Single-Pass Conformal Cross-Modal Anomaly Screening with Tabular Foundation Models

José Lucas De Melo Costa, Seong Woo Ahn, Fabrice Popineau, Arpad Rimmel, Bich-Liên Doan

LISN, Université Paris-Saclay, CNRS, CentraleSupélec

Oral, MultiTab workshop @ MICCAI 2026, Strasbourg, 27 September 2026.

Paper (open access)Code Slides (PDF)Poster (PDF)

Open-access version provided by the MICCAI Society in agreement with Springer; to appear in MICCAI 2026 Workshops and Challenges, LNCS 17263. Paper information and reviews · PDF (mirror)

In sixty seconds

A tabular foundation model used in context already encodes normality, so it becomes a training-free screen for records that are ordinary in each modality and off the rule that ties the modalities.

The obstacle is that the reference pool is both the model of normality and the calibration set, so exchangeability breaks and the naive gate over-flags (realized FPR 0.49 to 0.68 at nominal 0.10).

Masking the diagonal of the sample-attention matrix scores every reference point without itself and approximates all n leave-one-out calibration scores in one forward pass.

The guarantee is an empirical FPR certificate (at most 0.125), valid on 5 of 6 cross-modal pairs and 8 of 9 clinical tables, at 0.80 valid power.

Fusion must preserve correlation, blocks must be orthogonal, and the anomaly must break a dependency.

Pipeline: multimodal data is fused, scored by reconstruction, calibrated with a diagonal attention mask that yields all leave-one-out scores in one pass, and compared on power at valid FPR. — Fused record, reconstruction score, masked-attention leave-one-out calibration, power at valid FPR.

Let's talk

PhD (CIFRE, LISN / CentraleSupélec, Université Paris-Saclay) ending autumn 2027. On the postdoc market for 2028.

Looking for a host for a three-month research visit in 2027 (employer-funded mission): a group with paired modalities and an annotation of a rare subpopulation that is independent of the scored features.

What I bring: conformal and distribution-free uncertainty for anomaly screening, tabular foundation models in context, and the training dynamics of joint-embedding predictive architectures (NeurIPS 2026).

Email LinkedIn Google Scholar CV ORCID

Other recent work

Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures
NeurIPS 2026 (poster).
project tutorial arxiv code
Mitigating Convergence Collapse in Fixed-Target Anomaly Detectors via Kernel-Anchored Locality Regularization
Oral presentation at CIKM 2026, Rome.
arxiv code
Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets
CIKM 2026, Rome.
arxiv code
High Performance, Low Reliability: Uncertainty Benchmarking for Tabular Foundation Models
ESANN 2026, pp. 115-120.
doi arxiv code
T-JEPA: Augmentation-Free Self-Supervised Learning for Tabular Data
ICLR 2025. *Joint first authors.
openreview arxiv code

Workshop paper in the MICCAI 2026 workshop proceedings (Springer LNCS).