Abstract
A pretrained tabular foundation model already encodes the normal-data manifold; we turn it into a calibrated cross-modal screen with no training. Our target is a rare subpopulation of records that look ordinary in each modality yet sit off the rule binding them. We probe it by reconstructing each feature from the others under an in-context model (TabPFN-v2) whose fusion keeps the cross-modal correlations in the score. The obstacle is that the reference pool is at once the normality model and the calibration set, so exchangeability fails and the realized false-positive rate reaches 0.49-0.68 at a nominal 0.10. Masking the diagonal of the model’s sample-attention matrix largely corrects this in one forward pass, approximating all n leave-one-out calibration scores. This approximation does not exactly restore exchangeability, so the guarantee we report is empirical: a realized false positive-rate certificate rather than an exact conformal one. Across a six-pair medical evaluation suite this single-pass calibration holds the realized false-positive rate near the nominal level (within about one standard error, FPR ≤0.125) on five of six pairs, with no held-out split and no refitting, at power on par with the strongest training-free baselines. The fused screen beats both unimodal screens on rare cell states neither margin resolves. Code is available at: https://github.com/jose-melo/single-pass-conformal-calibration
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MultiTab_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=geFwxpPLWb
BibTex
@InProceedings{DeJos_SinglePass_MICCAISAT2026,
author = { De Melo Costa, José Lucas AND Ahn, Seong Woo AND Popineau, Fabrice AND Rimmel, Arpad AND Doan, Bich-Liên},
title = { { Single-Pass Conformal Cross-Modal Anomaly Screening with Tabular Foundation Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}