SurvICL: a Foundation Model for in-context Survival Analysis

Julie Alberge; Linus Bleistein; Judith Abécassis; Marine Le Morvan; Gaël Varoquaux; David Holzmüller

stat.MLOct 15, 2026

SurvICL: a Foundation Model for in-context Survival Analysis

Authors: Julie Alberge, Linus Bleistein, Judith Abécassis, Marine Le Morvan, Gaël Varoquaux, David Holzmüller

Organizations: Inria de Saclay · EPFL · Inria de Saclay, :probabl.

Abstract

Tabular foundation models (TFMs) now outperform gradient-boosted trees in classification and regression. Yet in survival analysis, which predicts when an event occurs from censored data, often on small clinical cohorts, we find that no existing TFM beats the half-century-old Cox model. We address this issue by introducing SurvICL, a TFM that predicts survival curves in-context. SurvICL is pretrained on purely synthetic datasets generated from a novel flexible proportional-odds prior that outperforms other prior types in our ablation study. Unlike most other survival models, SurvICL is trained with a plain quantile regression loss, which is possible because its synthetic pre-training data generator can provide uncensored event times. Additionally, we introduce 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: it wins 82.9% of pairwise comparisons in integrated Brier score, against 71.3% for the runner-up, the Cox model, and is the first model to rank above Cox on small datasets.

Figures & tables

Figure 1: SurvICL is the best survival model on 63 held-out datasets. Pairwise win rate in integrated Brier score (IBS): the fraction of (dataset, competitor) pairs on which a method achieves a lower IBS. SurvICL wins 82.9% of its comparisons, against 71.3% for the runner-up, the classical Cox model. All existing survival TFMs trail behind the Cox model.

Figure 2: (a) Overview over our synthetic data prior: building on Qu et al. (2026) , we construct a pipeline to simulate survival datasets by generating an event time and censoring time distribution. (b) Three example datasets drawn from our prior, which illustrate the high diversity in observed event time and censoring time distributions.

Figure 3: Overview of (a) the held-out datasets and (b) a sample of 500 pre-training datasets. In (b), the sample size is replaced by the context (train) set size (see Section 5 ).

Panel of Figure 4 — Figure 4: (a, b) Critical difference diagrams of average ranks on all held-out datasets (lower is better), for IBS and C-index at the median horizon $t_{50}$ ; methods joined by a bar are not significantly different. (c) Ablations of prior and training choices, by min–max normalized IBS (higher is better).

Model	Loss	Continuous Time	Prior	Prediction Target
SurvPFN	NLL + ranking	✓	Weibull	Observed times
SurvivalPFN	NLL	✗	Random MLP	Latent event + censoring times
Survival-IC	DeepHit Loss	✗	EH Model	Observed times
TFM4Survival	BCE	✗	✗	Binary survival labels
SurvICL (ours)	Pinball Loss	✓	PO prior	Latent event times

Table 1: Major differences between existing survival TFMs and SurvICL. “Latent” targets are uncensored times generated by the prior, unobserved in real data.

Appendix figures & tables55 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone
Column-wise transformer	3 blocks, 8 heads, 128 inducing points, QSSMax attention
Row-wise transformer	3 blocks, 8 heads, 4 CLS tokens
In-context (ICL) transformer	12 blocks, 8 heads, QSSMax attention
Max. sequence length (rows)	1024
Max. features per dataset	100 (min. 1)
Regression head	Pinball / quantile, 999 quantile levels

Appendix

Table 2: Backbone architecture and prior for SurvICL.

Panel of Figure 5 — Figure 5: Six generated datasets using our final PO prior, incorporating administrative censoring and time discretization. For each sample, we show the histogram of observed event (left), the censoring events (center), and the Kaplan-Meier curve (right).

Type	Model Name	Reference
TFM	SurvICL	This work
	SurvPFN	Böhm et al. 2026
	SurvivalPFN	Qi et al. 2026
	Survival-IC	Seletkov et al. 2026
	TFM4Survival	Kim et al. 2026
Tree	SurvivalBoost	Alberge et al. 2024

Appendix

Table 3: Overview of methods we consider in our experimental section.

model	parameter	range
Others
Kaplan–Meier	no hyperparameters (population-level estimator)
Cox (lifelines, ridge-penalized)	penalizer	loguniform (1e-04, 10)
Tree-based methods
Random Survival Forest	max features	sqrt, log2, 0.3, 0.5, 1.0
	min samples leaf	randint (1, 30)

Appendix

Table 4: Hyper-parameter search space for every baseline in our survival benchmark. Each cell is searched with RandomizedSearchCV for 50 iterations on a single stratified 80/20 split of the seed-0 training fold, scored by the integrated Brier score (IBS). randint / loguniform denote the scipy.stats distributions; a list denotes a categorical choice; a lone value is held fixed (not searched).

