pith. sign in

arxiv: 2606.03689 · v1 · pith:XVRZMVHInew · submitted 2026-06-02 · 💻 cs.LG · cs.AI

Staying Alive: Uncensored Survival Analysis with Tabular Foundation Models

Mariana Vargas Vieyra This is my paper

Pith reviewed 2026-06-28 11:31 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords survival analysisright-censoringtabular foundation modelsBuckley-James estimatortraining-freeaccelerated failure timein-context learningevent time prediction
0
0 comments X

The pith

Tabular foundation models perform survival analysis competitively with trained models by imputing right-censored times iteratively in a training-free regime.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a method that applies tabular foundation models to right-censored survival data without any dataset-specific training. It builds an accelerated failure time model around a single scalar and applies an iterative Buckley-James-style imputation step that uses the model to fill in unobserved event times. A sympathetic reader would care because survival analysis appears in healthcare and churn settings where right-censoring is routine and retraining models per dataset is expensive. Experiments on standard benchmarks indicate the resulting predictions match the accuracy of Cox regression and parametric accelerated failure time models that require fitting.

Core claim

A tabular foundation model can be used both to predict event times and to drive iterative imputation of right-censored observations, thereby constructing an accelerated failure time model that needs no training beyond one scalar and yields performance comparable to several parametric and semi-parametric survival regression methods that require training, including Cox regression.

What carries the argument

The non-parametric in-context Buckley-James-style estimator that iteratively imputes right-censored event times from tabular foundation model predictions.

If this is right

  • Survival regression becomes feasible in a purely in-context regime that needs no dataset-specific parameter fitting.
  • Only a single scalar parameter must be fit to turn the foundation model outputs into an accelerated failure time model.
  • Right-censored data can be handled by repeated non-parametric imputation driven by the same foundation model.
  • The approach reaches accuracy levels comparable to Cox regression and trained parametric AFT models on standard benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same in-context imputation loop could be tested on other forms of incomplete time-to-event data beyond classic survival analysis.
  • If the scalar fit remains stable across datasets, the method might reduce reliance on per-problem survival software libraries.
  • Larger-scale experiments could check whether the imputation accuracy degrades when event times are sparse or when the foundation model was not exposed to similar tabular distributions.
  • Combining the imputation step with other foundation-model strengths such as handling missing covariates might produce further gains without added training.

Load-bearing premise

Tabular foundation models already produce event-time predictions accurate enough to support the iterative imputation step for right-censored observations.

What would settle it

On the same standard survival benchmarks, the method would have to show consistently worse concordance or calibration than Cox regression or parametric AFT models to disprove the competitiveness result.

Figures

Figures reproduced from arXiv: 2606.03689 by Mariana Vargas Vieyra.

Figure 1
Figure 1. Figure 1: illustrates how the scalar parameter σ stabilizes, indicating our method successfully converges to a fixed point. Note how the IBS reaches a minimum in early iterations, suggesting our method would benefit from tuning of the number of iterations. 1 2 3 4 5 6 7 8 9 10 Iteration 0.550 0.600 0.650 0.700 0.750 0.800 C-index 1 2 3 4 5 6 7 8 9 10 Iteration 0.100 0.120 0.140 0.160 0.180 0.200 0.220 0.240 IBS 1 2 … view at source ↗
read the original abstract

