PaperlayerView exact revision on arXiv

Source-linked AI summary

SurvivalPFN: Amortizing Survival Prediction via In-Context Bayesian Inference

Shi-ang Qi, Vahid Balazadeh, Michael Cooper, Russell Greiner, Rahul G. Krishnan

arXiv:2605.15488v1cs.LGstat.ML

TL;DR

Survival modeling often requires practitioners to choose among estimators with different assumptions and substantial methodological expertise. SurvivalPFN uses pretrained in-context Bayesian inference to predict from censored data in one forward pass; across 21 models, 61 datasets, and five metrics, it achieves the best median rank.

  • Problem

    Choosing among survival estimators with distinct assumptions requires substantial domain and methodological expertise.

    Source: §1 · p. 2

  • Method

    SurvivalPFN pretrains on diverse identifiable survival processes and uses a single forward pass to produce posterior predictive survival distributions from right-censored data.

    Source: §1 · p. 2

  • Results

    Across 21 models, 61 datasets, and five evaluation metrics, SurvivalPFN achieves the best median rank.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results suggest that PFN-style in-context learning is a promising foundation for flexible, computationally efficient survival prediction.

    Source: §6 · p. 9

    SurvivalPFN relies on conditional independent censoring and is not valid under dependent censoring; its relative performance also declines on larger tables.

    Source: §6 · p. 9

Abstract

from arXiv · show

Survival analysis provides a powerful statistical framework for modeling time-to-event outcomes in the presence of censoring. However, selecting an appropriate estimator from the many specialized survival approaches often requires substantial methodological and domain expertise. We introduce SurvivalPFN, a prior-data fitted network that amortizes Bayesian inference for censored observations through in-context learning. SurvivalPFN is pretrained on a diverse family of synthetic, identifiable, and right-censored data-generating processes, enabling it to amortize survival analysis in a single forward pass during inference. As a result, the model adapts to the effective complexity of each dataset without task-specific training or hyperparameter tuning, avoids restrictive parametric assumptions, and produces calibrated survival distributions. In a large-scale benchmark spanning 61 datasets, 21 methods, and 5 evaluation metrics, SurvivalPFN achieves strong predictive performance and often improves upon established survival models. These results suggest that SurvivalPFN offers a principled and practical foundation model for survival analysis, with potential applications in high-impact domains such as healthcare, finance, and engineering (https://github.com/rgklab/SurvivalPFN).

1 Introduction

Right-censored survival modeling offers many estimators with distinct assumptions, leaving model selection and validation to practitioners with substantial expertise. SurvivalPFN amortizes Bayesian prediction through pretraining and a single inference pass, and achieves the best median rank in a 21-model, 61-dataset, five-metric benchmark.

1 Introduction

  • Right-censoring leaves some event times unobserved, while estimators have distinct assumptions and model selection, training, and validation require substantial expertise.Source: §1 · pp. 1–2
  • SurvivalPFN pretrains on millions of simulated processes to infer conditional survival distributions from right-censored data in a single forward pass.It adapts to data complexity without task-specific training or hyperparameter tuning and avoids restrictive parametric assumptions.Source: §1 · p. 2
  • The authors provide theoretical justification for asymptotic consistency under identifiable right-censored data-generating processes.Source: §1 · p. 2
  • Across 21 models, 61 datasets, and five metrics, SurvivalPFN achieves the best median rank.Figure 1 describes its speed as matching classical models.Source: Figure 1 · p. 1, §1 · p. 2
  • The authors release SurvivalPFN training code and a scikit-learn-style API.Source: §1 · p. 2

2 Background

With right-censored data, only the minimum of event and censoring times and an event indicator are observed. The section defines when event-time survival is identifiable and motivates amortized Bayesian prediction as a way to avoid repeated, difficult posterior computation.

