hallucinations. Tested on two diffusion model backbones across three benchmarks, DeMTS beat prior detection methods while staying efficient and transferable across tasks.
Hallucination detection is already messy for ordinary chatbots; diffusion models complicate it further because a bad token introduced early can corrupt tokens elsewhere in the sequence as the draft refines, not just the ones that come after it. Flattening the process into a single timeline or a single per-token score, which is what earlier methods did, throws away exactly the cross-token signal needed to catch that kind of error.
This is a benchmarks paper, not a shipping feature, so the real test is whether DeMTS holds up once someone bolts it onto a diffusion model people actually rely on.