PaperlayerView exact revision on arXiv

Source-linked AI summary

Tabular Foundation Models Can Do Survival Analysis

Da In Kim, Wei Siang Lai, Kelly W. Zhang

arXiv:2601.22259v2cs.LG

TL;DR

Right-censoring makes survival analysis difficult for tabular foundation models; this paper recasts static and dynamic survival analysis as binary classification of cumulative failure probabilities. Across real-world datasets, the classification-based models outperform baselines on average over survival metrics.

  • Problem

    Adapting tabular foundation models to survival analysis remains non-trivial because right-censoring challenges standard regression architectures, despite TFMs’ success in classification and regression.

    Source: §1 · pp. 1–2

  • Method

    The paper discretizes event times and casts static and dynamic survival analysis as binary classification of cumulative failure probabilities, enabling pretrained TFMs to use in-context learning without explicit training.

    Source: §1 · p. 2

  • Results

    Across 43 static datasets, classification-based models consistently outperform baselines, while MITRA-CF is best overall dynamically; the top 30% of datasets show 5–9% improvement over the best non-TFM baseline.

    Source: §5.2 · pp. 10–11, §5.3 · p. 12

  • Takeaways & Limitations

    The results support using pretrained TFMs for static and dynamic survival analysis without explicit model training.

    Source: §1 · p. 2, §5.2 · p. 10, §5.3 · p. 12

    Discretization loses temporal resolution, and the censored-label treatment is theoretically justified only under conditionally independent censoring; informative censoring and competing risks are out of scope.

    Source: §6 · p. 14

Abstract

from arXiv · show

While tabular foundation models have achieved remarkable success in classification and regression, adapting them to model time-to-event outcomes for survival analysis is non-trivial due to right-censoring, where data observations may end before the event of interest occurs. We utilize a classification-based framework that reformulates both static and dynamic survival analysis as a series of binary classification problems by discretizing event times. Censored observations are naturally handled as examples with missing labels at certain time points. This classification formulation enables existing tabular foundation models (TFMs) to perform survival analysis through in-context learning without explicit training. In contrast to classical approaches that use binary classifiers to model discrete-time hazards, our approach directly models cumulative failure probabilities, which we find empirically to be more robust to the number of discretization bins by avoiding multiplicative accumulation of per-bin errors. We prove that under standard censoring assumptions, minimizing our binary classification loss recovers the true survival probabilities as the training set size increases. We demonstrate through evaluation across 48 real-world datasets (43 static and 5 dynamic) that off-the-shelf TFMs with this classification formulation outperform classical and deep learning baselines on average over multiple survival metrics.

1. Introduction

The paper reformulates static and dynamic survival analysis as binary classification of cumulative failure probabilities, allowing tabular foundation models to use in-context learning without additional training. It reports that this formulation outperforms competing methods on average and is consistent under conditional independent censoring as sample size grows.

1. Introduction

  • Survival analysis is challenging because event times may be right-censored, datasets may be small, and covariates may be irregularly sampled; general-purpose practice uses models fitted from scratch per dataset.Right-censoring means the event time may be unknown when observation ends before the event occurs.Source: §1 · p. 1

(1) Classification formulation of survival analysis that models cumulative failure probabilities

  • We discretize event times and model cumulative failure probabilities as binary classification, enabling static and dynamic TFM survival prediction through in-context learning without additional training or tuning.Changing the number or placement of discretization bins does not require retraining the model.Source: §1 · p. 2, (1) Classification formulation of… · p. 2
  • Under conditional independent censoring, minimizing binary cross-entropy recovers true survival probabilities as sample size grows; lower loss also empirically corresponds to more accurate probabilities.The empirical relationship is shown in Figure 3.Source: (1) Classification formulation of… · p. 2
  • Across 43 static and 5 dynamic SurvSet datasets, TFMs with our formulation outperform competing TFM formulations and classical and deep survival models on average across multiple metrics.The comparison includes hazard-based and multi-class TFM formulations.Source: (1) Classification formulation of… · p. 2

2. Problem statement

The paper frames survival analysis in static and dynamic settings: static models use baseline features, while dynamic models update predictions as information arrives [Chen et al., 2024, Van Houwelingen and Putter, 2011].

