DOT-MoE: Differentiable Optimal Transport for MoEfication
Researchers have introduced DOT-MoE, a new framework that converts dense large language models into sparse Mixture of Experts (MoEs) architectures. This method frames the decomposition of dense layers as a Differentiable Optimal Transport problem, using differentiable Sinkhorn-Knopp iterations to manage expert capacity and Straight-Through Estimators for end-to-end learning of neuron-to-expert assignments and token routing. Experiments show DOT-MoE outperforms existing methods, maintaining 90% of dense model performance while halving active parameters. AI
IMPACT Enables more efficient inference for large language models by converting dense architectures to sparse MoEs.