acceptodds
Withdrawn from ICLR 2027

DialPFN: Assumption-Conditioned In-Context Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In time-to-event studies, some individuals leave follow-up before the event occurs, leaving its timing unknown. Predictions then depend on assumptions about how event times relate to when follow-up ends, a relationship that the observed records alone generally cannot determine. Analysts therefore need to understand how their predictions change under alternative assumptions, not just obtain one answer under a fixed assumption. We introduce DialPFN, a prior-fitted network that accepts an observed dataset, an individual's features, and a stated dependence assumption, and returns an event-time distribution under that assumption. Synthetic studies provide complete target times for supervision, while the network receives only the corresponding partially observed table and a stated dependence assumption. At inference, changing this input produces a sensitivity analysis for each query without updating model weights. A frozen-backbone quantile head and reference-sample bagging produce individual predictive distributions. For the reference predictor, supplying the generating assumption reduces strong-dependence survival-curve error by 48.3 to 54.7% across five reserved simulation banks. Ordinary prediction is preserved across 17 real datasets, and a separate Melanoma evaluation supports stable individual sensitivity curves. All six training runs pass seven grouped evaluation checks. DialPFN makes individual predictions conditional on explicit assumptions, allowing analysts to examine how their conclusions change across plausible alternatives.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
prior-fitted networks, tabular in-context learning, sensitivity analysis, uncertainty quantification, survival analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_X5C8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

DialPFN predicts individual event-time distributions conditional on observed context data and a specified dependence assumption. It learns from simulated studies with complete target times and produces sensitivity curves by varying the assumption input without retraining. The implementation combines a quantile head with context bootstrap averaging. Validation includes synthetic prediction and coverage tests, real-data prediction, and sensitivity-stability evaluations. In synthetic experiments, supplying the generating dependence assumption improves prediction accuracy compared with assuming independence.

Critical strengths
  1. A useful assumption-conditioned prediction task. Holding the context set and individual fixed while varying the dependence assumption provides individual sensitivity curves without repeated model fitting. This is a practically meaningful extension of survival PFNs.

  2. Evidence that the model uses the assumption input effectively. Supplying the generating assumption improves prediction accuracy in simulations, including the reserved Frank-copula evaluation. These results support the basic feasibility of the approach.

Critical weaknesses

Critical Weaknesses

  1. Limited novelty and insufficient empirical validation. Assumption-conditioned modeling and sensitivity analysis are well established, and the contribution appears mainly to extend these ideas to PFNs. Comparisons should extend beyond PFNs to dependency-aware neural methods, including HACSurv [1], which learns dependence structures, and KnownCopula [2], which uses a prespecified copula. Comparisons under comparable prediction targets and dependence information are needed to establish an accuracy or efficiency advantage.

  2. Substantial writing and reproducibility concerns. Many sentences are awkward or difficult to follow, and unfamiliar terminology adds to the reading burden. In my view, several sections require substantial rewriting for the paper to meet the standard expected for acceptance. Reporting is also incomplete: training hardware, actual training time, and sensitivity-sweep runtime are missing, and no accessible code is provided. These omissions make the practical benefits and reproducibility difficult to assess.

[1] Xin Liu, Weijia Zhang, and Min-Ling Zhang. HACSurv: A Hierarchical Copula-Based Approach for Survival Analysis with Dependent Competing Risks. AISTATS, 2025.

[2] Ali Hossein Foomani Gharari, Michael Cooper, Russell Greiner, and Rahul G. Krishnan. Copula-Based Deep Survival Models for Dependent Censoring. UAI, 2023.

Questions for authors
  1. Do you have comparisons with HACSurv and KnownCopula under comparable prediction targets and dependence information? Such results would materially affect my assessment of the empirical contribution.

  2. What hardware and actual training time were required, and what is the total runtime for a sensitivity sweep including bootstrap averaging? Is an anonymous implementation or pretrained checkpoint available for verification?

Reviewer_onNpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DialPFN, a method that combines the advantages of PFN based models like SurvivalPFN and copula-based models for prediction when some samples opt out of follow-ups (or experience a competing event) before the target event occurs. SurvivalPFN does not need training for each dataset. However, it assumes drop-out time and the target event are independent. Copula-based models on the other hand, allow the user to include their assumption about how drop-out time is related to the target event, but they need refitting for each dataset, each assumption, and the strength of the dependence. DialPFN provides the advantage of copula-based models with PFN-based models, creating a method that takes external assumptions, and does not need per-dataset training. Moreover, DialPFN does not need refitting for different assumptions and for two stopping processes, the model allows the user to see how these assumptions and their target-dependence strength can affect the final prediction.

Critical strengths
  1. Allowing the user to include their assumption about how dropout relates to the target event is an interesting idea. The same mechanism works for both dropout and competing events, and the user can see how the prediction for an individual changes with strength of target dependence for their stopping processes.

  2. In the simulation, supplying the correct assumption reduces error and the effect holds on a held-out copula family, and results replicate across six training seeds.

Critical weaknesses
  1. The presented DialPFN only supports two stopping processes (K=2). The authors mention that for further stopping processes, we would need retraining. Moreover, their model does not support settings where dropouts are healthier than average, or where several competing causes act together. Copula-based models already handle these cases. The potential benefits (small-sample performance, calibrated individual distributions) are plausible but untested.

