MulTaBench is a benchmark for multimodal tabular learning with text and images. Its purpose is to drive smarter encoding methods, towards our vision of truly Multimodal Tabular Foundation Models. Its 80 datasets balance image and text modalities, classification and regression, and how hard the text or images are to encode, with the Core slice focusing on the 40 datasets where frozen embeddings are not enough.
Frozen embeds text with e5-small and images with DINOv3-small, and projects each column to 30 dimensions with PCA. TAR, target-aware representations, first fine-tunes the encoder on the training labels, then projects the same way.
Elo rating
Have a model or an encoding method that should be on the board?
Open a GitHub issue with the model or encoding method, or a feature, bug or question about the board. Everything runs on the same hardware and protocol.
Open an issueFrozen embeds text with e5-small and images with DINOv3-small, and projects each column to 30 dimensions with PCA. TAR, target-aware representations, first fine-tunes the encoder on the training labels, then projects the same way.
Their Elo pits the two against each other on every model and dataset, leaving out E2E models.
Pareto Frontier
Each entry's Elo, under the filters above, against its median train + predict time per 1,000 rows (log scale), TAR's fine-tuning included; With encoding time adds an estimate of the encoders' time on those rows. ★ marks the Pareto front: no other model is both faster and stronger.
Cost
Median time per 1,000 rows over all successful runs, on the hardware each model is assigned. Runs read the cached embeddings, so this leaves out encoding; the text and image columns estimate it from the cached encoders' own time, and With encoding time adds it to the total. Fine-tune is the mean time TAR's LoRA fine-tuning took per 1,000 training rows, on an H100, and always counts.
Using the leaderboard in your paper
Cite MulTaBench
@misc{arazi2026multabenchbenchmarkingmultimodaltabular,
title={MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image},
author={Alan Arazi and Eilam Shapira and Shoham Grunblat and Mor Ventura and Elad Hoffer and Gioia Blayer and David Holzmüller and Lennart Purucker and Gaël Varoquaux and Frank Hutter and Roi Reichart},
year={2026},
eprint={2605.10616},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2605.10616},
}