UIUC & Harvard propose Explorative Modeling: a new scaling axis for generative models

机器之心 · wechat · 2026-08-01

Researchers from UIUC and Harvard introduce Explorative Modeling (XM), a new paradigm addressing exposure bias and mode blurring in generative models. The key idea: at each training step, generate K candidates and train only on the one closest to real data, thereby boosting generative expressivity without decomposing the generation process. Experiments show consistent improvements across image, video, and language modalities, with gains scaling: from 7% to 36% as data grows, 13% to 23% as model size increases, and efficiency gains doubling when compute triples. XM improves FLOP efficiency by 4.1x, sample efficiency by 6.2x, and parameter efficiency by 47%. In robot control, end-to-end XM matches DiffusionPolicy with 100 forward passes using just one. The authors argue exploration becomes a new scaling axis as models grow.

Related event: Explorative Modeling Introduces Third Pretraining Axis, Questioned as Reinvented avataRL(10 posts)→

Original post →

More from Models

Models channel →