When Repeated Measurements Hurt Tabular Foundation Models
Abstract
Tabular foundation models (TFMs) achieve state-of-the-art performance on IID benchmarks, but how do they respond to repeated measurements at the same inputs? Across six TFMs, synthetic and semi-synthetic experiments show that additional measurements can improve predictions at measured inputs while degrading them elsewhere. Harm is most frequent at low noise, and irrelevant features amplify it in several models; real repeated-measurement datasets show similar patterns. Statistically, the IID rule targeted by PFN pretraining treats each repeated input as fresh evidence about the input distribution. When the prior links this distribution to the response function, the resulting protocol mismatch can outweigh the benefit of additional measurements, even under exact inference. Computationally, interventions identify row-count dependence in TabPFN-v3's attention scaling as a contributor to harm. Resetting the count largely reverses distortion in a noiseless linear diagnostic and attenuates amplification by irrelevant features. Finally, we evaluate ways to retain the benefits of repeated measurements while limiting their harm. Averaging measurements usually outperforms supplying them separately at larger repetition counts, whereas count reset can suppress both harm and useful gains. Code, saved results and reproduction instructions are available at https://github.com/drizmiz/PFN-Repeated-Measurement-Supp
Related
- When Can Voting Help, Hurt, or Change Course? Exact Structure of Binary Test-Time Aggregation
- Conformal Prediction and Conditional Coverage for Tabular Foundation Models
- SurvFM enables tabular foundation models for right-censored survival prediction
- Can Tabular Foundation Models Amortize Statistical Inference?
- Estimating Uncoded Crash Factors with Tabular Foundation and System One Models: Kumo Tabular and Jev
BibTeX
@misc{liu2026when,
title = {{When Repeated Measurements Hurt Tabular Foundation Models}},
author = {Yixiang Liu and Ruizhe Deng and Bibhas Chakraborty and Qiong Zhang and Yan Shuo Tan},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.16801},
eprint = {2610.16801},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.16801},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - When Repeated Measurements Hurt Tabular Foundation Models
AU - Yixiang Liu
AU - Ruizhe Deng
AU - Bibhas Chakraborty
AU - Qiong Zhang
AU - Yan Shuo Tan
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2610.16801
UR - https://arxiv.org/abs/2610.16801
AB - Tabular foundation models (TFMs) achieve state-of-the-art performance on IID benchmarks, but how do they respond to repeated measurements at the same inputs? Across six TFMs, synthetic and semi-synthetic experiments show that additional measurements can improve predictions at measured inputs while degrading them elsewhere. Harm is most frequent at low noise, and irrelevant features amplify it in several models; real repeated-measurement datasets show similar patterns. Statistically, the IID rule targeted by PFN pretraining treats each repeated input as fresh evidence about the input distribution. When the prior links this distribution to the response function, the resulting protocol mismatch can outweigh the benefit of additional measurements, even under exact inference. Computationally, interventions identify row-count dependence in TabPFN-v3's attention scaling as a contributor to harm. Resetting the count largely reverses distortion in a noiseless linear diagnostic and attenuates amplification by irrelevant features. Finally, we evaluate ways to retain the benefits of repeated measurements while limiting their harm. Averaging measurements usually outperforms supplying them separately at larger repetition counts, whereas count reset can suppress both harm and useful gains. Code, saved results and reproduction instructions are available at https://github.com/drizmiz/PFN-Repeated-Measurement-Supp
ER -