Method	Dataset(s)	Cause
DeepSurv	divorce , actg , FRTCS	Training loss diverged (became non-finite) during optimization on these three small datasets.
TFM4Surival: TabICL	sac3 , DLBCL , gse1992	Out-of-memory issue on these larger/high-dimensional datasets.
SurvivalPFN	AML_Bull , gse1992 , vdv	The method’s internal dimensionality-reduction step (a truncated SVD) requires more components than the smaller of the fold’s row or feature count provides, and errors when that bound is violated.
TabSurv	DBCD , cesc tcga, kirc tcga, lgg tcga	The method’s obtained a MAE of value infty.

Appendix

Table 5: Baseline (method, dataset) pairs that created issues

name	source	n samples	n features	censoring frac	km survival at tmax
Framingham	survset	4699	18	68.7%	60.5 %
MCLcleaned	survset	92	574	30.4 %	0.0 %
Melanoma	survset	205	5	72.2 %	64.5 %
NSBCD	survset	115	549	67.0 %	57.4 %
Rossi	survset	432	29	73.6 %	73.6 %
Unemployment	survset	452	8	43.4 %	5.8 %

Appendix

Table 6: Overview of development datasets.

name	source	n samples	n features	censoring frac	km survival at tmax
AML Bull	survset	116	6283	42.2 %	28.3 %
Aids2	survset	2839	14	38.0 %	5.8 %
Bergamaschi	survset	82	10	65.9 %	58.3 %
DBCD	survset	295	4919	73.2 %	61.0 %
DLBCL	survset	240	7399	42.5 %	26.8 %
Dialysis	survset	6805	37	76.4 %	58.7 %

Appendix

Table 7: Overview of held-out datasets.

Backbone ( nanotabiclexsurvival )
Column-wise transformer	2 blocks, 4 heads, 128 inducing points, SSMax attention
Row-wise transformer	2 blocks, 4 heads, 4 CLS tokens
In-context (ICL) transformer	4 blocks, 4 heads, SSMax attention
Regression head	Pinball / quantile, 99 quantile levels
Target embedding	Sum
Duration rescaling (control)	Mean-duration

Appendix

Table 8: Backbone architecture shared by every arm.

Family	Arm(s)	Generative structure
Proportional hazards	cox	$H(t\mid x)=e^{\eta}H_{0}(t)$
Proportional odds	po	$O(t\mid x)=e^{\eta}O_{0}(t)$ , $O=(1-S)/S$
Transformation family	transform	$A_{\rho}(t\mid x)=e^{\eta}A_{\rho}(S_{0}(t))$ , $\rho\sim[0.05,3]$ per dataset (interpolates PH at $\rho\to 0$ and PO at $\rho=1$ )
Additive hazards	ah	$h(t\mid x)=h_{0}(t)+c\cdot\tanh(\eta)$ , clamped at 0
Accelerated hazards	ach	$h(t\mid x)=h_{0}(t\,e^{\eta})$ (curves leave $S{=}1$ together)
AFT, homoscedastic	aft-homo	$\log T=\mu(x)+(\log T_{0}-\mathbb{E}\log T_{0})$

Appendix

Table 9: Structural survival families compared on the family axis.

H

O

A

and

h

denote the cumulative hazard, cumulative odds, Aranda-Ordaz/Box-Cox transform, and hazard of

T\mid x

; a

0

subscript denotes the shared random baseline.

Axis	Control	Variant(s)
Discretization	Off	On (10% per-dataset chance of coarsening times to a grid)
Administrative censoring	Off	On (50% per-dataset chance of a study-end cutoff)
Label draws per row	1	5, 20
Quantile head width	99 levels	999 levels
Auxiliary censoring head	Event-only targets	$+$ censoring-distribution head (train-time only, dropped at inference)

Appendix

Table 10: Recipe axes ablated against the po control, one flag at a time.

