#37 · ML / AI methods

Single-cell foundation model gains depend on pretraining data and evaluation design

The dominant thread is that evaluation design, not only corpus or model scale, shapes what appears as a data-scaling effect. A systematic study of 400 single-cell foundation models pretrained on a 22.2M-cell corpus reports that performance plateaus early and that no clear data scaling laws emerge, with pretraining dataset size and diversity both examined. A companion preprint argues that standard finetuning conceals pretraining-scale gains, and that fully frozen and zero-shot evaluation reveals data-scaling dependence. Together these suggest that conclusions about dataset size depend on the evaluation regime. A separate preprint on biomedical claim verification is thematically adjacent: it reports that a SciFact artifact inflates scores and uses cross-domain tests to probe generalization. What is new is the explicit linkage of evaluation protocol to scaling conclusions in single-cell models.