2.1. Static setting

  • Static survival analysis estimates the probability that an event occurs after time t from baseline covariates, with event and censoring times observed through an event indicator.The target survival probability decreases from 1 at t = 0 to 0 as t approaches infinity; conditional independence of event and censoring times given baseline covariates ensures identifiability.Source: §2.1 · p. 3

2.2. Dynamic setting

  • Dynamic survival analysis uses covariate histories observed over time, potentially at irregular intervals, to estimate survival over a future horizon given information available at time t.Its conditional-independence censoring assumption conditions on the history and survival beyond t, ensuring dynamic survival probabilities are identifiable.Source: §2.2 · p. 3

3. Related work

Prior work spans established static survival methods, deep and discrete-time models, and several approaches to dynamic survival analysis. This section positions the paper’s cumulative-failure classification formulation against hazard-based classification, related discrete-time models, and concurrent survival-specific foundation models.

3.1. Overview of static and dynamic survival modeling approaches

  • Static survival work includes CoxPH and Random Survival Forests, deep extensions of parametric models, and discrete-time models; dynamic approaches include landmarking, sequence or joint models, and discrete-time models.The cited sequence and joint approaches include linear mixed-effects models with CoxPH, recurrent neural networks, and Neural ODEs for irregularly sampled data.Source: §3.1 · pp. 3–4

3.2. Classification models for survival analysis

  • Survival stacking models discrete hazards with binary classifiers, extending logistic regression approaches and recovering survival probabilities through a product of hazard-based terms.The passage notes that the observed-data likelihood includes censored and uncensored examples.Source: §3.2 · p. 4
  • With off-the-shelf TFMs, discrete-hazard survival stacking is competitive at small K but often degrades as K increases, motivating direct prediction of cumulative failure probabilities.The authors hypothesize that multiplying estimated per-bin hazards accumulates small errors over time.Source: §3.2 · p. 4
  • Yu et al. [2011] use K binary outcomes and multi-task logistic regression with cross-time dependencies, whereas this work uses binary cross-entropy with off-the-shelf TFMs through in-context learning.Yu et al. [2011] evaluated three large static datasets; this work studies static and dynamic survival across a broader collection of real-world datasets.Source: §3.2 · p. 4
  • DeepHit and Dynamic DeepHit significantly underperform the TFM-based approach in simulations; the authors hypothesize that training recurrent networks from scratch is less data efficient on their smaller datasets.The passage contrasts the models’ original evaluation on 800-10, 000 samples with a median dataset size of ≈500 in this work.Source: §3.2 · p. 5

3.3. Concurrent work on in-context learning models for survival analysis

  • SIC [Seletkov et al., 2026] and SurvivalPFN [Qi et al., 2026] pretrain survival-specific TFMs, while this work elicits survival analysis from general-purpose TFMs without survival-specific pretraining, fine-tuning, or architectural changes.SIC uses K = 10 bins and SurvivalPFN K = 1024; unlike those static-focused models, this method can vary K by reprocessing data and extend to dynamic analysis by re-querying at landmark times. The authors report that they could not directly compare these models because they were not publicly available.Source: §3.3 · p. 5

4. Methodology: Time-to-event modeling as supervised learning

The method formulates static and dynamic survival prediction as binary classification over discretized event horizons, with censored labels omitted and tabular foundation models usable through in-context learning.

4. Methodology: Time-to-event modeling as supervised learning

  • Time-to-event modeling becomes binary classification over discretized horizons, allowing off-the-shelf tabular foundation models to perform in-context learning without survival-specific training or tuning.The time boundaries use training-set observed-event-time quantiles to balance events across intervals; censoring-aware Kaplan–Meier quantiles yield nearly identical average performance.Source: §4 · pp. 5–6

4.1. Static setting

  • In the static setting, each subject becomes binary labels for whether the event occurred by each horizon, with labels available only before censoring and binary cross-entropy fit on observed labels.At inference, survival estimates are formed from the classifier’s predictions and clipped to remain monotonically decreasing over time.Source: §4.1 · pp. 6–7
  • Multi-class classification, including SurvivalPFN [Qi et al., 2026], significantly underperformed cumulative-failure prediction in the reported empirical comparison.The authors hypothesize that cumulative-failure prediction is easier to learn from finite data because it decomposes survival prediction into binary horizon-specific tasks.Source: §4 · p. 5, §4.1 · p. 7

