SurvICL: tabular foundation model achieves in-context survival analysis on clinical cohorts

Pretrained on synthetic datasets with a flexible proportional-odds prior and quantile regression loss, SurvICL outperforms 20 baselines and tops the 50-year-old Cox model on small cohorts.

Primary source: arXiv:2610.22242: https://arxiv.org/abs/2610.22242

What’s new: While tabular foundation models (TFMs) have overtaken gradient-boosted decision trees for tabular classification and regression, survival analysis on small clinical cohorts has remained dominated by the 1972 Cox proportional hazards model. Existing tabular foundation models and deep survival architectures consistently fail to beat Cox when sample sizes are limited and right-censoring obscures true event times. In SurvICL: a Foundation Model for in-context Survival Analysis (arXiv:2610.22242, 15 Oct 2026), researchers from Inria de Saclay, EPFL, and :probabl. present SurvICL, the first tabular foundation model designed specifically for in-context survival prediction.

SurvICL formulates survival estimation as in-context learning over patient feature tables. Rather than training on scarce or biased real-world clinical records, the model is pretrained entirely on synthetic datasets generated by a flexible proportional-odds prior that models diverse non-linear covariate interactions and baseline hazards. Because the synthetic generator produces uncensored ground-truth event times before synthetic censoring mechanisms are applied, SurvICL can be optimized directly with a plain quantile regression loss. At inference time, the model conditions on a context set of labeled, censored patient records and outputs calibrated survival curves for query patients in a single forward pass without gradient descent or hyperparameter tuning. Across the largest survival benchmark compiled to date (spanning 103 diverse datasets), SurvICL was evaluated on 63 datasets held out from all model design choices. SurvICL outperforms 20 competing methods, including Random Survival Forests, DeepHit, TabNet, TabPFN, and hyperparameter-tuned Cox models: winning 82.9% of pairwise comparisons in integrated Brier score (compared to 71.3% for standard Cox) and becoming the first machine learning model to systematically rank above Cox on small sample cohorts (fewer than 1,000 patients).

Why it matters: Prognostic risk scoring and treatment planning in clinical oncology and cardiology routinely rely on cohorts with only hundreds of patients, where high-capacity neural networks traditionally overfit and underperform classical statistical baselines. By demonstrating that pure synthetic pretraining with appropriate statistical priors enables robust zero-shot in-context survival modeling, SurvICL eliminates manual survival curve fitting and provides clinicians with immediate, calibrated prognostic curves without requiring local training pipelines.

Caveats: SurvICL is currently restricted to tabular baseline features and right-censored time-to-event outcomes; it does not process longitudinal time-varying covariates, competing risks, or high-dimensional multimodal imaging inputs. In-context inference requires feeding the entire training cohort through attention layers, which limits context size to approximately 1,000 to 2,000 patient records under standard GPU memory constraints. Finally, while the proportional-odds prior outperforms proportional-hazards and accelerated failure time priors on the benchmark, cohorts with complex non-proportional crossing hazards may still present residual calibration drift.