Figure 6: IBS on all held-out datasets, compared to the IBS score obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 7: Critical difference diagram of average ranks on all held-out datasets (lower is better) for IBS. Methods joined by a bar are not significantly different

Figure 8: Pairwise win rate IBS all held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower IBS. The scores are the same as in Figure 1 , but shown with a different grouping.

Figure 9: Pareto front of mean normalized IBS (higher is better) and fit

+

predict time. We report average performances over all held-out datasets. IBS was normalized to provide comparability across all datasets.

Figure 10: C-index on the 50th quantile on all held-out datasets, compared to the score obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 11: Critical difference diagram of average ranks on all held-out datasets (lower is better) for the C-index on the 50th quantile. Methods joined by a bar are not significantly different.

Figure 12: Pairwise win rate in the C-index on the 50th quantile on all held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a higher C-index.

Figure 13: Pareto front of mean C-index on the 50th quantile (higher is better) and fit

+

predict time. We report average performances over all held-out datasets.

Figure 14: KM calibration on all held-out datasets (lower is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 15: Critical difference diagram of average ranks on all held-out datasets (lower is better) for the KM calibration. Methods joined by a bar are not significantly different.

Figure 16: Pairwise win rate in KM calibration on all held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a better KM calibration.

Figure 17: MAE on all held-out datasets, compared to the MAE obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 18: Critical difference diagram of average ranks on all held-out datasets (lower is better) for the MAE. Methods joined by a bar are not significantly different

Figure 19: Pairwise win rate in MAE on all held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower MAE.

Figure 20: Pareto front of mean MAE (lower is better) and fit

+

predict time. We report average performances over all held-out datasets.

Figure 21: IBS on the small held-out datasets, compared to the IBS score obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 22: Critical difference diagram of average ranks on the small held-out datasets (lower is better) for the IBS. Methods joined by a bar are not significantly different.

Figure 23: Pairwise win rate in IBS on the small held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower IBS.

Figure 24: Pareto front of mean normalized IBS (higher is better) and fit

+

predict time. We report average performances over the small held-out datasets. IBS was normalized to provide comparability across all datasets.

Figure 25: C-index at the 50th quantile on the small held-out datasets, compared to the C-index score obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 26: Critical difference diagram of average ranks on the small held-out datasets (lower is better) for the C-index at the 50th quantile. Methods joined by a bar are not significantly different.

Figure 27: Pairwise win rate in the C-index at the 50th quantile on the small held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a higher C-index.

Figure 28: Pareto front of mean C-index on the 50th quantile (higher is better) and fit

+

predict time. We report average performances over the small held-out datasets.

Figure 29: KM calibration on the small held-out datasets (lower is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 30: Critical difference diagram of average ranks on the small held-out datasets (lower is better) for the KM calibration. Methods joined by a bar are not significantly different.

Figure 31: Pairwise win rate in the KM-calibration on the small held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a better KM-calibration.

Figure 32: MAE on the small held-out datasets, compared to the MAE obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 33: Critical difference diagram of average ranks on the small held-out datasets (lower is better) for the MAE. Methods joined by a bar are not significantly different.

Figure 34: Pairwise win rate in the MAE on the small held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower MAE.

Figure 35: Pareto front of mean MAE (lower is better) and fit

+

predict time. We report average performances over the small held-out datasets.

Figure 36: IBS on the large held-out datasets, compared to the IBS obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 37: Critical difference diagram of average ranks on the large held-out datasets (lower is better) for the IBS. Methods joined by a bar are not significantly different.

Figure 38: Pairwise win rate in the IBS on the large held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower IBS.

Figure 39: Pareto front of mean normalized IBS (higher is better) and fit

+

predict time. We report average performances over the large held-out datasets. IBS was normalized to provide comparability across all datasets.

Figure 40: C-index at the 50th quantile on the large held-out datasets, compared to the C-index obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 41: Critical difference diagram of average ranks on the large held-out datasets (lower is better) for the C-index at the 50th quantile. Methods joined by a bar are not significantly different.

Figure 42: Pairwise win rate in the C-index at the 50th quantile on the large held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a higher C-index.

