A language-model approach to species distribution reveals alternative stable states.

Ecological communities are reorganizing rapidly under accelerating global change, yet predicting how species assemble, persist, or collapse remains a major challenge.1,2,3,4 Most species distribution models treat species as independent entities, overlooking the dependencies that structure ecological communities. Here, we introduce a language-model-inspired approach5,6 to species distribution that generates ecological communities sequentially, predicting each species and its biomass conditional on previously assembled species. Trained on ∼27,000 reef fish surveys,7 the generative framework reveals multiple coherent community configurations emerging under identical environmental conditions.8,9 For instance, the model can generate both coral-associated and macroalgal-associated fish communities, consistent with alternative ecological regimes described in these ecosystems. The model also uncovers a hierarchy of emergent species positions within the learned dependency structure, including core species strongly associated with particular environmental conditions, bridging species linking alternative configurations, and satellite species occupying more peripheral positions. Despite this richer representation of ecological communities, the model also improves predictions of species occurrence and biomass relative to a species-by-species machine-learning baseline (XGBoost10) and a multispecies deep-learning baseline (MLP11). Together, these results show that ecological communities can be learned and generated as a structured syntax, from which multistability, directional assembly patterns, and hierarchical species positions naturally emerge.