2 Background

  • The observed time is T = min(E, C), and ∆ indicates whether the event occurred; predictors target conditional event-time densities or survival functions.Source: §2 · p. 3
  • Identifiability means that any two processes inducing the same observed law over (X, T, ∆) also induce the same event-time survival function.Conditional independent censoring and positivity are stated as sufficient assumptions for nonparametric identification.Source: §2 · p. 3
  • When event and censoring times are dependent given X, the event distribution is generally not identifiable from observed data alone without additional assumptions, such as a copula.Source: §2 · p. 3
  • Bayesian prediction integrates over plausible survival mechanisms, but posterior predictive integrals are generally intractable and MCMC or variational inference must be rerun for each dataset.Source: §2 · p. 3
  • PFNs replace dataset-specific posterior inference with a transformer that maps context data and query inputs to predictive distributions in one forward pass.Training samples identifiable DGPs, pairs observed context with query covariates and indicators, and minimizes likelihood loss.Source: §2 · p. 3, Figure 3 · p. 4

3 Method

SurvivalPFN amortizes Bayesian posterior predictive inference for right-censored survival data, moving dataset-specific inference to offline pretraining. Its consistency target depends on identifiable DGPs, and the synthetic prior is designed to cover diverse mechanisms within that constraint.

3.1 SurvivalPFN: Amortized Posterior Predictive Inference

  • A simulator prior over identifiable right-censored DGPs supplies synthetic contexts and latent query-time targets, training SurvivalPFN to approximate Bayesian posterior predictive distributions.At inference, one forward pass uses an observed dataset and query covariates; no dataset-specific gradient updates, posterior sampling, variational optimization, or hyperparameter tuning are required.Source: §3.1 · pp. 4–5
  • Asymmetric attention and no positional encodings make predictions invariant to context order and conditionally independent across queries; 1,024-bin transformed-time histograms represent PPDs.A query indicator selects event-time or censoring-time prediction, while monotone time transforms accommodate nonnegative support and distribute resolution across the context time range.Source: §3.1 · pp. 4–5

3.2 Consistency of the Bayesian Posterior Predictive Target

  • Under a prior supported on identifiable right-censored mechanisms, the Bayesian PPSD converges almost surely to the true conditional survival function; the model inherits this guarantee if it exactly amortizes that distribution.The argument groups DGPs by observed-data equivalence: within each class, identifiable mechanisms induce the same conditional event-time survival function.Source: §3.2 · pp. 5–6

3.3 Prior over Identifiable Survival DGPs

  • The prior must exclude non-identifiable mechanisms while remaining broad, and randomized MLPs generate covariates and conditionally independent event and censoring times across varied synthetic datasets.The prior samples among uniform, random, administrative, and conditional independent censoring mechanisms.Source: §3.3 · p. 6

4 Experiments and Results

Across 61 held-out datasets, SurvivalPFN achieves the strongest overall benchmark rank, with metric-specific strengths and variation by dataset size. It is also competitive in computational cost, stable across training-set sizes, and ranks highly against general-purpose tabular foundation models.

Experimental Setup

  • The benchmark uses 61 held-out datasets, 21 survival models, five metrics, and repeated 70%/30% train/test splits; 20 additional datasets are reserved for checkpoint selection.The five metrics assess probabilistic accuracy, discrimination, calibration, median-time prediction, and agreement with observed outcomes.Source: §4 · p. 7

RQ1: Predictive Performance

  • SurvivalPFN’s largest advantage is on small datasets, its relative performance decreases as dataset size grows, and it remains consistently strong across censoring regimes.The authors relate the small-dataset advantage to using structure learned in prior-data pretraining when downstream observations are limited.Source: §4 · p. 7
  • SurvivalPFN achieves the strongest overall rank across the benchmark, leading on IBS, MAE, and Log-Rank while remaining competitive on CI and D-calibration.Several baselines rank higher in the CI and D-calibration metric-specific rankings.Source: §4 · p. 7, Figure 6 · p. 8

RQ2: Computational Efficiency

  • SurvivalPFN is only modestly slower than CoxNet, the fastest baseline, while achieving the best overall ranking across five evaluation metrics.Tree-based and neural-network-based methods require substantially greater computation.Source: §4 · p. 8

RQ3: Sensitivity to Training-Set Size

  • SurvivalPFN maintains stable performance across training-data split ratios, while CoxNet, GB, and CWGB perform poorly at the smallest ratio and improve markedly with more data.The reported SurvivalPFN pattern includes consistently low IBS and high CI even when only a small fraction of data is used for training.Source: §4 · p. 8

RQ4: Comparison with General TFMs

  • Against general-purpose tabular foundation models trained only on uncensored instances, SurvivalPFN ranks first overall and across all metrics.The authors conclude that directly adapting generic TFMs is insufficient for equally reliable survival prediction.Source: §4 · p. 8

RQ5: Ablation Studies

  • Appendix ablations examine prior design, query schedules, monotone transformations, and objective functions.The supplied passage reports the ablation topics but not their outcomes.Source: §4 · p. 8

5 Related Work

Bayesian survival methods quantify uncertainty but remain tied to method-specific assumptions or per-dataset inference. Concurrent tabular-foundation approaches differ in how they adapt to censoring, with SurvivalPFN using a broader prior over identifiable right-censored processes.

