Dynamic Representational Synchrony through Collective Predictive Coding: A Computational Model of Parent-Infant Homeostatic Co-Regulation

Social synchrony, reflected in neural, behavioral, and physiological coupling, is fundamental to human interaction. In parent-infant relationships, early regulation of bodily states and emotional development depend on effective information exchange. Prior research highlights inter-brain synchrony (IBS) via hyperscanning, but models often assume full model sharing, neglecting real-world constraints like sensory limitations and knowledge asymmetry. The predictive coding framework, especially CPC, offers a promising avenue to model shared beliefs via symbols, yet its application to dyadic, asymmetric scenarios remains limited. Recent advances incorporate Bayesian inference and multi-agent systems, but lack mechanisms for asynchronous, local belief alignment. This study addresses this gap by proposing a model that captures how limited perception and asymmetric knowledge can still produce rapid, dynamic belief synchronization, crucial for early social development.

The core challenge is modeling how two individuals with different sensory modalities and knowledge structures can align their internal beliefs in real-time, despite limited and asymmetric information. Traditional models rely on complete model sharing, which is unrealistic in natural settings like parent-infant interactions. The question is how belief synchronization can occur through local, symbol-mediated interactions without requiring full model overlap. This involves understanding mechanisms for asynchronous belief updates, stochastic symbol acceptance, and dynamic learning of model parameters. Addressing this problem is vital for explaining early social cognition, emotion regulation, and the neural basis of interpersonal synchrony, especially under conditions of sensory and informational constraints.

The key innovations include: 1) integrating CPC principles with POMDP to model belief updates under sensory asymmetry; 2) employing MHNG for stochastic symbol acceptance, enabling belief fusion without full model sharing; 3) demonstrating rapid, early belief alignment preceding model convergence, highlighting the role of local interactions; 4) implementing online learning of generative matrices A and B, capturing knowledge asymmetry dynamically. These elements collectively enable a realistic simulation of parent-infant interaction, emphasizing how limited perception and stochastic communication can produce robust, sustained belief synchronization, advancing the understanding of social cognition mechanisms.

• �� Model two agents (parent and infant) as POMDPs with asymmetric knowledge: parent knows state transition matrix B, infant knows sensory matrix A. • Each agent maintains a latent belief zX, updated via Bayesian inference based on observations iX and shared symbol w. • Symbols w are proposed by the speaker, accepted or rejected by the listener based on MHNG acceptance probability, computed solely from the listener’s model. • The shared symbol constrains posterior beliefs, promoting belief alignment. • Both agents learn matrices A and B online through Dirichlet distributions, updating parameters after each interaction. • The model simulates visceral states in a 6×6 grid, with actions like eat, sleep, and temperature regulation, affecting internal states and observations. • Experiments compare MHNG with control conditions, measuring visceral regulation, KL divergence of matrices, and JSD of beliefs across iterations.

Simulations involved two agents interacting over 1000 iterations within a 6×6 visceral state grid, representing energy and temperature. The agents performed actions to regulate visceral states, with their belief states updated via Bayesian inference. The primary comparison was between MHNG-mediated interaction and one-sided control conditions (A-led and B-led). Metrics included the prior preference C, KL divergence of matrices A and B, and Jensen-Shannon divergence of latent beliefs. The experiments assessed how quickly and stably beliefs aligned, and how effectively visceral states were regulated. Multiple runs tested robustness, with parameters tuned to reflect realistic sensory asymmetries. Results consistently showed faster belief synchronization and better regulation under MHNG, validating the model’s efficacy.

The model achieved rapid belief alignment, with JSD dropping below 0.1 within 20 iterations, significantly faster than control conditions. Visceral regulation was more effective in MHNG, maintaining infant states near high-priority zones (p<0.01). KL divergence of matrices A and B decreased steadily, approaching true values within 500 iterations, indicating successful online learning. Transient spikes in belief divergence occurred during rare visceral state transitions but were quickly corrected, demonstrating the system’s resilience. Synchronization preceded model convergence, emphasizing the role of local belief constraints. These findings confirm that asynchronous, symbol-mediated interactions can produce sustained, dynamic belief coupling even with asymmetric knowledge.

The model informs early intervention strategies for developmental disorders by elucidating mechanisms of belief synchronization. It guides the design of social robots capable of adaptive, belief-based interactions under sensory limitations. Additionally, it supports the development of AI systems for social cognition, emphasizing local, belief-driven coordination. Broader applications include enhancing human-computer interaction, improving remote social therapy, and understanding neural mechanisms underlying social bonding.

The simplified 2D visceral state model limits biological realism; real physiological states are high-dimensional and continuous. Static prior preferences and interpretation matrices restrict modeling of social meaning evolution. The absence of neural or anatomical detail constrains biological interpretability; future work should incorporate neural data. The computational cost of online belief updates may limit scalability to complex scenarios. Addressing these limitations requires richer representations, adaptive symbol semantics, and integration with neural mechanisms for more accurate modeling of real-world social interactions.