dMoE: dLLMs with Learnable Block Experts

Researchers introduced dMoE, a block-level mixture-of-experts method for diffusion large language models.

Instead of routing experts per token, it aggregates token-level preferences into one decision per block.

This reduces the number of activated experts, cuts memory use, and improves end-to-end latency.

The method preserves nearly all baseline performance while making inference more efficient.