Concurrent Work on TFMs for Survival Analysis

  • Figure 8 compares SurvivalPFN with selected general-purpose tabular foundation models across 61 benchmark datasets, using Figure 6 plotting conventions.Source: Figure 8 · p. 9

Bayesian Survival Analysis

  • Bayesian neural survival methods use credible intervals, Bayesian updating, or variational inference, yet retain method-specific assumptions and per-dataset inference or updating.The examples include BNN-ISD, Bayesian LSTM-SURV, and NeuralSurv.Source: §5 · p. 9

Concurrent Work on TFMs for Survival Analysis

  • Kim et al. [46] reduce survival prediction to time-indexed binary tasks without survival-specific pretraining, while Seletkov et al. [91] pretrain on synthetic parametric extended-hazard mechanisms.The passage notes context-length and subsampling constraints for the former, and a parametric prior with random censoring for the latter; SurvivalPFN instead uses a broader identifiable-DGP prior and supports covariate-dependent censoring under conditional independence.Source: §5 · p. 9

6 Conclusions, Limitations, and Future Work

SurvivalPFN provides amortized Bayesian survival prediction from right-censored data and performs strongly across 61 real-world datasets. Its validity depends on conditional independent censoring, and its relative performance declines on larger tables; the section also recalls the paper’s notation convention.

6 Conclusions, Limitations, and Future Work

  • Across 61 real-world datasets, SurvivalPFN achieves strong overall performance while remaining computationally efficient.The authors describe PFN-style in-context learning as a promising foundation for flexible survival prediction.Source: §6 · p. 9
  • SurvivalPFN relies on conditional independent censoring; under dependent censoring, the event-time distribution is not nonparametrically identifiable from observed data alone, and the method is not valid.The authors identify identifiable dependent-censoring priors, such as copula methods, as a direction for extension.Source: §6 · p. 9
  • SurvivalPFN inherits a size-scalability trade-off, with reduced relative performance on larger tables.Improving long-context inference is left for future work.Source: §6 · p. 9

A Notation

  • The notation convention uses uppercase letters for random variables and lowercase letters for realizations.Table 1 summarizes the paper’s notation.Source: A Notation · p. 18, Table 1 · p. 18

B Identifiability and Non-identifiability under Right Censoring

Right-censored observations alone cannot distinguish some dependent-censoring models from independent-censoring models, so event-time survival is identifiable only under suitable assumptions. The appendix formalizes conditions for identifiability and states consistency of the Bayesian posterior predictive survival distribution under an identifiable prior.

B Identifiability and Non-identifiability under Right Censoring

  • Dependent and independent censoring models can generate identical observed data, but conditional independent censoring makes the event-time survival function identifiable.Without the independence assumption, the event-time distribution cannot be uniquely determined from observed data alone.Source: B Identifiability and… · pp. 19–20

C Posterior-Predictive Consistency

  • Under an identifiable survival prior, the Bayesian posterior predictive survival distribution is asymptotically consistent for the true conditional event-time survival function.The corresponding SurvivalPFN implication assumes the transformer exactly amortizes this Bayesian target.Source: C Posterior-Predictive Consistency · p. 20

C.1 Notation and regularity assumptions

  • The regularity assumptions are technical conditions that ensure priors, posteriors, conditional expectations, and the quotient construction are well-defined.The passage says they are satisfied by usual finite-dimensional parameter spaces and measurable simulators used in statistical and machine-learning models.Source: §C.1 · p. 21

C.2 Observed-law equivalence and survival identifiability

  • A prior is survival-identifiable at time t when observationally equivalent DGPs yield the same conditional event-time survival at t.Conditional independent censoring and positivity provide sufficient conditions on the identifiable time region.Source: §C.2 · p. 21

C.3 Consistency of the Bayesian PPSD

The Bayesian PPSD is consistent at time t if and only if the prior is survival-identifiable; without identifiability, its limit is an average over observationally equivalent DGPs. SurvivalPFN inherits this consistency only under exact amortization or vanishing amortization error, not as a finite-network guarantee.