Survival Analysis (SA) is a statistical framework that models the time span until some event of interest occurs. Widely used in several domains, including healthcare and churn prediction, a central challenge in its applicability stems from the time of the event being partially observed or \emph{right-censoring}. Tabular Foundation Models (TFM) have attracted significant interest in recent years due to their ability to perform prediction tasks in a single forward pass, requiring no dataset-specific parameter fitting. Despite their success, their application to prediction tasks on time-to-event data remains difficult due to right censoring. In this work, we present a training-free method to survival regression by leveraging TFMs to both predict the time of the event and iteratively impute right-censored data. Our method uses a TFM to construct an Accelerated Failure Time (AFT) model requiring no training beyond fitting a single scalar parameter. Subsequently, by building on the Buckley-James estimator, we introduce a non-parametric in-context estimator for right-censored data. Our experiments on standard survival analysis benchmarks show that our method is competitive with several parametric and semi-parametric survival regression models that require training, including Cox regression and parametric AFT models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a training-free (except for one scalar) method for right-censored survival regression that uses a tabular foundation model both to directly predict event times and to perform iterative Buckley-James-style imputation of censored observations, thereby constructing an accelerated failure time (AFT) model whose single free parameter is fitted once; experiments on standard SA benchmarks are reported to show competitiveness with trained parametric AFT models and Cox regression.

Significance. If the empirical competitiveness holds under the stated procedure, the result would be significant because it demonstrates that pre-trained TFMs can be leveraged for a classically difficult censored-regression task without dataset-specific model training, potentially enabling rapid deployment in data-scarce domains such as clinical survival analysis.

major comments (3)
  1. [Methods] Methods, paragraph on AFT construction: the central performance claim rests on the single scalar parameter that converts TFM outputs into an AFT model; the manuscript must specify exactly how this scalar is obtained (closed-form, grid search, or optimization on which data split) and whether the procedure remains strictly in-context or requires a labeled training subset.
  2. [Methods] Methods, Buckley-James imputation subsection: the competitiveness result depends on the iterative imputation loop; the paper must report the number of iterations performed, the convergence criterion, and any sensitivity analysis, because these choices directly affect the imputed targets used to evaluate the final AFT model.
  3. [Experiments] Experiments section, benchmark tables: the reported competitiveness is stated relative to Cox and parametric AFT baselines, yet no ablation isolating the contribution of the TFM predictions versus the imputation loop is provided; without this, it is impossible to attribute performance gains to the claimed training-free regime.
minor comments (2)
  1. [Abstract] Abstract, line 3: 'single forward pass' should be qualified to note the subsequent iterative imputation step.
  2. [Methods] Notation: the symbol used for the single scalar parameter should be introduced once and used consistently throughout the AFT construction equations.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below with clarifications and commit to revisions that strengthen the methodological transparency and experimental analysis without altering the core claims of the work.

read point-by-point responses

  1. Referee: [Methods] Methods, paragraph on AFT construction: the central performance claim rests on the single scalar parameter that converts TFM outputs into an AFT model; the manuscript must specify exactly how this scalar is obtained (closed-form, grid search, or optimization on which data split) and whether the procedure remains strictly in-context or requires a labeled training subset.

    Authors: We agree that the exact procedure for the scalar must be stated explicitly. The revised manuscript will specify that the scalar is obtained via a closed-form solution to the AFT scale parameter using ordinary least squares on the log-event times predicted by the TFM for uncensored observations; this uses only the in-context examples already present in the dataset and requires no separate labeled training subset or gradient-based optimization of the TFM, preserving the strictly training-free regime for the foundation model itself. revision: yes

  2. Referee: [Methods] Methods, Buckley-James imputation subsection: the competitiveness result depends on the iterative imputation loop; the paper must report the number of iterations performed, the convergence criterion, and any sensitivity analysis, because these choices directly affect the imputed targets used to evaluate the final AFT model.

    Authors: We will expand the Buckley-James subsection to report that the iterative imputation is performed for a fixed maximum of 20 iterations with a convergence criterion of mean absolute change in imputed times below 0.01; a sensitivity analysis varying the iteration count and threshold will be added to the supplement to demonstrate that final performance remains stable across reasonable choices of these hyperparameters. revision: yes

  3. Referee: [Experiments] Experiments section, benchmark tables: the reported competitiveness is stated relative to Cox and parametric AFT baselines, yet no ablation isolating the contribution of the TFM predictions versus the imputation loop is provided; without this, it is impossible to attribute performance gains to the claimed training-free regime.

