Survival analysis predicts when an event occurs from censored data, often in small clinical cohorts. The authors of a new arXiv paper note that tabular foundation models now beat gradient-boosted trees in classification and regression, yet no existing such model outperforms the half-century-old Cox model in survival analysis. To close that gap, they introduce SurvICL, a foundation model that predicts survival curves in-context.

SurvICL is pretrained entirely on synthetic datasets generated from a novel flexible proportional-odds prior, which the authors say outperforms other prior types in their ablation study. Because the synthetic generator can provide uncensored event times, SurvICL can be trained with a plain quantile regression loss, unlike most other survival models.

The paper also introduces what it calls the largest survival benchmark to date, spanning 103 datasets. On the 63 datasets held out from all design decisions, SurvICL outperforms 20 tree-based, deep, and foundation-model baselines without any tuning, winning 82.9% of pairwise comparisons in integrated Brier score versus 71.3% for the runner-up Cox model. It is, the authors report, the first model to rank above Cox on small datasets.