C.3 Consistency of the Bayesian PPSD

  • At fixed time t, Bayesian PPSD consistency holds if and only if the prior is survival-identifiable, on full-measure sets of query covariates and parameters.The result also holds simultaneously over any finite or countable evaluation grid.Source: §C.3 · p. 22
  • Without survival-identifiability, the PPSD converges only to the prior-average survival function over DGPs observationally equivalent to the truth.The observed data distinguish equivalence classes, but not DGPs within a class.Source: §C.3 · p. 22

C.4 Implication for SurvivalPFN

  • The proposition does not establish finite-sample or finite-capacity consistency for the trained transformer; it gives consistency in the idealized exact-amortization limit.The proposition characterizes the Bayesian posterior predictive target that SurvivalPFN is trained to approximate.Source: §C.4 · p. 23
  • Under Proposition C.1 conditions, SurvivalPFN is consistent when it exactly amortizes the Bayesian target or its amortization error vanishes.The result follows by combining the amortization condition with consistency of the Bayesian PPSD.Source: §C.4 · p. 23

D.1 Synthetic Prior and DGP Simulation

This section describes SurvivalPFN’s synthetic prior and data-generating processes, then outlines its training and inference alongside the benchmark baselines, evaluation protocol, metrics, and compute environment. It also specifies how runs exceeding the 72-hour limit were handled.

D.1 Synthetic Prior and DGP Simulation

  • Each synthetic task draw specifies covariates, event-time and censoring-time mechanisms, censoring type, and target censoring rate; its latent times supply supervised query targets during training.A generic tabular generator can use sources such as random multilayer perceptrons, structural causal models, or trees.Source: §D.1 · p. 23
  • Censoring mechanisms include uniform, random tabular, administrative censoring from staggered entry, and event- and censoring-times generated independently conditional on covariates.Conditional independent censoring permits censoring to depend on covariates while preserving the standard assumption used for nonparametric identification.Source: §D.1 · p. 24
  • Four prior families provide irregular tabular times, smooth monotone distributions, Weibull/lognormal mixtures with heavy tails and multimodality, and a configurable mixture of these priors.The kitchen-sink default places most mass on direct table-output and monotone survival-distribution priors.Source: §D.1 · pp. 24–25
  • Across four priors, I(E; C | X) concentrates near zero while generated data span nearly all censoring rates and varied time dispersion, uncertainty, and survival-curve shapes.The diagnostics sampled 500 datasets per prior; curve summaries show different time scales, decay rates, and tail behaviors.Source: Figure 9 · p. 26, §D.1 · pp. 26–27, Figure 10 · p. 27

D.4 Training Objective

  • Training uses simulator-provided latent query times rather than the standard censored observed-data likelihood, and evaluates smoothed cross-entropy as an alternative to discrete NLL.The smoothed target is a narrow histogram in transformed-time space, implemented using a Gaussian smoothing density.Source: §D.4 · p. 30

D.5 Inference Procedure

  • At inference, SurvivalPFN produces a discretized posterior predictive event-time distribution in one forward pass, without dataset-specific gradient updates, posterior sampling, or hyperparameter tuning.The posterior predictive survival distribution is computed by summing probability mass in bins whose raw-time support exceeds the queried time.Source: §D.5 · p. 31

E.1 Baseline Model Details

  • The benchmark compares SurvivalPFN with 20 representative methods spanning classical statistical models, tree ensembles, discrete- and continuous-time neural survival models, and tabular foundation models.The authors describe baseline implementations and summarize their methodological properties in a qualitative comparison table.Source: §E.1 · p. 31, Table 2 · p. 32

E.2 Unified Time Grid for Consistent Evaluation

  • Methods are evaluated through a common survival-probability matrix, with time grids defined using training data only to avoid test-set leakage.Discrete-time models use event-time quantile bins, while continuous-time models are queried on empirical training-time support.Source: §E.2 · pp. 34–35

E.3 Hyperparameter Tuning

  • Hyperparameters are tuned only for neural-network baselines; selection uses 10 sampled configurations and shuffled 5-fold cross-validation on training-side data, with held-out testing reserved for final evaluation.Classical scikit-survival models, StaticSurvivalTFM, and SurvivalPFN remain fixed at benchmark settings.Source: §E.3 · p. 35

