TabPFN-3.5, a Tabular Foundation Model for messy real-world tables

Prior Labs' TabPFN-3.5 predicts categories or numerical outcomes from tables, including tables with many columns, text fields or numerous distinct category values. Its technical report describes revised input encodings and synthetic training data designed to represent more of these difficult cases.

The reported gains extend beyond clean, randomly split datasets. TabPFN-3.5 ranks first overall on the report's BeyondArena comparison, which includes predicting later observations or previously unseen groups. However, tuned combinations of conventional neural networks retain the best results on grouped, temporal and large datasets in the full comparison. The result supports broader usefulness on messy tables, with the best choice still depending on the data regime.