High Performance, Low Reliability: Uncertainty Benchmarking for Tabular Foundation Models

Researchers benchmarked tabular foundation models across 112 TALENT datasets and found they delivered the strongest AUC versus gradient-boosted decision trees and other baselines.

However, under conformal prediction, the models showed lower conditional coverage, indicating weaker uncertainty reliability despite better accuracy.

Synthetic experiments suggested the performance–uncertainty trade-off can become even more pronounced in some settings.

The results highlight a calibration gap that matters for high-stakes systems needing trustworthy uncertainty estimates.