Abstract: Real-world joint actions inherently possess a hierarchical structure, requiring co-actors to coordinate both concrete, low-level sensorimotor mappings and abstract, high-level task rules. However, prevailing theories—such as co-representation and referential coding—conceptualize self-other integration as a static, one-shot mapping, largely confining investigations to a single hierarchical level. This static, single-level approach fails to capture the dynamic information accumulation and cross-level interactions essential for complex interpersonal coordination. To address these critical gaps, the current study adopts a “hierarchical task representation” perspective, aiming to achieve a paradigm shift from “static single-level representation” to “dynamic hierarchical prediction.” The core innovations span three dimensions: phenomenological framing, computational mechanism modeling, and the identification of asymmetric social modulation effects.
First, we introduce a hierarchically nested paradigm to reveal the dissociation and bidirectional interaction of self-other integration. Previous studies have typically examined high-level (e.g., joint statistical learning) and low-level (e.g., joint Simon) tasks in isolation, leaving their dynamic interplay unexplored. By embedding statistical regularities within a joint Simon task, we simultaneously capture both levels within a single paradigm. Across five experiments combining behavioral measures and EEG hyperscanning, we provide systematic evidence that high-level and low-level integrations exhibit distinct temporal trajectories. Crucially, we reveal bidirectional interactions: high-level regularity acquisition exerts top-down modulation on low-level spatial compatibility, while low-level conflict influences high-level rule updating. Methodologically, we dissociate their neural signatures using time-resolved ERPs (frontal-central N2 for low-level conflict vs. parietal-occipital P3 for high-level prediction) and inter-brain synchrony, thereby overcoming the signal overlap inherent in behavioral reaction times.
Second, we construct a Bayesian computational model to quantify the cognitive mechanisms underlying level specificity. Transcending the descriptive limitations of prior theories, we formalize self-other integration as a dynamic Bayesian updating process. We propose that the observed level specificity arises from distinct modes of prior information accumulation, operationalized by a weight parameter (η). For high-level abstract tasks characterized by high uncertainty, the cognitive system adopts a "data-driven fast update" strategy (η>1), amplifying new observational evidence to rapidly revise prior beliefs. Conversely, for low-level concrete tasks with stable mappings, a “prior stability maintenance” strategy (0<η<1) is employed, diminishing the impact of new input to preserve behavioral coherence. Furthermore, to model cross-level interactions, we introduce regulation parameters (α and β) that mathematically define how the posterior distribution at one level dynamically modulates the prior of the other. By manipulating task difficulty and working memory load, and by linking these computational parameters to frontoparietal theta connectivity (a neural marker of evidence accumulation), we furnish computable and biologically verifiable evidence for the model's validity. This framework also reconciles the long-standing debate between co-representation and referential coding, re-conceptualizing them as top-down “social priors” and bottom-up “sensory inputs” within a unified predictive processing stream.
Third, we identify the asymmetric modulation effects of interpersonal common ground on hierarchical integration. While prior literature assumes a uniform facilitating role for common ground, we reveal its level-specific paradoxical effects. Across four experiments manipulating action-based and trait-based common ground, we demonstrate that action constraints (e.g., restricting response hands) specifically impair low-level spatial integration while sparing high-level abstract rule integration. More strikingly, we discover a counter-intuitive effect of trait common ground: similarity in prosocial traits (e.g., extraversion, agreeableness) facilitates low-level integration by enhancing behavioral predictability, yet paradoxically hinders high-level integration. This occurs because trait similarity induces a social projection bias, leading individuals to over-rely on self-knowledge and substitute their own abstract task representations for the partner’s, thereby diminishing the distinct encoding of the other’s high-level rules. Furthermore, by testing the robustness of our Bayesian model across these varying interpersonal contexts, we validate the generalizability of the proposed computational mechanisms.
In summary, this research constructs a comprehensive “phenomenon-mechanism-modulation” framework for hierarchical self-other integration. By integrating behavioral experiments, EEG hyperscanning, and cognitive computational modeling, we deliver a computable and verifiable theoretical tool. These innovations not only fundamentally advance the psychological understanding of dynamic interpersonal coordination but also provide critical theoretical inspiration and architectural parameters for the development of adaptive, collaborative multi-agent systems and human-machine teams.

Key words: joint action, self-other integration, hierarchical task representation, Bayesian model