TabPFN and TabICL Outperform Tuned XGBoost on Grinsztajn Benchmark
September 27, 2026
TabPFN and TabICL achieve superior performance over tuned XGBoost across 14 datasets in the Grinsztajn benchmark. These in-context learning models maintain an advantage for datasets up to 32,000 rows without requiring traditional training or hyperparameter optimization.
HOW THIS AFFECTS YOU
●
builderYou may be able to skip expensive hyperparameter tuning for tabular data using in-context learning models.
●
researcherThis validates the scalability of tabular foundation models to medium-sized datasets without retraining.