Multi-level integration for predictive inference: mPFC connectivity in action anticipation across representational profiles

Action anticipation has been primarily attributed to the action observation network (AON) and dual visual streams that process action information, yet this sensorimotor account overlooks the inferential processes required when predicting outcomes from incomplete information. How this sensorimotor system interacts with the medial prefrontal cortex (mPFC), a region supporting inference under uncertainty, remains unclear. We investigated the connectivity architecture between mPFC and sensorimotor systems, and how different representational profiles shape this architecture. Novices received either observation-based or execution-based training on table tennis serve anticipation, alongside an untrained control group. Participants completed anticipation tasks using point-light displays and full-body videos during fMRI scanning, with behavioral modeling, neural activation, and dynamic causal modeling (DCM) analyzed. DCM revealed architectural invariance across all groups: mPFC receives inputs from both AON-integrated action representations and direct feature-level information from dual visual streams, enabling flexible predictive inference. However, representational profiles systematically modulated this shared architecture. Perception-dominant representations prioritized contextual integration with enhanced prefrontal engagement when contextual cues were available, while motor-dominant representations emphasized kinematic features with format-invariant extraction, stronger top-down connectivity, and enhanced mPFC recruitment for impoverished stimuli. These findings demonstrate that action anticipation relies on a universal multi-level connectivity architecture enabling cognitive inference in mPFC, with experiential learning shaping information weighting and modulatory dynamics within this shared system.