The Computational and Neural Basis of Zero-Shot Control in Dynamic Pursuit

Biological agents flexibly adapt their behavior to novel goals and environmental demands without additional training, yet the computational principles enabling such control remain unclear. Here, we propose that three cognitive constructs constitute minimal computational motifs for such flexible control: relational structure, spotlight attention, and affordance computation. We examine whether these constructs underpin flexible control in an embodied dynamic pursuit task that requires continuous integration of inter-entity relations, reward, and action feasibility, making it a suitable testbed for real-time control. By implementing these constructs within a multi-module graph convolutional network, we show that the model achieves zero-shot transfer across novel pursuit scenarios that vary in physics, target properties, and interaction policies such as fleeing or chasing, without additional training. Although not explicitly trained to do so, the model also exhibits change-of-mind (CoM) behavior, or mid-course target revision, a hallmark of flexible control exhibited by biological agents. Neural recordings from the primate dorsal anterior cingulate cortex revealed population-level signatures that link these constructs to neural dynamics, providing biological support for the proposed computational architecture.