4.2. Dynamic setting

  • In the dynamic setting, classifiers use covariate histories and survival through the current time to predict later event horizons, retaining examples only before censoring and while the subject remains at risk.Inputs across history lengths are featurized into fixed-dimensional vectors, and predicted survival probabilities are clipped to be nonincreasing over time.Source: §4.2 · pp. 7–8

4.3. Theory: Consistent survival models via loss minimization

  • Under conditionally independent censoring, minimizing the static and dynamic binary losses recovers true survival probabilities as training data grows, subject to positivity for the static setting.For static predictions, the population loss is minimized exactly when the classifier models the conditional cumulative event probability at each horizon.Source: §4.3 · p. 8

5. Experimental results

Across 43 static and 5 dynamic datasets, classification-based survival models performed strongly, with MITRA-CF leading overall in both settings. Direct cumulative-failure modeling outperformed discrete-hazard variants, while classification loss tracked survival-probability accuracy and supported model selection.

5. Experimental results

  • The experiments covered 43 static and 5 dynamic SurvSet datasets and compared TFM, XGBoost, and classical or deep-learning survival approaches.Static comparisons included CoxPH, RSF, DeepSurv, and DeepHit; dynamic comparisons included landmarked CoxPH and RSF, Dynamic DeepHit, and Joint Modeling.Source: §5 · p. 9

5.1. Evaluation metrics

  • Evaluation used IPCW estimators for the time-dependent C-index, Integrated AUC, and Integrated Brier Score to account for conditionally independent censoring.C-index and Integrated AUC assess discrimination, while IBS assesses calibration and overall predictive accuracy.Source: §5.1 · pp. 9–10

5.2. Static results

  • MITRA-CF had the best overall static performance, while XGBoost-CF also performed strongly and direct variants outperformed discrete-hazard and multi-class variants on aggregate metrics.Table 1 averages results across four time-discretization granularities, K ∈ {5, 10, 15, 20}.Source: §5.2 · p. 10, Table 1 · p. 11
  • On the median dataset, discrete-hazard gains over baseline methods monotonically decreased as K increased; by contrast, cumulative-failure gains, especially with MITRA, remained relatively stable or even increased.Source: §5.2 · p. 11
  • The top 30% of static datasets showed 5–9% improvement and the top 10% showed > 20% improvement over the best non-TFM baseline.These are relative gains for MITRA-CF; the median gain was described as modest.Source: §5.2 · p. 11, Figure 2 · p. 12
  • Binary cross-entropy loss was strongly correlated with Integrated Brier Score (r = 0.89), but classification metrics had only moderate correlation with C-Index.The authors suggest selecting among candidate classifiers by choosing the one with the lowest validation binary cross-entropy loss.Source: §5.2 · p. 11, Figure 3 · p. 12

5.3. Dynamic results

  • MITRA-CF was best overall in the dynamic setting at both K = 5 and K = 8, leading on Integrated AUC and IBS and achieving the highest average C-Index.It achieved the best raw values across metrics; the Joint Model was the strongest non-TFM baseline.Source: §5.3 · pp. 12–13, Table 2 · p. 13
  • MITRA-CF’s dynamic-setting lead was most pronounced at t2; it led in C-Index at t1 and t2 but was surpassed by the Joint Model and Landmark CoxPH at t3.It consistently led in Integrated AUC over time and was better than or highly competitive in IBS, particularly when Dynamic DeepHit struggled.Source: §5.3 · p. 13, Figure 4 · p. 14

6. Discussion

The classification-based formulation enables pretrained TFMs to perform survival analysis efficiently, but discretization limits temporal resolution and its censoring theory excludes informative censoring and competing risks. The discussion proposes extending the approach to continuous-time modeling, dependent censoring, and treatment-effect analysis.

Discussion

  • ≈30 minutes on one NVIDIA Tesla A100 GPU (80GB) covered static experiments across 43 datasets and 4 discretization granularities; dynamic experiments across 5 datasets took ≈5 minutes.The static experiments used MITRA-CF and TabPFN-CF.Source: §6 · p. 14

Limitations

  • Discretizing event times into K bins loses temporal resolution, especially for concentrated or non-uniform times; increasing K mitigates this loss but raises computational cost.The missing-label treatment is theoretically justified only under conditionally independent censoring; informative censoring and competing risks fall outside Theorem 4.2.Source: §6 · p. 14

Future work

  • Future work includes fine-tuning TFMs for ranking or C-Index, continuous-time modeling, handling informative censoring, and treatment-effect analysis for time-to-event outcomes.Source: §6 · p. 14

A. Theoretical Results · A.1. Consistent static survival model

Under conditionally independent censoring, the static survival loss is minimized exactly when predictions recover the cumulative failure probabilities at each discretized time, wherever censoring positivity holds.

