Causal Pre-training Under the Fairness Lens: An Empirical Study of TabPFN

AI summary

Overview

Research area: Algorithmic fairness and tabular foundation models — specifically the Tabular Prior-data Fitted Network (TabPFN), a transformer pre-trained on synthetic datasets generated by structural causal models (SCMs).

Technical level: Intermediate. Readers should be comfortable with group fairness metrics (Demographic Parity, Equalized Odds), covariate shift, and in-context learning.

Scope: A systematic empirical evaluation of TabPFN and a fine-tuned variant (FT-TabPFN) against three classical baselines across four tabular fairness benchmarks, measuring accuracy, fairness, and robustness under spurious-correlation and missing-not-at-random (MNAR) covariate shift.

What This Paper Is About

Foundation models for tabular data such as TabPFN are pre-trained on millions of synthetic datasets produced by structural causal models rather than on real, potentially biased data — a design that could plausibly reduce reliance on spurious correlations and improve fairness. The problem is that this fairness potential has not been systematically tested, especially under distributional shifts. The authors therefore benchmark TabPFN and FT-TabPFN against Logistic Regression, Random Forest, and a Multi-Layer Perceptron to ask how much fairness causal pre-training alone actually buys.

Key Contributions

1. A comprehensive empirical evaluation of TabPFN and FT-TabPFN on predictive accuracy and group fairness (Demographic Parity and Equalized Odds, computed with FairLearn version 0.7.0) across four tabular fairness benchmarks — Heart (303 samples, 13 features; sensitive: age, sex), Bank (5,000, 14; education, family), Law (21k, 8; race, sex), and Adult (48k, 14; race, sex) — including dataset-size sweeps at 500, 1k, and 10k samples on Adult and Law.
2. A spurious-correlation stress test using a pseudo-correlation perturbation: a synthetic feature Z_spur ~ N(mu_y, 0.5), correlated with the label in training and with its correlation sign reversed in the test set, evaluated via accuracy, DP, EO, and a Flip Consistency Rate (the proportion of unchanged predictions).
3. A covariate-shift stress test using MNAR selection bias, in which biased subsets of N = 500 samples remove 70% of instances from specific outcome- and attribute-dependent subgroups, compared against a size-matched "reduced-clean" unbiased subset.
4. A framing of the results as an asymmetry: causal pre-training mitigates spurious statistical shortcuts but delivers limited to no fairness improvement under realistic MNAR selection bias, with the paper's stated implication that causal pre-training is "helpful but insufficient" for algorithmic fairness.

Main Findings

• TabPFN leads on accuracy. FT-TabPFN achieves the highest or near-highest accuracy across all dataset sizes and datasets, reaching 0.99 on both Bank and Heart. On Bank, RF, MLP, TabPFN and FT-TabPFN all reach accuracy 0.99 (LR: 0.95); on Heart, TabPFN reaches 0.96 and FT-TabPFN 0.99, versus 0.84 (LR), 0.91 (RF), and 0.90 (MLP).
• Fairness gains are moderate and inconsistent. TabPFN follows closely in accuracy while showing superior fairness, particularly in small-data regimes — for example, at 500 samples on Adult, average DP for TabPFN and FT-TabPFN is significantly lower than other baselines. But on Heart, FT-TabPFN has the highest DP in Table 1 (0.35) and TabPFN 0.30, above LR (0.23).
• Fairness drifts with scale. Beyond TabPFN's previously validated 10k-sample regime, both TabPFN and FT-TabPFN maintain strong accuracy, but Equalized Odds shows greater variability — moderate increases on Adult and slight fluctuations on Law — while remaining at competitive levels.
• Robustness to spurious correlation. Under pseudo-correlation perturbation, FT-TabPFN and TabPFN achieve the best flip consistency (0.82–0.97 across datasets), leading on Heart (0.89 and 0.87) and Bank (both 0.97). RF follows (0.70 on Heart, 0.96 on Bank), while MLP (0.56 on Heart) and LR (0.51 on Heart) lag. On Bank, all models except LR reach accuracy 0.97 (FT-TabPFN, TabPFN) or 0.94 (RF) and 0.93 (MLP).
• One dataset collapses under flipping. Law shows anomalously low consistency (<0.62) and accuracy (<0.56) across models post-flip — FT-TabPFN and TabPFN both fall to 0.27 consistency and 0.21 accuracy — despite strong non-perturbed utility. The authors attribute this to extreme imbalance in sensitive attributes and labels that may foster proxy reliance, noting prior work has also flagged anomalies for this dataset.
• MNAR shift: accuracy holds, fairness does not. TabPFN and FT-TabPFN achieve the highest accuracy on biased (Acc_b: 0.89–0.99) and clean (Acc_c: 0.86–0.99) data, outperforming baselines by 2–10%. However, fairness improvements are inconsistent: they attain the lowest EO in Heart (0.08 for TabPFN on biased data) and Law (0.33), but EO remains notably higher in Bank (up to 0.59) and Adult shows only marginal gains. DP is particularly unstable, ranging from near-optimal in Law and Adult to the highest values in Heart (0.41–0.42) and Bank. On Heart, all models show DP in the 0.38–0.42 range under shift.
• The central asymmetry. Causal pre-training effectively mitigates spurious statistical shortcuts but provides limited to no fairness improvement under MNAR selection biases, possibly because the pre-training cannot compensate for systematic under-representation of causal data from subpopulations. The authors also note that shortcomings appear in entangled scenarios such as imbalanced sensitive attributes.

