Toward Calibrated Mixture-of-Experts Under Distribution Shift

arXiv·medium signal

Wong, Prinster, and Saria study why Mixture-of-Experts models become miscalibrated when inputs drift from the training distribution and propose methods to realign predictive uncertainty with empirical outcome frequencies. As MoE has become the dominant architecture in frontier open-weight models, calibration under shift is a practical concern for anyone running MoE models in production with non-stationary inputs.