Behavior-Driven Marine Larval Dispersal and Settlement with AI Agent-Based Modeling

Abstract

Larval dispersal models are central to mapping and predicting ichthyoplankton dynamics in the ocean, yet despite decades of refinement they remain fundamentally limited by their ability to represent adaptive behaviors, relying instead on static trait parameterizations. This deficiency constrains our capacity to design effective restoration and mitigation strategies in an increasingly stressed ocean. SWARM (Simulating Waterborne Agent Routes for Marine connectivity) overcomes this barrier by integrating Large Language Model (LLM)-based behavioral agents with a standard biophysical model to simulate active decision-making during the pelagic larval stage. In both idealized and realistic conditions focusing on Red Snapper larvae in the Gulf of Mexico, agents develop adaptive behaviors that improve settlement and generate explainable vertical distribution patterns. SWARM demonstrates that LLMs can overcome long-standing limitations in dispersal modelling by explicitly representing behavioral drivers of movement, opening new pathways for predicting connectivity and designing effective marine-ecosystem restoration.