Multiple-Prediction-Powered Inference

View PDF

Abstract:Statistical estimation often involves tradeoffs between expensive, high-quality measurements and a variety of lower-quality proxies. We introduce Multiple-Prediction-Powered Inference (MultiPPI): a general framework for constructing statistically efficient estimates by optimally allocating resources across these diverse data sources. This work provides theoretical guarantees about the minimax optimality, finite-sample performance, and asymptotic normality of the MultiPPI estimator. Through experiments across three diverse large language model (LLM) evaluation scenarios, we show that MultiPPI consistently achieves lower estimation error than existing baselines. This advantage stems from its budget-adaptive allocation strategy, which strategically combines subsets of models by learning their complex cost and correlation structures.

Comments: ICLR 2026, 45 pages, 17 figures

Subjects:

Statistics Theory (math.ST); Artificial Intelligence (cs.AI)

ACM classes: G.3

Cite as: arXiv:2603.27414 [math.ST]

(or arXiv:2603.27414v1 [math.ST] for this version)

https://doi.org/10.48550/arXiv.2603.27414

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Charlie Cowen-Breen [view email] [v1] Sat, 28 Mar 2026 21:24:00 UTC (2,043 KB)