MulTaBench · frozen encoders · 10K rows

Which tabular models handle text and images best?

Tabular models on 80 multimodal datasets. Text columns are embedded with frozen e5-small and images with frozen DINOv3-small, each reduced to 30 PCA components per column; ConTextTab, TabSTAR and TabSTAR-v1.1 read raw text themselves. Default hyperparameters, no bagging, 3 folds × 2 repeats, up to 10K training rows.

Code · Datasets · Paper · CSV

Datasets

Task

Modality

Elo rating

Elo vs cost

Each model's Elo, under the filters above, against its median train + predict time per 1,000 rows (log scale). ★ marks the Pareto front: no other model is both faster and stronger.

Per-dataset results

Sort

best worst cell = test error, mean of 6 folds · outlined = best on the dataset · shading normalized per dataset text image image + text

Cost

Median time per 1,000 rows over all successful runs, on the hardware each model is assigned. Embedding time is excluded: every run reads the cached embeddings.

How to read this

Elo

As in TabArena, every fold is its own battle: on each of a dataset's 6 folds, each pair of models plays one battle, won by the lower test error. The battles are fit as a Bradley–Terry model, with RandomForest fixed at 1000, by TabArena's bencheval. The interval is the 95% range over 1,000 bootstrap resamples of datasets, each with all its folds, bootstrapped before anchoring, so each is the model's own. Win rate, average rank, normalized error and best-on count are per fold as well.

Missing results are not imputed

Datasets

The benchmark's 80 datasets are 40 core and 40 extended ones, as their Hugging Face repos are named (core-…, extended-…); the Datasets filter keeps either half.

Data

Every run is in multabench/leaderboard/models/<model>.csv. The tables here are plain CSVs in this Space's data folder: leaderboards.csv (one leaderboard per tier, task and modality filter), per_dataset.csv and cost.csv.

Metrics

Test error is lower-is-better: 1 − ROC AUC for binary tasks, log loss for multiclass, RMSE for regression. Normalized error rescales each dataset to 0 (best model) – 1 (worst).

Hardware

Each run gets 8 CPUs and 32 GB. Tree models run on CPU (Xeon 8581C); the others on one H100 80 GB (Xeon 8481C).

Using the leaderboard in your paper

Elo vs cost plot

Ranking tableCSV · every filter

Per dataset and foldPer dataset CSV Every run, per model (GitHub)

Version history

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},
}

Have a model that should be on the board?

Open a GitHub issue with the model, or a feature, bug or question about the board. Models run on the same hardware and protocol as everyone else.

Open an issue