A Complementarity-Enhanced Mixture of Human-AI Teams for Decision-Making

Abstract

With the rapid development of deep learning, Artificial Intelligence (AI) has evolved from a mere tool to a collaborator in decision-making, sparking increasing attention to the human-AI cooperation. The Mixture of Experts (MoE) framework, originally proposed to capture domain-specific expertise and now widely adopted in large-scale models, naturally aligns with the requirements of human-AI teams. However, deploying MoE in human-AI cooperation involves two challenges: 1) While machine experts can be continuously optimized during training, human experts remain fixed, significantly reducing the effectiveness of traditional sparse activation strategies; 2) Some existing methods fuse all expert predictions during training phase but select only the highest weighted expert during testing phase, thereby introducing inconsistencies between the two phases. To overcome this, we propose the Complementarity-Enhanced Mixture of Human-AI Teams (CE-MoHAIT) framework. Our approach decomposes the gating network’s output into two branches, i.e., a human expert branch and a classifier branch, thereby explicitly modeling the complementarity between human and AI capabilities. Moreover, we introduce a method called Adaptive and Complementary Construction (ACC) that directly optimizes the gating network by constructing weighted labels, enabling the classifier model to compensate for the deficiencies of human experts and ensuring consistent task allocation across training and testing. Experiments on CIFAR-100 and two real-world medical image datasets show that our approach surpasses the existing methods, improving test accuracy by up to 20%, especially with larger teams and weaker experts. Code is available in the repository at https://github.com/H-F-Liang/CE-MoHAIT .