    Authors: We acknowledge that an explicit ablation would improve attribution. The revised experiments section will include a new table comparing (i) direct TFM predictions without imputation, (ii) the imputation loop using a non-TFM baseline predictor, and (iii) the full combined method, thereby isolating the contribution of each component while retaining the training-free character of the TFM usage. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper's central claim is an empirical statement of competitiveness on standard benchmarks after fitting one scalar parameter for the AFT construction and performing Buckley-James-style imputation. This is not a derivation that reduces to its inputs by construction; the performance numbers are obtained from external benchmark evaluation rather than being forced by the fitting step itself. No self-citation chains, uniqueness theorems, or ansatzes imported from prior author work are invoked as load-bearing. The single scalar fit is explicitly acknowledged and does not rename or smuggle in the target result. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The claim rests on one fitted scalar and two domain assumptions about censoring and model capability; no new entities are introduced.

free parameters (1)
  • single scalar parameter for AFT
    Fitted once to turn foundation-model predictions into an accelerated failure time model.
axioms (2)
  • domain assumption Only right-censoring is present
    Standard survival-analysis premise stated in the abstract.
  • domain assumption Tabular foundation models can predict event times usefully in one forward pass
    Required for the training-free construction.

pith-pipeline@v0.9.1-grok · 5735 in / 1266 out tokens · 37455 ms · 2026-06-28T11:31:00.141977+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 4 canonical work pages · 3 internal anchors

  1. [1]

    Journal of Machine Learning Research , year=

    Time-to-Event Prediction with Neural Networks and Cox Regression , author=. Journal of Machine Learning Research , year=

  2. [2]

    IEEE Transactions on Cloud Computing , author =

    Predicting. IEEE Transactions on Cloud Computing , author =. 2020 , note =

    2020

  3. [3]

    and Novakovic, Stanko and Hsu, Lisa and Ernst, Dan and Zardoshti, Pantea and Shah, Monish and Rajadnya, Samir and Lee, Scott and Agarwal, Ishwar and Hill, Mark D

    Li, Huaicheng and Berger, Daniel S. and Novakovic, Stanko and Hsu, Lisa and Ernst, Dan and Zardoshti, Pantea and Shah, Monish and Rajadnya, Samir and Lee, Scott and Agarwal, Ishwar and Hill, Mark D. and Fontoura, Marcus and Bianchini, Ricardo , month = oct, year =. Pond:

  4. [4]

    Yasin, Ahmad , month = mar, year =. A

  5. [5]

    Grill, Jean-Bastien and Strub, Florian and Altché, Florent and Tallec, Corentin and Richemond, Pierre H. and Buchatskaya, Elena and Doersch, Carl and Pires, Bernardo Avila and Guo, Zhaohan Daniel and Azar, Mohammad Gheshlaghi and Piot, Bilal and Kavukcuoglu, Koray and Munos, Rémi and Valko, Michal , month = sep, year =. Bootstrap your own latent:

  6. [6]

    Geyer, Fabien and Scheffler, Alexander and Bondorf, Steffen , month = feb, year =. Network

  7. [7]

    Conditional

    Oord, Aaron van den and Kalchbrenner, Nal and Vinyals, Oriol and Espeholt, Lasse and Graves, Alex and Kavukcuoglu, Koray , month = jun, year =. Conditional

  8. [8]

    Neural Information Processing Systems , year=

    Temporally-Consistent Survival Analysis , author=. Neural Information Processing Systems , year=

  9. [9]

    Annual Review of Public Health , volume=

    Survival analysis in public health research , author=. Annual Review of Public Health , volume=. 1997 , publisher=

    1997

  10. [10]

    2002 , edition=

    The Statistical Analysis of Failure Time Data , author=. 2002 , edition=

    2002

  11. [11]

    Playing Atari with Deep Reinforcement Learning

    Playing Atari with Deep Reinforcement Learning , author=. arXiv preprint arXiv:1312.5602 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Machine learning , volume=

