acceptodds
Under review as a conference paper at ICLR 2027

Towards Principled Data Generation for Prior-Data Fitted Networks

Abstract

As natural sources of data are becoming exhausted, synthetic data is playing an increasingly important role for pre-training foundation models. Perhaps surprisingly, synthetic data has been shown to surpass natural data in downstream performance. For example, prior-data fitted networks (PFNs), which create an implicit prior through data generation, are trained entirely on synthetic data, but provide state-of-the-art results in tabular prediction. Despite this success, the design of synthetic data remains largely heuristic, requiring expensive trial-and-error rather than principled understanding. In this paper, we work towards a science of data generation, using PFNs to systematically understand the relationship between properties of the data and its performance in downstream tasks. We demonstrate that the synthetic data generation procedure shapes the training dynamics and inductive biases of the model, and thus the optimal pre-training data distribution may be distinct from the downstream task distribution. We identify two key properties of the prior, diversity and complexity, and show that increasing either consistently improves generalization, even if the prior is moved farther from the downstream task. To operationalize these ideas, we also introduce practical proxies and demonstrate that they are predictive of real-world downstream performance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
synthetic data, data generation, in-context learning, prior-data fitted networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ByJ7OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focus on prior-data fitted network (FPN) that are designed for data generation. It observes that the FPN can deviate from the true posterior they are approximating on OOD data. This can make sense, since the OOD data distribution itself can deviate from the training data. In that instance, FPN can provide better generalization than the true posterior they were trained to optimize on. Knowing this, the authors decided to train FPN by increasing diversity and complexity of the data even if it hurt the true bayesian posterior prediction capabilities. They analyze the trade-off between complexity and diversity over the FPN performances. The authors highlight that FPN does not perform exact bayesian inference and that they are learning to leverage inductive bias that are shaped by the training distribution.

One interesting experimental result is that having diverse priors can outperform prior that are less diverse but closer to the downstream task. However, this might come at a cost, since performance could become a bit worse on the training task. They also empirically show that complexity of the training data can also be a predictor of the FPN performance on real world data.

The authors conclude that their experiments allow them to say that FPN are mostly shaped by their architectures and training data, and that they might deviate from Bayesian inference. The authors argue that their experiment will help practitioners designing better prior.

Critical strengths

Good experimental ablations. The experimental results are interesting and they are replicated across different model capacities.

Critical weaknesses

Mostly 1D tasks for the experiments (except one 2D experiment), not sure what the results might be on more complex data distributions. Also, the paper is about more principled data generation, but it seems that the main recommendations of the paper are adding more diverse data and parameters to the model, which is kind of well known. The paper provide an interesting view around what's the effect of more data and more parameters on the FPNs, but I am not sure about the novelty criteria there.

Questions for authors
  • You claim in the abstract that design of synthetic data remains largely heuristic. However I am not sure that this paper make it less heuristic since there are additional dimension a practioner might need to play with: Data diversity and model complexity. So how do you reconcile this idea of providing guided principle while showing this in fact require empirical evaluations?
1 comment

Thank you for your questions. We’re glad you found the experimental results interesting and appreciated the controlled ablations and replication across model capacities.

Most experiments are one-dimensional. What happens on more complex data distributions?

We appreciate the importance of testing these findings beyond simple settings. In Section 5, both the synthetic SCM tasks and UCI experiments use 5-dimensional inputs, and we additionally evaluate on multivariate TabArena regression datasets which go up to 80 dimensions. Our complexity findings are therefore tested directly on more complex synthetic and real-world data.

For the experiments in Sections 3 and 4, we intentionally use low-dimensional settings where the Bayesian predictive distribution can be computed efficiently.

Adding diverse data is already well known. What is the novelty?

The natural intuition is that diversity helps by bringing the training distribution closer to the test distribution. Our novel finding is that increasing diversity can improve PFN generalization even when it moves the training distribution farther from the downstream task and makes exact Bayesian prediction worse (Section 4.1). The same change to the prior therefore has opposite effects on the PFN and what we would expect from a true Bayesian predictor. This demonstrates that diversity can benefit the learned predictor beyond improving distribution alignment.

This finding changes how synthetic priors should be evaluated: a prior that is better matched to the downstream task can nevertheless train a worse PFN. Understanding this distinction is especially valuable for synthetic data generation, where the training distribution itself is a design choice.

How can the paper make prior design less heuristic if empirical evaluation is still required?

