HACD-H: Human-AI coevolution theory with 14,700 interaction turns

Jingyi Zhou, Senlin Luo and Haofan Chen present HACD-H, the Human-AI Coevolution Dynamics Framework, in an arXiv submission dated 17 June 2026. The paper builds a formal model of long-term human-AI interaction, tests it on a constructed conversational dataset of approximately 14,700 interaction turns, and reports that social intelligence is negatively correlated with social cognitive energy (r = -0.391, p < 0.001).

What is HACD-H and how was it evaluated?

HACD-H is a formal dynamical model that treats human-AI interaction as a self-organizing social cognitive system, integrating emotional adaptation, relational organization, social memory and personality consistency, and introducing principles such as multi-timescale social cognition, relational attractors, trust basins, developmental phase transitions and social cognitive energy. The authors constructed a conversational dataset with approximately 14,700 interaction turns and developed a theory-driven empirical evaluation framework to test the model's predictions.

The framework explicitly combines component concepts that many prior systems handle separately: emotion modeling, memory retrieval and persona conditioning. The paper frames these elements within dynamics language, naming constructs like relational attractors and trust basins to describe stable relationship patterns and regions of interaction space. The empirical evaluation pairs that theory with the dataset and a formal metric landscape the authors call the social cognitive energy landscape.

What did the study find?

The empirical results show a hierarchy of temporal persistence in social cognition, stable relational attractors, phase-transition-like developmental patterns and a structured social cognitive energy landscape; interaction trajectories exhibit progressive energy reduction and social intelligence shows a significant negative correlation with social cognitive energy (r = -0.391, p < 0.001). These findings support the paper's central claim that social intelligence emerges from long-term coevolution rather than isolated conversational capabilities.

Beyond the correlation, the authors report organized, multi-timescale behavior in their evaluation: some social-cognitive features persist across longer temporal scales than others, and interaction trajectories tend to move toward lower-energy regions of the modeled landscape. The paper phrases these dynamics as relational attractors and developmental phase transitions, suggesting that relationships and social competences stabilize as interactions accumulate.

Why it matters

The HACD-H formulation reframes social conversational behavior as an evolving system rather than a stack of independent modules. If the framework and its reported empirical patterns hold up, designers of conversational agents would need to model long-term interaction dynamics explicitly, not only per-turn quality or isolated persona signals. The observed negative correlation between social intelligence and social cognitive energy (r = -0.391, p < 0.001) gives a concrete, testable link between a theoretical energy landscape and an operational measure of social competence.

This matters for teams building assistants, companions or therapeutic agents because it shifts evaluation priorities: metrics that capture temporal persistence, attractor formation and trajectory-level energy reduction become central to judging whether systems truly develop social intelligence in situ.

What to watch

Look for independent replication of the paper's core empirical signals, in particular the reported correlation r = -0.391 (p < 0.001) between social intelligence and social cognitive energy on other conversational datasets. Also watch for follow-up work that applies HACD-H principles to larger or longitudinal deployment data to confirm the claimed hierarchy of temporal persistence, stable relational attractors and phase-transition-like development.

Details: the paper was submitted to arXiv on 17 June 2026 and lists Jingyi Zhou, Senlin Luo and Haofan Chen as authors. The dataset used in the evaluation contains approximately 14,700 interaction turns.