Explorative Modeling (XM) is a new paradigm for handling multimodality in generative models: instead of factoring the generation procedure into smaller steps during training (the trick every existing scalable approach relies on), a model explores many candidate generations and reinforces only the best-matching one, so its predictions capture individual modes instead of blurring them. This is useful in two distinct ways. First, exploration is a new scaling axis for existing generative models, yielding up to 53% better FLOP efficiency, 2.5× better data efficiency, and 47% better parameter efficiency—and, like data and parameters, these gains grow monotonically with scale. Second, as a standalone model, XM is the first reconstructive generative model to succeed with end-to-end training, achieving 20–250× inference speedups while matching diffusion on certain tasks.
@misc{gladstone2026explorativemodeling,
title={Explorative Modeling: Unlocking A New Scaling Axis and End-to-End Generative Modeling},
author={Alexi Gladstone and Heng Ji and Yilun Du},
year={2026},
note={Preprint. Under review.}
}