Can a foundation model beat Cox regression on small survival datasets?
SurvICL ranked first across 63 held-out datasets without dataset-specific tuning. The unreviewed benchmark is broad, but it does not test clinical decisions, competing risks or changing patient data.
The short version
- What happened
- The benchmark covered 103 real-world datasets: 40 for development and 63 held out from all design decisions.
- Who it affects
- A more reliable small-cohort survival model could help researchers use scarce follow-up data, but an unvalidated probability could also distort counselling or treatment choices.
- What remains unproven
- Unreviewed preprint; benchmark rather than clinical deployment; single-event right-censoring only; no time-varying covariates or treatment-effect test; code and weights were not yet released.
Keep in mind
Unreviewed preprint; benchmark rather than clinical deployment; single-event right-censoring only; no time-varying covariates or treatment-effect test; code and weights were not yet released.
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
A modern model finally moved ahead of a 50-year-old baseline
A foundation model trained on synthetic survival data ranked ahead of 20 established alternatives across 63 real datasets that were kept out of every design decision. SurvICL won 82.9% of pairwise comparisons on integrated Brier score, which tests the accuracy of predicted survival probabilities over time. The runner-up Cox proportional-hazards model won 71.3%. On a discrimination measure, SurvICL won 77.7% of comparisons, against 71.6% for random survival forests. The result matters because Cox regression has remained unusually hard to beat on small time-to-event datasets.
The full benchmark contained 103 datasets from healthcare, genomics, economics and engineering, with up to 52,000 rows and 15,000 features. Forty were used during development; the reported comparison used the other 63. More than 70% of the collected datasets had fewer than 1,000 samples. For each dataset, methods were tested over ten event-stratified 80/20 splits on four metrics. SurvICL used up to 4,096 training rows as context and made predictions without fine-tuning or dataset-specific hyperparameter searches.[1]
Synthetic training taught the model what censored records omit
Survival records are incomplete by design. If a person leaves a study alive after three years, their eventual event time is unknown; the record only establishes that it is later than three years. SurvICL was pretrained on simulated datasets where the hidden event times were available to the trainer, even though the model's context contained censored observations. It learned to output a continuous set of event-time quantiles, which the researchers converted into a monotonic survival curve.
That architecture addresses a real statistical problem, but the study remains a benchmark rather than a clinical trial. The datasets mix very different outcomes and populations, and a win rate counts whether one method beats another—not the size or medical importance of the gap. The authors also report a boundary to the result: gradient boosting retained a slight edge on the largest datasets. The classical Cox model therefore remains a serious comparator, not an obsolete method displaced in every setting.[1]
The distinction that matters
How the evidence was separated
The researchers kept model design apart from the final comparison, then tested every method repeatedly on unseen rows.
Develop
Choose the model on 40 datasets
No final scores taken here
Hold out
Freeze 63 other datasets
Never used for design decisions
Compare
Run ten 80/20 splits and four metrics
SurvICL led overall
Better forecasts are not yet better decisions
The preprint covers right-censored, single-event tabular data. It does not handle competing events such as death from different causes, covariates that change through time, images, free text or treatment effects. Every baseline used its hyperparameters selected on the first split for the remaining splits, while SurvICL needed no tuning; that pragmatic protocol may not represent the strongest possible tuning for every competitor. The promised code and model were not yet available with the paper, limiting independent reproduction today.
Public and university programmes in Switzerland, France and the EU supported the work, and Probabl provided Claude API access. No separate competing-interest statement appears in the accessible manuscript. For a clinician or researcher, the immediate value is a candidate method to compare during local validation—not a survival estimate to place directly in front of a patient. Confidence would rise after peer review, open code and weights, independent replication, prospective decision studies and extensions to competing risks and changing patient histories.[1]