Prior Labs Unveils TabPFN-3.5: A Tabular Model Surpassing the Otto Kaggle Champion

Prior Labs has introduced TabPFN-3.5, the latest iteration of its innovative tabular foundation model. This model operates by predicting outcomes directly from tabular data without requiring any specific training or tuning for each dataset. Impressively, it has secured the top position across seven different tabular benchmarks. Notably, it has surpassed the winning entry from the renowned 2015 Otto Group Product Classification Challenge on Kaggle. This competition involved 3,505 teams and tasked participants with categorizing products into nine distinct groups using 93 obfuscated features. The winning team, comprised of Kaggle grandmasters, utilized a complex ensemble of 36 models based on meticulously crafted features. In contrast, TabPFN-3.5 achieved a score of 0.375 on the Otto leaderboard using only raw data and default settings, demonstrating its efficiency. The model was pretrained solely on synthetic data and did not require prior exposure to the Otto dataset. Furthermore, it offers open weights for research purposes, while commercial applications necessitate a license or API access. The advancements in TabPFN-3.5 include an increase in model parameters and improved encoding techniques, enhancing its predictive capabilities significantly.