A.1. Consistent static survival model

  • Under conditional independence of censoring, the static loss is minimized if and only if predictions equal the conditional failure probabilities at each time point.The target is P(T_i ≤ t_k | X_i,0 = x) for each k and supported x satisfying the stated positivity condition.Source: §A.1 · p. 19
  • The result applies to supported covariate values x only when the probability of remaining uncensored through t_k is positive.The theorem states positivity as P(C_i ≥ t_k | X_i,0 = x) > 0; its proof states P(C_i > t_k | X_i,0 = x) > 0.Source: §A.1 · p. 19
  • The proof uses conditional independence of censoring and the event indicator given baseline covariates to justify the relevant expectation equality.The passage derives C_i ⊥⊥ Y_i,k | X_i,0 from Assumption 2.1, which states C_i ⊥⊥ T_i | X_i,0.Source: §A.1 · p. 19
  • For a binary outcome with success probability q, binary cross-entropy is uniquely minimized by predicting q.This pointwise result underlies the theorem’s identification of the loss-minimizing prediction.Source: §A.1 · p. 19
  • When q is 0 or 1, the binary cross-entropy minimum occurs at the corresponding boundary prediction, b = 0 or b = 1.The proof treats these edge cases separately.Source: §A.1 · p. 19
  • For interior q, the stationary prediction b = q is the global minimum because the loss has a strictly positive second derivative.The proof first identifies b = q from the zero derivative, then invokes strict convexity.Source: §A.1 · p. 19

A.2. Consistent dynamic survival model

In dynamic survival analysis, minimizing the population loss recovers the true conditional survival probabilities under the dynamic conditionally independent censoring assumption. The result applies to event-free histories with positive probability of remaining uncensored through the prediction horizon.

A.2. Consistent dynamic survival model

  • As training sample size n tends to infinity, minimizing loss (7) leads the classifier to learn the true underlying dynamic survival probabilities.The population loss minimizer is unique.Source: §A.2 · pp. 20–21
  • Under Assumption 2.2, the population loss is minimized if and only if predictions equal the conditional future-event probability, or one minus conditional survival.The conditioning uses the observed history and event indicators through time t_k.Source: §A.2 · p. 20
  • The equivalence holds for histories where the event has not occurred by t_k and the probability of remaining uncensored through t_k+Δ is positive.These restrictions specify the support on which the theorem identifies the conditional probabilities.Source: §A.2 · p. 20

B. Metrics

The appendix documents evaluation metrics for static and dynamic survival settings and briefly explains ELO calculation.

B. Metrics

  • The appendix formalizes metric definitions, estimators, and implementation details for static and dynamic settings; survival metrics use sksurv, with ELO calculation outlined in Section B.3.Source: B. Metrics · p. 21

B.1. Static Metrics … D. Additional Experimental Details and Results

This appendix documents survival-analysis evaluation metrics and Elo aggregation, model implementations, and dynamic baselines that use longitudinal history to predict landmark-time risk or survival.

B.1.1. Concordance Index (C-Index)

  • Static C-Index measures risk-ranking concordance; its IPCW estimator truncates evaluation at the maximum observed training time to keep estimated uncensoring probabilities positive.The censoring function is estimated with Kaplan-Meier; IPCW is consistent under stated assumptions, including Assumption 2.1.Source: §B.1.1 · p. 21

B.1.2. Integrated AUC; B.1.3. Integrated Brier Score (IBS)

  • Static integrated AUC averages time-specific discrimination using survival-based weights, while IBS integrates time-specific squared errors between predicted survival and observed outcomes.The integrated AUC uses Kaplan-Meier marginal survival estimates for its weights; IBS uses IPCW and trapezoidal integration [Kvamme and Borgan, 2023].Source: §B.1.2 · p. 22, §B.1.3 · p. 22

B.2. Dynamic Metrics; B.2.1. Concordance Index (C-Index); B.2.2. Integrated AUC; B.2.3. Integrated Brier Score (IBS)