E.4 Evaluation Metrics

  • Evaluation includes Harrell’s concordance index [33], IPCW integrated Brier score, pseudo-observation MAE [80], and D-calibration [29].Higher CI indicates better discrimination; lower IBS, MAE, and D-calibration chi-square indicate better performance or less evidence against calibration.Source: §E.4 · pp. 36–37
  • A log-rank goodness-of-fit test assesses alignment between predicted event times and observed outcomes; a smaller statistic indicates closer agreement.Predicted median survival times are treated as uncensored event-time predictions.Source: §E.4 · p. 37

E.5 Compute Environment and Runtime Protocol

  • Benchmark experiments use the same fixed resource allocation across models and datasets: one NVIDIA L40S GPU, one CPU core, 16 GB system memory, and a 72-hour maximum.Jobs run independently on a Slurm-managed GPU cluster.Source: §E.5 · p. 37
  • The 72-hour budget covered fitting, enabled hyperparameter tuning, final refitting, test-set prediction, and metric computation; incomplete runs were ranked worst or tied for worst on that dataset.Source: §E.5 · p. 37

E.6 Performance Reporting Protocol

The protocol evaluates models across repeated train/validation/test splits, ranks them within each dataset and metric, and summarizes performance with median ranks and bootstrap confidence intervals. It also records benchmark coverage and model-specific failures, including why several methods did not produce valid results.

E.6 Performance Reporting Protocol

  • Models were evaluated over 10 repeated random splits, with 70% of each split assigned to training/validation and 30% to testing.Scores were reported as the mean and standard deviation across the 10 splits.Source: §E.6 · p. 38
  • Each model was ranked separately within each dataset and metric; invalid results received one rank below the worst completed method.Ties received the minimum tied rank.Source: §E.6 · p. 38
  • 28 of 1,281 model-dataset runs failed (2.19%); failures included invalid predictions, training limits, memory exhaustion, and nonconvergence.The benchmark evaluated 22 models on 61 datasets.Source: §E.6 · p. 38
  • CoxPH failed on 13 datasets because singular Hessians prevented the Newton–Raphson optimizer from computing valid updates.Source: §E.6 · p. 38
  • For each model and metric, performance was summarized by median rank across datasets, with uncertainty estimated using a nonparametric dataset bootstrap and 95% confidence intervals.Source: §E.6 · p. 38
  • Overall rank pooled ranks across evaluation metrics before applying the median-rank and bootstrap procedure; IBS and D-calibration were inapplicable to cSVR.cSVR produces scalar event-time predictions rather than a full survival distribution.Source: §E.6 · pp. 38–39

F Benchmark Dataset Details

The benchmark collection combines screened SurvSet data with additional datasets from textbooks, papers, and software packages. Twenty datasets support checkpoint selection, while 61 are held out for final evaluation.

F Benchmark Dataset Details

  • The collection contains 81 real-world survival datasets: 57 SurvSet datasets [15] after excluding longitudinal, competing-risk, and feature-count-exceeds-sample datasets, plus 24 additional datasets with at most 100,000 samples.Source: F Benchmark Dataset Details · p. 39
  • 20 datasets were reserved for validation and checkpoint selection, leaving 61 held out for the final benchmark without checkpoint-selection use.Validation splits used 70% of samples as context and 30% for inference; checkpoints were selected by weighted average integrated Brier score.Source: F Benchmark Dataset Details · p. 39
  • Validation datasets were selected to span sample size, censoring rate, and Kaplan-Meier tail survival probability.Source: F Benchmark Dataset Details · p. 39

G.1 RQ1: Predictive Performance

Across benchmark comparisons, SurvivalPFN remains among the strongest methods, with its clearest advantage on small datasets and a smaller edge as dataset size grows. It also performs strongly across censoring strata and against general tabular foundation models.

G.1 RQ1: Predictive Performance

  • On small datasets, SurvivalPFN is clearly separated from most baselines in overall rank and performs strongly across IBS, CI, D-calibration, MAE, and Log-Rank.Traditional statistical and tree-based methods form a second tier, while neural-network methods generally perform worst.Source: §G.1 · p. 42, Figure 11 · p. 43
  • As dataset size increases, SurvivalPFN remains strong on medium datasets but loses relative advantage on large datasets, where it stays competitive on MAE and Log-Rank.The authors attribute the size-regime trade-off to deep-learning estimators benefiting from more observations while SurvivalPFN faces finite context length and large-table summarization constraints.Source: §G.1 · p. 42, Figure 12 · p. 44, Figure 13 · p. 45
  • Across low-, medium-, and high-censoring subsets, SurvivalPFN remains in the top overall group and retains strong IBS, MAE, and Log-Rank performance.No single baseline matches its overall consistency across dataset strata, censoring strata, and metrics.Source: §G.1 · p. 42

