Neural and behavioral evidence for hierarchical and counterfactual reasoning in non-human primates

When faced with challenging multi-step decisions, humans are unable to evaluate all branches of the decision tree simultaneously and instead rely on piecemeal hierarchical and counterfactual reasoning strategies. To investigate the neural basis of these cognitive strategies, we developed a challenging multi-step decision task for monkeys in which they had to track an invisible ball through a maze using partial cues about the times when the ball makes turns within the maze. Animals were able to use temporal cues to infer the ball’s exit point. To evaluate the animals’ decision strategy, we compared their behavior to models implementing different inference strategies including (1) optimal maximum-likelihood strategy, (2) hierarchical strategy that handles decisions sequentially, (3) postdictive strategy that conditions early decisions on late evidence, and (4) hierarchical strategy with potential revisions using counterfactuals. Comparing behavioral responses to these cognitive models revealed that monkeys, like humans, solved the task using a combination of hierarchical and counterfactual strategies. Several independent behavioral tests involving out-of-distribution maze geometries, targeted temporal perturbations, and post-error eye movements provided additional evidence that animals solved the task flexibly using these strategies. To investigate the underlying neural computations, we recorded simultaneously from large populations of single neurons in the anterior cingulate (ACC) and dorsomedial frontal cortex (DMFC). Single neurons in both areas had highly heterogeneous response profiles and formed structured low-dimensional activity patterns across the population. Further analysis of neural response dynamics with respect to the temporal cues provided clear evidence of hierarchical information processing in low uncertainty conditions and counterfactual processing in high uncertainty conditions when the late sensory cues could help revise potentially incorrect earlier decisions. Together, these results indicate that monkeys, like humans, rely on rapid hierarchical and counterfactual decision strategies and highlight a potential role of ACC and DMFC in supporting the underlying computations.