Search papers, labs, and topics across Lattice.
Masked diffusion models (MDMs) suffer from positional factorization in their reverse process, hindering few-step generation quality. This work introduces Enhanced Mixture-of-Experts (E-MoE), which conditions a mixture of factorized distributions on a discrete shared latent derived from MoE routing decisions, avoiding posterior collapse and parameter increase. E-MoE achieves improved few-step generation across synthetic, binarized MNIST, and LM1B benchmarks compared to factorized baselines.
Non-factorized diffusion language models achieve superior few-step generation by leveraging a discrete latent space learned via Mixture-of-Experts routing, bypassing posterior collapse issues.
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.