G.2 RQ2: Computational Efficiency

  • The runtime aggregation excludes datasets with any failed run so each method is compared on the same completed datasets.This avoids making methods appear faster because they crashed, timed out, or failed to return valid predictions.Source: §G.2 · p. 43

G.3 RQ3: Sensitivity to Training-Set Size

  • Increasing the training ratio improves all methods across 16 additional datasets, with the sharpest IBS decreases and CI increases from very small to moderate context sizes.SurvivalPFN remains stable across ratios and is frequently among the best, especially in low-data regimes.Source: §G.3 · p. 43

G.4 RQ4: Compare with General Tabular Foundational Models

  • SurvivalPFN achieves the best aggregate rank against general tabular foundation models and ranks first or near-first across nearly all metrics.Its largest gains are on IBS and D-calibration; it also performs best on CI and log-rank.Source: §G.4 · p. 44
  • StaticSurvivalTFM performance depends on its backbone: MITRA improves over TabDPT, but neither wrapper matches SurvivalPFN across distributional metrics and log-rank.Direct TFM regression baselines are consistently worse despite their strength as general tabular predictors.Source: §G.4 · p. 44
  • The comparison supports survival-specific supervision and censoring-aware synthetic tasks over generic tabular in-context learning alone.Source: §G.4 · p. 44

G.5 RQ5: Ablation Studies

The ablations compare selected SurvivalPFN configurations across pretraining, prior, time transformation, loss, query schedule, and train/query ratio. The selected v01 checkpoint performs best overall among those evaluated.

G.5 RQ5: Ablation Studies

  • The selected v01 checkpoint achieves the best median rank across metrics among evaluated configurations and ranks best on IBS and D-calibration.It uses predictive pretraining, the survival-distribution prior, lognormal2normal transformation, one-hot NLL, the Both query schedule, and a fixed train/query ratio.Source: §G.5 · p. 46
  • The survival-distribution prior is consistently stronger than the naive prior, while the kitchen-sink prior does not clearly dominate it.The comparison suggests that matching the prior family to survival prediction matters beyond simply increasing prior diversity.Source: §G.5 · p. 46
  • Replacing lognormal2normal with time2quantile in v01 versus v02 substantially worsens overall rank and all five metric-specific ranks.The authors suggest the smooth positive-time coordinate and tail extrapolation of lognormal2normal help transfer across heterogeneous time scales.Source: §G.5 · p. 47
  • Predictive pretraining is generally helpful but not decisive in every evaluated configuration, and varying the train/query ratio shows no clear monotonic benefit.Source: §G.5 · p. 47

H Concurrent Works on Tabular Foundation Models for Survival Analysis

The concurrent approaches adapt general tabular foundation models to censored survival prediction through classification conversion or survival-specific in-context pretraining. Their priors, context requirements, and empirical scope differ from SurvivalPFN.

H Concurrent Works on Tabular Foundation Models for Survival Analysis

  • StaticSurvivalTFM converts survival prediction into time-indexed binary classification, treating labels after censoring as missing and using off-the-shelf tabular models.Under conditional independent censoring and positivity, population binary cross-entropy recovers survival probabilities at the discretization points.Source: H Concurrent Works on Tabular… · pp. 47–48
  • The classification conversion expands each subject into up to K−1 examples, so medium or large datasets may exceed current TFM input limits and require subsampling.It also depends on the time grid and requires post-hoc monotonicity correction of predicted survival curves.Source: H Concurrent Works on Tabular… · p. 48
  • SIC also pretrains an in-context model for survival, but uses a parametric extended-hazard prior and random censoring, unlike SurvivalPFN's broader identifiable right-censored DGP family.SurvivalPFN includes covariate-dependent censoring mechanisms satisfying conditional independence and avoids explicit parametric hazard or survival families.Source: H Concurrent Works on Tabular… · pp. 48, 50
  • SurvivalPFN has a posterior-predictive consistency guarantee for identifiable survival priors, whereas SIC provides empirical evidence only.Source: H Concurrent Works on Tabular… · p. 51
  • The study evaluates 61 held-out datasets, five metrics, and 21 baselines, but cannot directly compare with SIC because its weights and code are unreleased.Source: H Concurrent Works on Tabular… · p. 51
Loading 2605.15488v1…