Agentic Discovery of Multi-Channel Bioacoustic Association Algorithms

Multi-channel call association is an understudied yet essential problem in bioacoustics, serving as a critical building block for localization and density estimation. The complexity of the task stems from the fact that a robust algorithm would involve integrating spectral and temporal features across receivers. As a result, conservation researchers currently rely on time-consuming manual annotation. The challenge of searching this large, multi-constraint algorithmic space makes it an ideal testbed for agentic AI, which has recently shown a strong ability to discover novel algorithms. In this work, we investigate the effectiveness of agentic approaches for discovering generalizable algorithms for multi-channel association. We formulate multi-channel association as an agentic program synthesis problem, where candidate algorithms are discovered through iterative empirical evaluation and feedback. To address the scarcity of labeled real-world data, we rely on physics-inspired simulation and introduce a regime-scheduled evaluation and feedback strategy with warm restarts that expose candidate programs to synthetic conditions of increasing difficulty. The resulting constraint-aware algorithms generalize to unseen simulation regimes and transfer effectively to real whale call datasets, improving association accuracy.