To address your feedback, we’ve added a new Section 6 to the updated manuscript that explicitly provides guidance for practitioners. We examine two ways to use the proxies before committing to the full training procedure:

  1. Use early proxy measurements to screen candidate priors. We adapt the complexity and diversity proxies to partial training curves and compare their rankings with final UCI performance. Using only the first 10% of logged training observations, the complexity and diversity proxies achieve Spearman correlations of 0.94 and 0.74, respectively, with final downstream performance. Both proxies identify the three best-performing priors among 18 candidates, showing that early measurements can guide prior selection well before training finishes (Figures 8a and A.11, left).

  2. Use smaller models to guide prior selection for larger architectures. We compare the rankings of our complexity and diversity proxies from a smaller model (2 layers, width 128) with the rankings from a larger model (8 layers, width 512) across various priors. Using the full available training curves, the rankings agree closely across architectures, with Spearman correlations of 0.98 for complexity and 0.93 for diversity. These results support using a smaller model to identify priors that remain highly ranked by either proxy in a larger architecture (Figures 8b and A.11, right).

We hope these additions clarify how practitioners can narrow the search before committing to full training and downstream evaluation.

Reviewer_Y9fKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates and contradicts a set of commonly held beliefs in the world of tabular foundation models by providing counterexamples to past PFN-related claims. The paper shows that in OOD settings, PFN can perform better than exact Bayesian prediction, identifies diversity as a driver of PFN performance (similar to insights in the LLM-related ICL literature), shows examples where simplicity is not better, and provides proxies for translating some of these findings to priors used for SOTA models.

Critical strengths

The paper's argumentative flow and findings are its critical strengths.

The paper is well written and makes it easy to follow the authors' arguments. The arguments are generally convincing and straightforward, making them easy to follow. The paper maintains a clear focus throughout its sections.

I strongly suspect the findings in this paper will spark highly interesting discussions within the community. They align with common folklore in the community while seemingly contradicting some of the strong assumptions in the initial PFN works. Overall, they help the community move forward and keep trying to understand PFNs' true behavior, not just their idealized Bayesian versions.

Critical weaknesses

It is hard to judge whether all claims in this paper will survive rigorous community scrutiny. Some of the claims the paper contradicts were not rigorously proven in their original work either, so the discussions around them may also be strongly connected to discussions of assumptions and (Bayesian) beliefs. The authors did their job by finding potential counterexamples that are not easily refutable. Moreover, the code for these counterexamples is unavailable, which further complicates verification.

Questions for authors

n/a

2 comments

Thank you for your feedback. We’re glad you found the arguments clear and convincing, and see the findings as a useful contribution to understanding PFNs.

To address your feedback, we have added text in Section 3 of the updated manuscript to better clarify how our findings relate to the Bayesian interpretation of PFNs in prior work.

We also provide our code here: https://anonymous.4open.science/r/pfn-iclr-2026/

During the rebuttal period, we have also added a new Section 6 to the updated manuscript on practical guidance for prior selection. The new experiments show that early training measurements identify promising priors and that smaller models give similar proxy rankings to larger architectures (Figures 8 and A.11). These results connect the paper's findings to concrete ways of screening priors before committing to full training.

Thank you for the additional changes. The changes and code do help, but they don't change my mind. I will maintain my score.

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

This paper studies how synthetic training distributions shape the behavior and generalization of prior-data fitted networks (PFNs). It shows that PFNs can depart from Bayesian prediction under their training priors and that these departures can improve performance under distribution shift. The authors investigate prior diversity and complexity through controlled synthetic experiments and real-world regression benchmarks, and propose loss-based proxies for predicting downstream performance.

Critical strengths
  1. The paper offers a useful perspective on synthetic prior design. It demonstrates that a prior influences not only the Bayesian prediction target but also the behavior learned by a finite neural network. The experiments where greater diversity improves generalization despite increased distribution mismatch are particularly interesting and challenge a simple distribution-matching view of pretraining.

  2. The observations are examined across complementary settings. The paper combines controlled problems with tractable Bayesian comparisons and evaluations on UCI and TabArena regression tasks. Additional architecture experiments strengthen the investigation of how the training distribution interacts with the learner. These results provide useful empirical insights without requiring a new prediction architecture.

Critical weaknesses
  1. The experiments vary graph connectivity, node retention, and noise separately, but it remains unclear why these changes improve downstream performance. They may help through different mechanisms, such as broader task coverage or better uncertainty calibration. Grouping them under “complexity” therefore provides limited guidance on which property of a new prior should be changed.

  2. The proposed proxies remain dependent on the trained model. Validation-loss variability and training-loss statistics reflect model capacity and optimization as well as the data-generating distribution. Their correlation with downstream performance is useful, but it does not yet establish that they can reliably guide selection across different prior families. Since these proxies require PFN training, their practical value depends on how much downstream evaluation they can replace, which is not yet clear.