Figure 43: Pareto front of mean C-index on the 50th quantile (higher is better) and fit

+

predict time. We report average performances over the large held-out datasets.

Figure 44: KM calibration on the large held-out datasets (lower is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 45: Critical difference diagram of average ranks on the large held-out datasets (lower is better) for the KM calibration. Methods joined by a bar are not significantly different.

Figure 46: Pairwise win rate in the KM calibration on the large held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a better KM calibration.

Figure 47: MAE on the large held-out datasets, compared to the MAE obtained with a Kaplan-Meier estimator (higher is better). The points show individual dataset performances, and a boxplot summarizes the overall distribution. The dot color represents the dataset size.

Figure 48: Critical difference diagram of average ranks on the large held-out datasets (lower is better) for the MAE. Methods joined by a bar are not significantly different.

Figure 49: Pairwise win rate in the MAE on the large held-out datasets: the fraction of (dataset, competitor) pairs on which a method achieves a lower MAE.

Figure 50: Pareto front of mean MAE (lower is better) and fit

+

predict time. We report average performances over the large held-out datasets.

Survival Analysis

Survival analysis models when events such as disease progression, failure, or death occur, accounting for incomplete follow-up and censoring.

See all 106 papers

SurvPFN: Towards Foundation Models for Survival Predictions
Jun 3, 2026Samuel Böhm, Lennart Purucker, Frank Hutter +1Tabular Foundation Models Survival Analysis
Institute of Epidemiology and Prevention, Medical Center - University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany · Institute of Epidemiology and Prevention, Medical Center – University of Freiburg, Freiburg, Germany · Department of Computer Science, University of Freiburg, Freiburg, Germany +7
Survival In-Context: Amortized Bayesian Survival Analysis via Prior-Fitted Networks
Mar 31, 2026Dmitrii Seletkov, Paul Hager, Georgios Kaissis +3Survival Analysis Prior-Data Fitted Networks
Institute of Diagnostic and Interventional Radiology, Technical University of Munich, Germany · Chair for AI in Healthcare and Medicine, Technical University of Munich, Germany · Hasso Plattner Institute for Digital Engineering, University of Potsdam, Germany +3
Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation
Jun 10, 2026Minh-Khoi Pham, Luca Cotugno, Alina Sirbu +3Tabular Foundation Models Survival Analysis
ADAPT Centre, Dublin City University, Dublin, Ireland · School of Computing, Dublin City University, Dublin, Ireland · Department of Computer Science and Engineering, University of Bologna, Bologna, Italy

Jun 3, 2026J/K move · Enter open · S save

SurvPFN: Towards Foundation Models for Survival Predictions

Samuel Böhm, Lennart Purucker, Frank Hutter, Pascal Schlosser

Institute of Epidemiology and Prevention, Medical Center - University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany · Institute of Epidemiology and Prevention, Medical Center – University of Freiburg, Freiburg, Germany · Department of Computer Science, University of Freiburg, Freiburg, Germany+7

Tabular Foundation Models Survival Analysis Prior-Data Fitted Networks Mortality Prediction

Open paper AskPDF arXiv

Tabular foundation models (TFMs) have made rapid progress in standard classification and regression, but time-to-event survival prediction tasks have remained largely untouched. Unlike in standard regression tasks, survival prediction models must account for censored data. Standard TFMs cannot handle natively censored data, leading to biased and inaccurate predictions, making them unsuitable for real-world applications. To overcome this fundamental limitation, we propose SurvPFN, a prior-data fitted network (PFN), for survival prediction tasks. We pretrain SurvPFN on millions of synthetic survival prediction tasks to learn survival via distributional regression that accounts for censored data. SurvPFN works by (1) generating data with Weibull event times and a non-informative censoring mechanism; (2) integrating a censored event indicator; and (3) minimizing a censored negative log-likelihood. On SurvSet, a collection of real-world survival tasks, SurvPFN is highly competitive with classical and deep survival baselines without per-dataset fitting, a survival-specific architecture, or feature engineering. We show that survival can be treated as a continuous-time distributional regression problem with censored loss, unlocking the power of PFNs for time-to-event predictions.

Abstract

Survival Analysis

Explore similar work

SurvPFN: Towards Foundation Models for Survival Predictions

Survival In-Context: Amortized Bayesian Survival Analysis via Prior-Fitted Networks

Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation