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

A language-model-inspired species distribution framework trained on ~27,000 reef fish surveys generates multiple coherent community configurations under identical environmental conditions, including coral- and macroalgal-associated regimes.

• Why it matters: Species distribution models often treat species as independent, missing dependencies that shape how ecological communities assemble, persist, or collapse under rapid global change.
• What they did: The approach sequentially predicts each species and its biomass conditional on previously assembled species, learning a dependency structure that yields core, bridging, and satellite species positions.
• The result: It improves species occurrence and biomass predictions versus species-by-species XGBoost and multispecies deep-learning MLP baselines, while naturally revealing multistability and hierarchical assembly patterns.