Autonomous Field Guides: Sparse Autoencoders and Heuristic Ranking for Dichotomous Key Learning

Morphological traits used to distinguish between species, also known as dichotomous keys, are widely used by ecologists and researchers. Despite their utility, they are hard to extract or discover at scale due to their qualitative nature. We test whether sparse autoencoders (SAEs) trained on vision foundation model embeddings can recover dichotomous keys automatically, using Heliconius butterfly mimicry as a case study. We train Matryoshka SAEs on DINOv3 patch embeddings and rank the resulting features using two simple heuristics: selectivity, a classifier-free score based on the Mann-Whitney U statistic, and Wasserstein distance between activation distributions. Across three mimic pairs with documented dichotomous keys, the top-ranked SAE features consistently correspond to the known dichotomous keys. We also investigate two failure cases: feature splitting, and a mimic pair for which our method does not surface a convincing candidate trait. Overall, our results suggest SAE features can serve as reasonable, quantified candidates for automatic dichotomous key discovery.