E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models
E-MoE improves few-step non-factorized diffusion language models via Mixture-of-Experts routing as a discrete shared latent, boosting sample quality without extra active parameters.
Published Sep 29, 2026▲ 66 on Hugging FacearXiv ↗
Only vote on papers you've read. Sign in with GitHub to vote.
E-MoE's MoE routing acts as a discrete shared latent that breaks factorized diffusion without extra active parameters, though it remains unclear whether gains come from genuine correlation or routing noise, and continuous baselines are missing.
Abstract
Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored. We propose Enhanced Mixture-of-Experts (E-MoE), which builds the reverse process as a mixture of factorized distributions over a discrete shared latent given by the expert-routing decisions of a Mixture-of-Experts (MoE) backbone, without increasing active parameters over the factorized baseline. Across synthetic multi-modal benchmarks, binarized MNIST, and LM1B, E-MoE improves few-step generation over factorized baselines.