MoE-Based Model Fusion for Efficient Pareto Set Approximation

This method addresses multi-objective optimization in large neural networks by using a Mixture of Experts (MoE) based model fusion approach. It provides a scalable alternative to computationally expensive evolutionary or scalarization algorithms for learning Pareto fronts.