TabMF: Learning Dataset Representations with Tabular Foundation Models
TabMF turns pretrained tabular predictors into dataset encoders, then post-trains them for representation work.
The paper adds a dataset-level readout to tabular foundation models and adapts them with contrastive training on synthetic supervised datasets. It tests TabMF versions built on TabICLv2 and TabPFN-2.5 across six tasks on unseen real datasets. The authors report that even frozen TFM readouts are competitive, suggesting the predictors already carry useful dataset-level structure. The TabMF encoders lead the evaluated methods overall, especially for dataset retrieval, compact benchmark construction, and distribution-shift severity assessment. ArXiv · AI/CL/LG's note
The paper adds a dataset-level readout to tabular foundation models and adapts them with contrastive training on synthetic supervised datasets. It tests TabMF versions built on TabICLv2 and TabPFN-2.5 across six tasks on unseen real datasets. The authors report that even frozen TFM readouts are competitive, suggesting the predictors already carry useful dataset-level structure. The TabMF encoders lead the evaluated methods overall, especially for dataset retrieval, compact benchmark construction, and distribution-shift severity assessment. ArXiv · AI/CL/LG's note
score 4