    Learning to predict by the methods of temporal differences , author=. Machine learning , volume=. 1988 , publisher=

    1988

  13. [13]

    Hepatology , volume=

    Primary biliary cirrhosis: prediction of short-term survival based on repeated patient visits , author=. Hepatology , volume=. 1994 , month=

    1994

  14. [14]

    New England Journal of Medicine , year=

    A comparative trial of didanosine or zalcitabine after treatment with zidovudine in patients with human immunodeficiency virus infection , author=. New England Journal of Medicine , year=

  15. [15]

    BMC Medical Research Methodology , volume=

    DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network , author=. BMC Medical Research Methodology , volume=. 2018 , publisher=

    2018

  16. [16]

    IEEE Journal of Biomedical and Health Informatics , volume=

    Deep Survival Machines: Fully Parametric Survival Regression and Representation Learning for Censored Data With Competing Risks , author=. IEEE Journal of Biomedical and Health Informatics , volume=

  17. [17]

    37th Conference on Neural Information Processing Systems (NeurIPS 2023) , year=

    Neural Frailty Machine: Beyond Proportional Hazard Assumption in Neural Survival Regressions , author=. 37th Conference on Neural Information Processing Systems (NeurIPS 2023) , year=

    2023

  18. [18]

    , author=

    Differentiable sorting for censored time-to-event data. , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  19. [19]

    Proceedings of AAAI Spring Symposium on Survival Prediction - Algorithms, Challenges, and Applications 2021 , year =

    Deep Parametric Time-to-Event Regression with Time-Varying Covariates , author =. Proceedings of AAAI Spring Symposium on Survival Prediction - Algorithms, Challenges, and Applications 2021 , year =

    2021

  20. [20]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Survival Permanental Processes for Survival Analysis with Time-Varying Covariates , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  21. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Deep Recurrent Survival Analysis , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  22. [22]

    Playing Atari with Deep Reinforcement Learning , year =

    Mnih, Volodymyr and Kavukcuoglu, Koray and Silver, David and Graves, Alex and Antonoglou, Ioannis and Wierstra, Daan and Riedmiller, Martin , keywords =. Playing Atari with Deep Reinforcement Learning , year =

  23. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

    DeepHit: A Deep Learning Approach to Survival Analysis with Competing Risks , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year=

  24. [24]

    2021 , eprint=

    Survival stacking: casting survival analysis as a classification problem , author=. 2021 , eprint=

    2021

  25. [25]

    Ellis and Brian Whitman and Paul Lamere , title =

    Thierry Bertin-Mahieux and Daniel P.W. Ellis and Brian Whitman and Paul Lamere , title =. 2011 , owner =

    2011

  26. [26]

    AAAI Conference on Artificial Intelligence , year=

    Deep Reinforcement Learning with Double Q-Learning , author=. AAAI Conference on Artificial Intelligence , year=

  27. [27]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Bias in Cross-Entropy-Based Training of Deep Survival Networks , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  28. [28]

    Double Q-learning , volume =

    Hasselt, Hado , booktitle =. Double Q-learning , volume =

  29. [29]

    Continuous control with deep reinforcement learning

    Continuous control with deep reinforcement learning , author=. arXiv preprint arXiv:1509.02971 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Nature , year=

    Human-level control through deep reinforcement learning , author=. Nature , year=

  31. [31]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Attention is All You Need , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  32. [32]

    Adam: A Method for Stochastic Optimization

    Adam: A Method for Stochastic Optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Categorical deep Q-networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  34. [34]

    Sociological Methodology , year=

    Discrete-time methods for the analysis of event histories , author=. Sociological Methodology , year=

  35. [35]

    2008 International Conference on Prognostics and Health Management , pages=

    Damage propagation modeling for aircraft engine run-to-failure simulation , author=. 2008 International Conference on Prognostics and Health Management , pages=. 2008 , organization=

    2008

  36. [36]

    D. R. Cox , title =. 1972 , journal =

    1972

  37. [37]

    Statistical Methods in Medical Research , year=

    Multi-state models for event history analysis , author=. Statistical Methods in Medical Research , year=

  38. [38]

    Machine Learning , year=

    Q-learning , author=. Machine Learning , year=

  39. [39]

    Zame and Jin Long and Uri Shalit and Mihaela van der Schaar , title =

    Changhee Lee and William R. Zame and Jin Long and Uri Shalit and Mihaela van der Schaar , title =. IEEE Transactions on Biomedical Engineering , year =

  40. [40]

    2014 , eprint=

    On the Properties of Neural Machine Translation: Encoder-Decoder Approaches , author=. 2014 , eprint=

    2014

  41. [41]

    2025 , eprint=

    TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models , author=. 2025 , eprint=

    2025

  42. [42]

    Nature , year =

    Accurate predictions on small data with a tabular foundation model , author=. Nature , year=. doi:10.1038/s41586-024-08328-6 , publisher=

    work page doi:10.1038/s41586-024-08328-6

  43. [43]

    International Conference on Learning Representations 2023 , year=

    TabPFN: A transformer that solves small tabular classification problems in a second , author=. International Conference on Learning Representations 2023 , year=

    2023

  44. [44]

    Qu, Jingang and Holzm. Tab. International Conference on Machine Learning , year=

  45. [45]

    International Conference on Machine Learning , year=

    Qu, Jingang and Holzm. International Conference on Machine Learning , year=

  46. [46]

    Tabular Foundation Models Can Do Survival Analysis , doi =

    Kim, Da and Lai, Wei and Zhang, Kelly , year =. Tabular Foundation Models Can Do Survival Analysis , doi =. Arxiv , pages =

  47. [47]

    Future Internet , YEAR =

    Papathanasiou, Dimitris and Demertzis, Konstantinos and Tziritas, Nikos , TITLE =. Future Internet , YEAR =

  48. [48]

    The Stata Journal , year=

    Regression Analysis of Censored Data Using Pseudo-observations , author=. The Stata Journal , year=

  49. [49]

    Biometrika , year=

    Linear regression with censored data , author=. Biometrika , year=

  50. [50]

    Random survival forests , author=

  51. [51]

    2008 , publisher =

    Applied Survival Analysis: Regression Modeling of Time to Event Data , author =. 2008 , publisher =

    2008

  52. [52]

    Journal of Clinical Oncology , volume =

    Randomized 2 x 2 Trial Evaluating Hormonal Treatment and the Duration of Chemotherapy in Node-Positive Breast Cancer Patients , author =. Journal of Clinical Oncology , volume =. 1994 , month = oct, doi =

    1994

  53. [53]

    Nature , volume =

    The Genomic and Transcriptomic Architecture of 2,000 Breast Tumours Reveals Novel Subgroups , author =. Nature , volume =. 2012 , month = apr, doi =

    2012

  54. [54]

    Knaus, W. A. and Harrell, F. E. and Lynn, J. and Goldman, L. and Phillips, R. S. and Connors, A. F. and Dawson, N. V. and Fulkerson, W. J. and Califf, R. M. and Desbiens, N. and Layde, P. and Oye, R. K. and Bellamy, P. E. and Hakim, R. B. and Wagner, D. P. , year =. The. Ann Int Med , doi =

  55. [55]

    Mayo Clinic Proceedings , volume =

    Use of Nonclonal Serum Immunoglobulin Free Light Chains to Predict Overall Survival in the General Population , author =. Mayo Clinic Proceedings , volume =. 2012 , publisher =

    2012

  56. [56]

    Statistics in Medicine , volume =

    Assessment and Comparison of Prognostic Classification Schemes for Survival Data , author =. Statistics in Medicine , volume =. 1999 , month = sep, doi =

    1999

  57. [57]

    Wei, L. J. , title =. Statistics in Medicine , volume =