Diffusion models for discrete data typically face a choice: learn an unnormalized score, which is good for generating samples but not for assigning probabilities, or use a fully normalized model, which supports likelihood evaluation but may be restrictive. The new paper attempts to combine the strengths of both through an evolving variational autoregressive network.
The approach builds a variational model that changes over the diffusion steps, approximating the reverse process in a normalized way. This allows the model to produce samples and directly evaluate likelihoods, a capability that standard score-based methods lack.
The abstract mentions a recent tensor-network approach as prior related work, but the specifics of how the present method differs are not fully detailed in the available excerpt. The authors plan to archive the full paper under arXiv:2609.27306, where the technical details will be available.