  • Dynamic metrics assess risk ranking, discrimination, and squared-error accuracy at future horizons conditional on survival to the current time, using conditional censoring adjustment.Dynamic AUC and IBS evaluate selected future time points from discretized horizons; dynamic C-Index uses a truncated IPCW estimator.Source: §B.2.1 · p. 23, §B.2.2 · p. 23, §B.2.3 · p. 24

B.3. ELO Rating

  • Elo ratings begin at 1000 and average model performance across datasets after within-dataset aggregation over discretization granularities and, dynamically, time points.Pairwise ratings use the default K-factor of 32; static aggregation covers K∈{5, 10, 15, 20}, while dynamic aggregation covers K∈{5, 8} and t1, t2, t3.Source: §B.3 · p. 24

C.1.1. Classification-based models; C.1.2. Alternative classification formulations

  • When needed, training rows are randomly subsampled to model limits—10,000 for MITRA and 50,000 for TabPFN—and survival probabilities are post-processed to be non-increasing; DH and MC variants differ only in prediction target and inference formula.DH estimates discrete-time hazards and reconstructs survival by a multiplicative product; at inference, δ=1 is always input, without requiring true δ.Source: §C.1.1 · pp. 24–25, §C.1.2 · p. 25

C.1.3. Cox Proportional Hazards (CoxPH); C.1.4. Random Survival Forests (RSF); C.1.5. DeepSurv; C.1.6. DeepHit

  • The conventional static baselines include CoxPH and DeepSurv proportional-hazards models, RSF tree ensembles, and DeepHit, which estimates a discrete event-time distribution without proportional hazards.CoxPH estimates coefficients with the partial likelihood; RSF averages tree-level cumulative hazards [Ishwaran et al., 2008]; DeepSurv replaces CoxPH’s linear risk function with a neural network; DeepHit directly models event-time probabilities.Source: §C.1.3 · p. 26, §C.1.4 · p. 26, §C.1.5 · p. 27

C.2. Dynamic Models; C.2.1. Classification-Based Models

  • Dynamic classification models use last-observation-carried-forward covariates and history-based features, then enforce non-increasing predicted survival over future horizons.Feature representations can include elapsed time since the most recent observation and the prediction-horizon index.Source: §C.2 · p. 27, §C.2.1 · p. 28

C.2.2. Landmark CoxPH

  • Landmark CoxPH fits separate models among individuals still event-free at each landmark time and predicts future survival conditional on survival to that landmark.It uses last-observation-carried-forward covariates with running means, and estimates conditional baseline survival with Breslow.Source: §C.2.2 · pp. 28–29

C.2.3. Landmark Random Survival Forest

  • At each landmark, a separate RSF is fit on individuals still event-free, using their history representation to estimate cumulative hazards and conditional survival.Its time-dependent risk score and survival probability are defined from the estimated hazard over future event times.Source: §C.2.3 · p. 29

C.2.4. Dynamic DeepHit

  • Dynamic DeepHit models future event-time distributions from longitudinal history using fixed time bins, with next-interval event probability as its risk score.The implementation outputs cumulative incidence estimates and conditional survival probabilities.Source: §C.2.4 · p. 29

C.2.5. Joint Model

  • The joint model links longitudinal and survival outcomes through shared subject-specific random effects, using a linear mixed-effects model and CoxPH model.Given survival to a landmark and observed history, it predicts future event probabilities; its time-dependent risk score is the next-interval event probability.Source: §C.2.5 · pp. 29–30

D.1. Datasets · D.2. Data Preprocessing · D.3. Hyperparameters

The appendix reports the scope of the static and dynamic datasets, the preprocessing pipeline, and the validation-based hyperparameter selection and search spaces used for baseline models.

D.1. Datasets

  • Across 43 static datasets, sample counts ranged from 26–6805, covariates from 3 to 160, and censoring rates from 6.6% to 94.4%.The median number of unique samples was 461.Source: §D.1 · p. 30, Table 3 · p. 31
  • Across 5 dynamic datasets, sample counts ranged from 103–4603, covariates from 4 to 7, and patient-wise censoring rates from 27.2% to 66.2%.The median number of unique samples was 467.Source: §D.1 · p. 30, Table 4 · p. 32

D.2. Data Preprocessing

  • Each dataset was split 70:15:15 into training, validation, and test sets using event-indicator stratification to preserve event–censoring proportions.Source: §D.2 · p. 30
  • Categorical encoding and missing-value imputation were fit on training data, with held-out feature matrices aligned to the training design matrix and imputed without held-out information.Categorical covariates used one-hot encoding; missing values were handled with a SimpleImputer.Source: §D.2 · p. 30

