Reasoning Across Space: Tiny Recursive Models for Spatial Omics

Spatial omics has enabled the study of gene expression within intact tissue, but existing models are limited by local processing, scarce annotations, and poor generalization across tissue sections. Current approaches perform spatial reasoning in a single pass over local regions, restricting the integration of long-range dependencies that underlie many biological signals. In this work, we investigate whether recursive computation provides a natural inductive bias for spatial omics, enabling global reasoning over entire tissue samples without test-time training. We apply Tiny Recursive Models (TRM) for spatial omics tasks and show that it consistently outperforms parameter-matched supervised baselines across robust validation splits. TRM also surpasses several slice-specific unsupervised methods in domain identification and achieves state-of-the-art performance among supervised methods in gene expression prediction. Further analysis reveals that recursion enables test-time scaling and adaptive computation, and that the model is partially generalizable across tissue types for domain identification. Overall, our results demonstrate that recursive computation provides a scalable, data-efficient framework for spatial omics, overcoming the locality constraints of existing approaches without sacrificing generalization.