Prediction-powered Inference by Mixture of Experts

A new paper from researchers Yanwu Gu, Linglong Kong, and Dong Xia introduces a semi-supervised inference framework that leverages a mixture of experts (MOE) to improve prediction-powered inference (PPI). The core idea is to treat multiple AI prediction tools as a committee of experts, each with different architectures, training strategies, and strengths. Rather than relying on a single predictor, the framework dynamically weights the experts to minimize variance—the key source of uncertainty in PPI. This yields smaller confidence intervals and more reliable estimates, even when labeled data is scarce.

The framework is theoretically grounded, with the authors proving non-asymptotic upper bounds on coverage error for confidence intervals. It extends beyond simple mean estimation to linear regression, quantile estimation, and general M-estimation. Empirical tests validate the approach, showing that MOE-powered inference consistently outperforms standard PPI and achieves the best-expert guarantee—meaning it performs at least as well as the single best predictor in the mixture. For practitioners, this means more accurate statistical inference from limited labeled data, a critical capability in domains like healthcare, finance, and scientific research where labeling is expensive.