Scaling Amortized Inference to Large Observation Sets

A new method enables neural posterior estimation to scale to arbitrarily large observation sets without prohibitive memory or compute costs. The approach trains a mean-pool Deep Set encoder on sets of size at most two, then fine-tunes the inference head on pre-aggregated embeddings, making training cost essentially independent of deployment set size N. This decouples representation learning from posterior modeling. The technique is demonstrated across scalar, image, and multi-view 3D tasks. The paper is available on arXiv (2605.07972).