  2. DialPFN supports 400 context samples. fitting a copula-based model on a dataset with 400 samples, takes only seconds. In practice, the difference is “a quick refit per value” versus “100 forward passes per value.” Time complexity analysis is not theoretically, or empirically discussed in the paper, so it is unknown whether the method offers any computational advantage. Moreover, Copula-based methods don't have the 400 sample limit and therefore have the ability to fit the data better. Settings where copula-based models could outperform DialPFN, especially in cases where training data includes more than 400 samples are not tested.

Questions for authors
  1. How does DialPFN compare with a copula-based model given the same assumption, in accuracy and computation, as sample size grows beyond 400?

  2. Why is dependence restricted to [0, 0.9] and the correlation between stopping processes fixed at zero?

  3. Could stopping processes be padded to a larger maximum with a mask token? How would the constraint Σaⱼ² < 1 limit representable dependence as K grows?

  4. How were the inference constants (0.75 and 1.05) selected, and on which data?

Reviewer_XuFKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper extends survival prior-data fitted networks (PFNs) to settings with up to two dependent censoring mechanisms. The model is pretrained on synthetic data in which event and censoring times are drawn from several parametric families, including Weibull, given the covariates, and their dependence is generated by a Gaussian copula. The model outputs a log-normal distribution for the event time, which a small quantile head refines by adjusting fifteen of its quantiles. The model assumes the two censoring times are conditionally independent of each other given the event time. It takes the two correlation parameters, ρ(T,C1)\rho(T, C_1) and ρ(T,C2)\rho(T, C_2), as inputs and predicts the event-time survival distribution given those values. Evaluation consists of: (i) the RMSE of predicted survival curves against the known true curves on synthetic data, and (ii) the mean concordance index (C-index) and integrated Brier score (IBS) across 17 real-world datasets.

Critical strengths

The paper explores an important problem of accounting for dependent censoring in survival PFN estimators.

Critical weaknesses
  • Limited novelty and motivation: The method appears to be a direct extension of survival PFNs: the training data are simulated with copula-based dependent censoring, and the copula correlations are supplied by the user as assumptions at inference time. The motivation for this particular setup is not clearly articulated, and the formulation seems tailored to the paper's specific design rather than to a general class of dependent-censoring problems.

  • Related work and baselines: The paper overlooks several copula-based survival methods for dependent censoring (e.g., [1, 2]), as well as standard non-PFN approaches such as CoxPH, AFT models, and their deep-learning extensions.

  • Data-generating process (Section 4.1): The DGP would benefit from both clearer exposition and a broader scope. A graphical (causal-diagram) presentation, with each assumption explicitly stated and motivated, would greatly improve readability.

    • It's unclear why the paper restricts itself to K=2K = 2 censoring mechanisms with ρ(C1,C2)=0\rho(C_1, C_2) = 0. This choice also implicitly constrains the admissible dependence levels to ρ(T,C1)2+ρ(T,C2)2<1\rho(T, C_1)^2 + \rho(T, C_2)^2 < 1, which is not discussed.
    • The additional events are treated as censoring mechanisms rather than following the standard competing-risks formulation, in which multiple event causes are modelled jointly (e.g., via cause-specific hazards or cumulative incidence functions).
    • Only the Gaussian copula is considered. Alternatives with different tail behaviour (e.g., Clayton, Gumbel, Frank) are not explored, even though the copula family is not identifiable from the data and is likely to affect the results.
  • Sensitivity analysis: Although the paper claims the approach supports sensitivity analysis, this appears impractical: predictions are conditional on user-supplied values of ρ(T,C1)\rho(T, C_1) and ρ(T,C2)\rho(T, C_2), which are unknown in practice, so users must sweep over many values without guidance on which are plausible. It's unclear why the paper does not instead pursue a bounds-based approach, as is typical in causal sensitivity analysis (e.g., [3]), which would directly return the range of survival curves compatible with a given level of dependence.

  • Clarity of the predictive model (Section 4.2): The precision and clarity of this section should be improved.

    • The model predicts a Gaussian distribution on log target time, which corresponds to a log-normal event-time distribution, but this is not stated explicitly.
    • The construction of the event-time distribution from fifteen prespecified quantiles, with bounded adjustments and fixed post hoc shrinkage and widening factors, needs clearer motivation and justification.
    • It is also unclear why alternative distributional approaches, such as other quantile-regression formulations or discrete-time prediction as in DeepHit, were not considered or would not work in this setting.
  • Experimental evaluation: The experimental results are underwhelming.

    • Real data: The paper provides no description of the 17 real-world datasets and reports only summary C-index and Brier scores, without comparison against alternative time-to-event baselines.
    • Synthetic data: The experimental setup is not described, and it's unclear why RMSE is the only metric used. Also, the synthetic results appear to assume correctly specified dependence (i.e., the true ρ\rho values and copula family are supplied), which cannot be reproduced in practice, where both are unknown. Experiments under misspecification, such as incorrect ρ\rho values, a different copula family, or non-log-normal event times, would be more informative.

References

  • [1] Foomani et al., Copula-based deep survival models for dependent censoring, UAI, 2023.
  • [2] Zhang et al., Deep copula-based survival analysis for dependent censoring with identifiability guarantees, AAAI, 2024.
  • [3] Javurek et al., Amortizing causal sensitivity analysis via prior data-fitted networks, arXiv, 2026.
Questions for authors

Could the authors motivate the main modeling assumptions, the choices made in the data-generating process, and address the concerns raised in the weaknesses section?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.