Neural representational geometries reflect behavioral differences in monkeys and recurrent neural networks

Animals may adopt different strategies to solve the same laboratory task, which can be obscured when pooling behavior and neural data across subjects. We analyzed two male monkeys performing a visually cued rule-based task and found no behavioral performance differences by standard analyses. However, population-level neural activity in dorsolateral prefrontal cortex (PFdl) revealed strikingly different representational geometries: one monkey represented visual cue shape in an abstract format, while the other represented the rule abstractly. Guided by these neural differences, a refined behavioral analysis uncovered corresponding differences in reaction time patterns. Recurrent neural networks (RNNs) trained on the same task reproduced these distinct geometries and reaction-time patterns, with strategy differences correlating with training duration. Our results show that representational geometry can reveal individual strategies and link neural coding formats to behavior, suggesting training history as a mechanistic factor underlying neural and behavioral variability.