Questions for authors
  1. Can the authors distinguish the benefits of greater structural complexity from those of increased noise? In particular, do the existing experiments indicate whether the gains reflect improved predictive accuracy, improved uncertainty calibration, or both?

  2. Under what conditions should the proposed proxies transfer across prior families or model architectures? How would practitioners use them to choose a new prior without repeating the full training-and-evaluation process?

2 comments

Thank you for your review and questions. We’re glad you found the perspective on prior design useful and the diversity results particularly interesting. Your questions about practical use and predictive uncertainty motivated the new prior-selection and calibration experiments, as detailed below.

How would practitioners use the proxies to choose a new prior without repeating the full training-and-evaluation process?

To address your feedback, we’ve added a new Section 6 to the updated manuscript that explicitly provides guidance for practitioners. We examine two ways to use the proxies before committing to the full training procedure:

  1. Use early proxy measurements to screen candidate priors. We adapt the complexity and diversity proxies to partial training curves and compare their rankings with final UCI performance. Using only the first 10% of logged training observations, the complexity and diversity proxies achieve Spearman correlations of 0.94 and 0.74, respectively, with final downstream performance. Both proxies identify the three best-performing priors among 18 candidates, showing that early measurements can guide prior selection well before training finishes (Figures 8a and A.11, left).

  2. Use smaller models to guide prior selection for larger architectures. We compare the rankings of our complexity and diversity proxies from a smaller model (2 layers, width 128) with the rankings from a larger model (8 layers, width 512) across various priors. Using the full available training curves, the rankings agree closely across architectures, with Spearman correlations of 0.98 for complexity and 0.93 for diversity. These results support using a smaller model to identify priors that remain highly ranked by either proxy in a larger architecture (Figures 8b and A.11, right).

We hope these additions clarify how practitioners can narrow the search before committing to full training and downstream evaluation.

Do the gains reflect improved predictive accuracy, improved uncertainty calibration, or both?

To better distinguish improvements in predictive accuracy from uncertainty effects, we have now added evaluations of both MSE and calibration on UCI in Figure A.8 of the updated manuscript. We find in these new experiments that more complex priors improve both point-prediction accuracy and uncertainty calibration: they achieve lower MSE and a smaller gap between observed and nominal interval coverage. These results show that the benefits include both more accurate predictions and better-calibrated uncertainty, providing further support for prior complexity as a useful design principle.

Under what conditions should the proxies transfer across prior families or model architectures?

Our new rebuttal experiments in Section 6 of the updated manuscript directly examine transfer across model architectures. We find that small and large models give closely aligned complexity and diversity rankings across various SCM priors, supporting the use of smaller models to guide prior selection for larger architectures. The expanded comparison in Figure A.9 further shows that the downstream benefits of more complex priors persist across model sizes, strengthening the evidence that these findings extend across architectures.

Thank you for your thoughtful response and the additional experiments. I will maintain my weak accept rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the AC and reviewers for their thoughtful feedback, which helped us strengthen the paper’s practical guidance and experimental evidence.

Our paper identifies diversity and complexity as useful principles for designing synthetic training data. In particular, we find that a more diverse training prior can improve PFN generalization even when it moves farther from the downstream task and its exact Bayesian posterior predictive becomes worse. We then show that the complexity of the training distribution is a strong predictor of downstream performance: PFNs trained on more complex priors perform better on real-world regression tasks. These findings show why closer distribution matching and simpler priors can lead to worse trained models, and provide concrete directions for improving synthetic data generation.

Rebuttal updates

To address the reviewers' feedback, we have made the following additions to the updated manuscript:

  • Practical guidance for prior selection. We have added a new Section 6 that explicitly provides guidance for practitioners: use early proxy measurements to screen candidate priors, and use smaller models to guide prior selection for larger architectures. The new experiments show that early measurements identify promising priors well before training finishes, and that complexity- and diversity-proxy rankings agree closely across architectures (Figures 8 and A.11).
  • Point-prediction accuracy and uncertainty calibration. To better distinguish improvements in predictive accuracy from uncertainty effects, we have added evaluations of both MSE and calibration on UCI. More complex priors achieve lower MSE and a smaller gap between observed and nominal interval coverage (Figure A.8).
  • Performance across model architectures. We have expanded the comparison across models with different numbers of layers and hidden widths. The downstream benefits of more complex priors persist across model sizes, supporting prior complexity as a useful design principle across architectures (Figure A.9).
  • Code availability. We make our code available here: https://anonymous.4open.science/r/pfn-iclr-2026/

The additions are marked in blue in the updated manuscript. All section and figure references below refer to that version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.