TabPFN 3.5 matches species distribution models without tuning, and fine-tuning makes it the benchmark's best single model

An untuned tabular foundation model reached the level of the best of 21 published methods on the standard benchmark for species distribution prediction. A re-analysis of that benchmark with 21 current methods had found no single best method; the best results came from a five-model ensemble, down-sampled random forests, boosted regression trees and MaxEnt, each depending on choices of regularisation, tree complexity or class weighting that strongly affect performance. On the benchmark of 225 species in six regions with independent presence–absence surveys, TabPFN 3.5 with balanced bagging reached a mean AUC of 0.732, the same as the five-model ensemble's 0.732 and numerically above down-sampled random forests (0.730), MaxEnt (0.728) and boosted regression trees (0.725), though the differences from the best single methods were not significant.

Fine-tuned on presence–background tasks from other regions, TabPFN 3.5 became the best single model on the benchmark. Fine-tuning used a leave-one-region-out design in which the evaluated region contributed no data, with all settings fixed in advance; the only previous study fine-tuned on survey data from the evaluated regions and compared against five methods. TabPFN 3.5-FT had a mean AUC of 0.736 and, in aligned-rank tests among 13 approaches, significantly outperformed MaxEnt (+0.008, higher for 147 of 225 species, P = 0.021) and boosted regression trees (+0.012, P = 0.005), but did not differ significantly from MaxNet, down-sampled random forests or the five-model ensemble.

The model passes the training table together with the points to be predicted into a pre-trained network and returns class probabilities in a single forward pass, without fitting parameters or tuning hyperparameters. Presence–background data are highly imbalanced, their "absences" are background points of unknown status, and presence records carry sampling bias, unlike the synthetic tabular tasks TabPFN was pre-trained on. To fit that imbalance, each balanced bag held all presences and an equal number of random background points; predictions were averaged over 100 bags on the logit scale, with test-time augmentation adding Gaussian noise of standard deviation 0.01 to continuous covariates.

The conclusions rest on 225 species in six regions, records collected decades ago, and independent presence–absence surveys; evaluation measures discrimination at survey sites, not transfer to new environments. Under spatial-block cross-validation the fine-tuned model lost no more accuracy than established methods, but its advantage over MaxEnt was not significant (+0.003, P = 0.46). The authors suggest that if a few hundred real presence–background tasks improve a generic tabular prior, a model pre-trained on many thousands of species could become a foundation model for species distributions, and say they are building such a model for North American birds.

The authors state that six regions are too few for the benchmark to show the advantage holds in new regions; with region as a random effect, the mixed model's interval for the difference from MaxEnt includes zero (−0.012 to 0.020), and the significant contrasts required the 100-bag setting, with a difference of 0.005 (−0.002 to 0.011) at five bags. Among virtual species, the foundation model recovered true probabilities less well than regression-based methods such as MaxEnt under additive relationships, was worse by 0.05–0.07 under environmentally biased sampling for reasons that remain unclear, and fine-tuning removed most of its advantage under complex relationships. TabPFN weights are under a non-commercial licence, and the 100-bag setting takes about 2 minutes per species on one GPU against about 20 seconds for MaxEnt.