D.3. Hyperparameters

  • For each dataset, hyperparameters were selected using the validation-split model with the best IPCW-estimator C-Index.Source: §D.3 · p. 30
  • Static baseline searches covered CoxPH, RSF, DeepSurv, and DeepHit; no static classifier-model hyperparameters were defined.CoxPH α ∈ {10^-6, 10^-3, 10^-1}; RSF n_estimators ∈ {50, 100, 200}, min split ∈ {2, 5, 10, 20}. DeepSurv and DeepHit used dropout ∈ {0, 0.1}, hidden layers ∈ {(128, 32), (256, 32), (512, 32)}, and learning rate ∈ {10^-3, 10^-4, 10^-5}.Source: §D.3 · pp. 30, 32
  • Dynamic baseline searches covered Landmark CoxPH, Landmark RSF, Dynamic DeepHit, and a joint model.Landmark CoxPH α ∈ {10^-6, 10^-3, 10^-1}; Landmark RSF n_estimators ∈ {100, 200, 500, 1000}, max depth ∈ {3, 10, 50, None}, min split ∈ {10, 50}. Dynamic DeepHit tuned RNN hidden units ∈ {50, 100}, loss parameter σ ∈ {1, 3}, and learning rate ∈ {10^-4, 10^-3, 5 × 10^-3, 10^-2}; the joint model tuned prior standard deviation σα ∈ {10^-6, 10^-3, 10^-1}.Source: §D.3 · p. 32
  • Classifier-based dynamic models tuned two binary temporal-feature flags, while XGBoost additionally tuned its tree-ensemble parameters.The flags indicate whether to include the prediction-horizon index (quantile_info) and elapsed time since the most recent observation (time_since), yielding four feature representations. XGBoost tuned n_estimators ∈ {100, 200}, learning_rate ∈ {0.1, 0.3}, and max_depth ∈ {3, 6}.Source: §D.3 · p. 32

D.4. Sensitivity to discretization granularity 𝐾 · D.5. Censoring-aware binning via Kaplan–Meier quantiles · D.6. TabPFN v3

The appendix finds cumulative-failure formulations comparatively robust as discretization granularity grows and under Kaplan–Meier-quantile binning, and reports TabPFN v3 results for completeness.

D.4. Sensitivity to discretization granularity 𝐾

  • DH reconstructs survival by multiplying K−1 per-bin estimates, so small estimation errors compound as K grows, analogous to iterated-forecast error accumulation.Source: §D.4 · p. 32
  • Across K∈{5, 10, 15, 20}, cumulative-failure models are among the best across metrics, particularly on IBS at larger K, while discrete-hazard models degrade.Figure 5 compares the best validation-C-index model in each group; Figure 6 shows per-model TFM results, with MITRA-CF strongest at larger K and TabPFN-DH degrading sharply, especially on IBS.Source: §D.4 · p. 32, Figure 5 · p. 33, Figure 6 · p. 33
  • MITRA favors direct cumulative-failure modeling at larger K, while TabPFN-DH can compete at small K but degrades sharply at large K, especially on IBS.The authors hypothesize that differing pretraining inductive biases may explain this interaction and note that survival-specific pretraining could close the gap.Source: §D.4 · p. 33, Figure 6 · p. 33

D.5. Censoring-aware binning via Kaplan–Meier quantiles

  • To address possible bias in observed-event quantiles under heavy censoring, the alternative uses equally spaced quantiles of the training-set Kaplan–Meier survival curve, following Kvamme and Borgan [2021].This places more bin boundaries in high-event-density regions even when censoring obscures events.Source: §D.5 · pp. 33–34
  • Under KM-quantile binning, MITRA-CF remains best overall and cumulative-failure modeling outperforms discrete hazards for each classifier type, preserving the qualitative ranking across 43 datasets.The reported conclusion is robustness to the binning rule despite heavy, heterogeneous censoring.Source: Table 5 · p. 34, §D.5 · p. 34

D.6. TabPFN v3

  • The paper primarily uses TabPFN v2.5 for experimental consistency and additionally reports TabPFN v3 results in Table 6 for completeness.Table 6 averages results over K∈{5, 10, 15, 20} and reports mean ± standard error.Source: §D.6 · p. 35, Table 6 · p. 35
Loading 2601.22259v2…