The abstract is read at the publisher; the summary is JClub's.
A multi-module graph convolutional network achieves zero-shot transfer in an embodied dynamic pursuit task, revising mid-course target choice (change-of-mind) without additional training.
- Why it matters: Flexible control in biological agents adapts to new goals and environments without extra training, but the minimal computational motifs and their neural basis are unclear.
- What they did: The model implements relational structure, spotlight attention, and affordance computation to integrate inter-entity relations, reward, and action feasibility in real time, and primate dorsal anterior cingulate cortex recordings test neural signatures.
- The result: The network transfers across novel pursuit scenarios with different physics, target properties, and interaction policies, and neural population dynamics link the three constructs to the observed control behavior.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.