Methodology in Plain English

The authors take four well-known tabular datasets used in fairness research — Heart, Bank, Law, and Adult — and put five models through the same pipeline (standardized features, one-hot encoded categoricals). The baselines are Logistic Regression with L2 regularization (C = 0.1), Random Forest with 50 trees of maximum depth 5 and minimum samples per split of 10, and a Multi-Layer Perceptron with one hidden layer of 50 units (L2 alpha = 0.01, up to 300 training iterations). The two models under study are TabPFN in zero-shot mode with pretraining limits disabled so it can run on datasets exceeding 10k samples (n_estimators = 2, using 5000 subsampled training examples as context) and FT-TabPFN, fine-tuned for 10 epochs via Adam at learning rate 1 × 10⁻⁵ with meta-batch size 1, inner batch size 5000, and cross-entropy loss on batched in-context predictions, evaluated with the same 5000-sample context.

For each setting they measure accuracy plus Demographic Parity and Equalized Odds, averaged over race and sex where relevant. To test shortcut reliance, they add an artificial feature that predicts the label in training but has its correlation reversed at test time, and check how much accuracy, DP, EO, and the fraction of unchanged predictions (Flip Consistency Rate) degrade. To test selection bias, they build 500-sample training subsets in which 70% of specific outcome- and attribute-dependent subgroups are removed — women with high cholesterol less willing to seek medical attention (Heart), high-income middle-aged and older individuals reluctant to disclose finances (Bank), women with long working hours less willing to report income (Adult), and minority students with high LSAT scores less likely to be tracked (Law) — and compare against randomly downsampled unbiased data of the same size. All results are averaged over five independent runs with different random seeds for statistical reliability. Code is available at https://github.com/ql909/An-Empirical-Study-of-TabPFN.

Why This Matters

Impact on research. The paper tests a widely assumed claim — that causally grounded pre-training buys fairness for free — and finds the benefit is partial. It also draws a distinction the authors call previously overlooked: robustness to spurious correlation and robustness to MNAR selection bias are not the same thing, and success on the first does not imply success on the second. The authors position this as a gap between causal robustness and algorithmic fairness rather than a solved problem. Notably, the study deliberately excludes fairness-specific adaptations such as FairPFN, which directly optimizes causal fairness metrics during adaptation but is restricted to binary sensitive attributes; this isolates how far causal pre-training alone goes.

Real-world applications.

• High-stakes tabular decision systems (credit, admissions, benefits) where TabPFN-style models are attractive because they need no large labeled training set.
• Clinical or health screening models trained on small, selectively collected samples where women or other subgroups are under-represented.
• Financial services, where income and education attributes drive both predictions and disclosure behavior, matching the MNAR scenario tested on Bank and Adult.
• Educational or legal tracking decisions, where the Law dataset's flip-induced collapse shows how badly proxy reliance can fail in imbalanced data.

Industry relevance. Practitioners evaluating whether to adopt tabular foundation models should treat the accuracy and fairness claims separately: strong accuracy and shortcut robustness do not extend to MNAR selection conditions, so fairness auditing and targeted interventions remain necessary before deployment.

Future Directions

• Integrating causal imputation for missing attributes, which the authors name as a direction for closing the MNAR fairness gap.
• Developing dynamic fairness regularization that adapts as dataset scale and distribution change, given that DP and EO were unstable across dataset sizes.
• Investigating why the Law dataset collapses under spurious-correlation flipping (consistency 0.27 for both TabPFN and FT-TabPFN) while retaining strong non-perturbed accuracy, and how extreme sensitive-attribute and label imbalance drives proxy reliance more broadly.
• Extending the evaluation to fairness-specific adaptations such as FairPFN and beyond its current restriction to binary sensitive attributes, to see whether explicit fairness objectives close the gap that causal pre-training alone leaves open.

Target Audience

Researchers and practitioners in algorithmic fairness, trustworthy machine learning, and tabular machine learning, particularly those assessing foundation models for high-stakes or small-sample decision tasks. It also suits applied data scientists deciding whether pre-trained tabular models reduce the need for explicit fairness interventions, and anyone studying the relationship between causal reasoning, distribution shift, and group fairness.

Authors’ abstract

Foundation models for tabular data, such as the Tabular Prior-data Fitted Network (TabPFN), are pre-trained on a massive number of synthetic datasets generated by structural causal models (SCM). They leverage in-context learning to offer high predictive accuracy in real-world tasks. However, the fairness properties of these foundational models, which incorporate ideas from causal reasoning during pre-training, remain underexplored. In this work, we conduct a comprehensive empirical evaluation of TabPFN and its fine-tuned variants, assessing predictive performance, fairness, and robustness across varying dataset sizes and distributional shifts. Our results reveal that while TabPFN achieves stronger predictive accuracy compared to baselines and exhibits robustness to spurious correlations, improvements in fairness are moderate and inconsistent, particularly under missing-not-at-random (MNAR) covariate shifts. These findings suggest that the causal pre-training in TabPFN is helpful but insufficient for algorithmic fairness, highlighting implications for deploying TabPFN (and similar) models in practice and the need for further fairness interventions.