Evaluating the learnability of single-cell large language models on multiple tasks.

The rise of single-cell foundation models (scFMs) has sparked interest in their potential to unify diverse biological tasks. However, their practical utility and the validity of scaling laws-the assumption that performance improves with model and data size-remain under-examined. Here, we systematically evaluate two representative scFMs, Geneformer and scGPT, across perturbation prediction and cell type annotation tasks. Our findings suggest that the benefits of large-scale pretraining are strongly task-dependent, conferring substantial advantages in cell type annotation but limited gains in perturbation prediction. Furthermore, our results indicate that increasing model size does not guarantee improved performance and can even be detrimental, challenging the "bigger is better" paradigm for the models and tasks examined here. By comparing model performance on real versus synthetic data with different levels of complexity, our analysis suggests that for perturbation prediction, the tested scFMs may capture little more than simple summary statistics, suggesting limited capacity to learn complex biological interactions within our experimental design. Based on our evaluation of Geneformer and scGPT, these results highlight the need to move beyond scaling and toward developing models that integrate deeper biological knowledge. We suggest that a renewed focus on task-specific architectures and biologically-informed priors may be critical for unlocking the true potential of foundation models in single-cell biology.