A tabular foundation model rivals established species distribution models from presence-only records

TabPFN 3.5 matches the best species distribution model on the standard benchmark (mean AUC 0.732) without per-dataset tuning, using presence-only records across 225 species in six regions.

• Why it matters: SDMs are widely used to map species, disease vectors, and other phenomena from presence data, but they typically require dataset-specific tuning, and prior tests of tabular foundation models were limited.
• What they did: The study evaluated TabPFN 3.5 under the 21-method SDM benchmark protocol (independent presence-absence surveys) and on 120 virtual species, comparing results with MaxEnt and other methods, with and without fine-tuning.
• The result: Without tuning, TabPFN 3.5 equalled a five-model ensemble (mean AUC 0.732); with fine-tuning on presence-background tasks it reached 0.736, exceeded MaxEnt (0.728), and improved calibration, but showed reduced performance with environmentally biased virtual records.

AI summary of the abstract · may contain errors · not medical advice · read the paper.