Ground beetles are some of the best indicators of ecosystem change, yet identifying them takes an expert with a microscope. I benchmark vision and vision-language encoders on long-tailed, real collection data, and measure what matters to the people who would deploy them: accuracy on rare species, labeling cost, and robustness when the camera moves from the lab to the field.
12 vision and vision-language encoders, linear-probed on about 100K images of 1,769 carabid species from four long-tailed datasets. ViLT reaches 97% genus and 94% species accuracy, half the training data costs little, and lab-to-field transfer drops 37%.
The first version of the benchmark, presented at CV4Animals: where frozen encoders succeed on curated collections, and how sharply they fail once the same species are photographed in the field.
With hundreds to thousands of labels, the regime most science lives in, multimodal LLMs plateau early while vision-only encoders keep improving. On NeWT, 10x more training samples beats 10x more inference compute.