Interpretable Tabular Foundation Models via In-Context Kernel Regression

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Miftachov, Ratmir, Charron, Bruno, Valentin, Simon
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918319471198208
author Miftachov, Ratmir
Charron, Bruno
Valentin, Simon
author_facet Miftachov, Ratmir
Charron, Bruno
Valentin, Simon
contents Tabular foundation models like TabPFN and TabICL achieve state-of-the-art performance through in-context learning, yet their architectures remain fundamentally opaque. We introduce KernelICL, a framework to enhance tabular foundation models with quantifiable sample-based interpretability. Building on the insight that in-context learning is akin to kernel regression, we make this mechanism explicit by replacing the final prediction layer with kernel functions (Gaussian, dot-product, kNN) so that every prediction is a transparent weighted average of training labels. We introduce a two-dimensional taxonomy that formally unifies standard kernel methods, modern neighbor-based approaches, and attention mechanisms under a single framework, and quantify inspectability via the perplexity of the weight distribution over training samples. On 55 TALENT benchmark datasets, KernelICL achieves performance on par with existing tabular foundation models, demonstrating that explicit kernel constraints on the final layer enable inspectable predictions without sacrificing performance.
format Preprint
id arxiv_https___arxiv_org_abs_2602_02162
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Interpretable Tabular Foundation Models via In-Context Kernel Regression
Miftachov, Ratmir
Charron, Bruno
Valentin, Simon
Machine Learning
Tabular foundation models like TabPFN and TabICL achieve state-of-the-art performance through in-context learning, yet their architectures remain fundamentally opaque. We introduce KernelICL, a framework to enhance tabular foundation models with quantifiable sample-based interpretability. Building on the insight that in-context learning is akin to kernel regression, we make this mechanism explicit by replacing the final prediction layer with kernel functions (Gaussian, dot-product, kNN) so that every prediction is a transparent weighted average of training labels. We introduce a two-dimensional taxonomy that formally unifies standard kernel methods, modern neighbor-based approaches, and attention mechanisms under a single framework, and quantify inspectability via the perplexity of the weight distribution over training samples. On 55 TALENT benchmark datasets, KernelICL achieves performance on par with existing tabular foundation models, demonstrating that explicit kernel constraints on the final layer enable inspectable predictions without sacrificing performance.
title Interpretable Tabular Foundation Models via In-Context Kernel Regression
topic Machine Learning
url https://arxiv.org/abs/2602.02162