Researchers develop tiny 10-30M parameter models with recursive structures to overcome biomedical AI's data scarcity bottleneck

New approach challenges the industry trend of building ever-larger foundation models by focusing on parameter-efficient small models instead

UCell uses recursive structures in its forward computation graph to achieve better parameter efficiency for single-cell segmentation tasks

Models contain only 10-30M parameters—tiny by modern AI standards—yet designed to perform competitively on biomedical vision tasks

Addresses the practical constraint in biomedical research where limited training data and high validation costs make model scaling difficult

Shifts focus from scaling up models to improving the capability of small models, an area previously under-explored in the field