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

Predicting where something occurs from records of where it has been observed, without reliable records of absence, is a common problem in science. Species distribution models (SDMs) such as MaxEnt address it in fields from biological conservation and climate-change assessment to the mapping of disease vectors and archaeological sites, but must be tuned for each dataset. Tabular foundation models, pre-trained on synthetic prediction tasks, predict new datasets without per-dataset training. Only one study has tested them as SDMs, and it fine-tuned on survey data from the evaluated regions. We evaluated tabular foundation models on the standard SDM benchmark (225 species in six regions, with independent presence-absence surveys), under the protocol used to compare 21 methods, and on 120 virtual species with known distributions. Without tuning, a tabular foundation model, TabPFN 3.5, equalled the best published approach, a five-model ensemble (mean AUC 0.732). Fine-tuned on presence-background tasks from other regions, it became the best single model on the benchmark (0.736), outperformed MaxEnt (0.728) and boosted regression trees after correction for multiple comparisons, and was better calibrated than MaxEnt, although six regions were too few to establish the advantage for new regions. Fine-tuning on presence-absence surveys of other species helped only when they came from the evaluated regions. On virtual species, it recovered simple additive relationships less well than regression-based methods but complex ones at least as well when records were unbiased, and was worse with environmentally biased records. Tabular foundation models are competitive SDMs without per-dataset tuning, but they are sensitive to environmentally biased records.