01
Performance and inference time
BeyondArena results on 142 datasets.
142 datasets · 58-method Elo field
Loading measured results…
KNOW WHERE TO LOOK.
Search less.
Choose better.
Good hyperparameters don’t have to be found from scratch.
- Download a portfolio of 25, 64, or 200 configurations from Hugging Face. Use a shorter prefix, or filter your own from our released results.
- Fit each configuration on your dataset.
- Use validation to choose a model or build an ensemble.
Beyond Foundation Models: PFN Priors Improve
Tabular Hyperparameter Optimization
Artem Chistyakov* · Renat Sergazinov* · Artem Babenko
* Equal contribution · 2026LEARN OFFLINESynthetic tasks
→REUSE ONLINEYour portfolio
THE EVIDENCE
Results on real datasets
The figures use released results from BeyondArena and TabArena.
02
Random search doesn’t catch up
From 1 to 4,000 configurations, the average random search stays below PriorTune.
Synthetic-selected · 64 configsTabRepo-selectedRandom search
What is this win rate? +
03
Real data starts strong. Synthetic scales.
We compare synthetic tasks with 182 TabRepo datasets at similar offline training budgets.
Synthetic · TabICLv2 priorReal · TabRepo
How budgets are compared +
04
Where does the extra compute help?
Increase the number of candidate configurations, synthetic tasks, or both.
Read the grid & color scale +
05
Which hyperparameters get selected?
Compare the original sampling distribution with the 200 selected configurations.
Original sampling priorSelected configurationsChoose a parameter, or pull for three new ones.
Distribution details +
CODE AND DATA
Try PriorTune
Get the portfolios and experiment results on Hugging Face, run the code, or read the paper.
CITE THE WORK
@article{chistyakov2026priortune,
title={Beyond Foundation Models: PFN Priors
Improve Tabular Hyperparameter Optimization},
author={Chistyakov, Artem and Sergazinov, Renat
and Babenko, Artem},
journal={arXiv preprint arXiv:2611.05141},
year={2026}
}