关键词: 有限理性, 动态博弈, 心理理论双系统, 递归心理理论, 互动协同

Abstract: Research on theory of mind (ToM) has developed primarily along two relatively independent lines. Dual-process research examines whether individuals process others’ mental states through fast, low-cost automatic processing or slow, resource-demanding reflective processing, thereby addressing how social reasoning is carried out. Research on recursive ToM, by contrast, focuses on the nested levels of mental-state representations and addresses how deeply social reasoning proceeds. Existing accounts have not adequately explained how these two dimensions jointly vary under cognitive resource constraints, why individuals adopt different processing modes and recursive depths across interactive contexts, or how these choices affect decision benefits and interpersonal coordination. To bridge this explanatory gap, we draw on bounded rationality to propose an integrative perspective of resource-adaptive ToM reasoning. This perspective treats processing mode and recursive depth as interrelated but not isomorphic dimensions and conceptualizes ToM performance as a dynamic adjustment of processing investment and recursive level in response to individual resources, task constraints, environmental structure, cognitive costs, and expected benefits.
The central innovation of this perspective is that resource constraints are not reduced to the simple claim that fewer resources necessarily produce poorer performance. Instead, processing mode and recursive depth are assumed to be probabilistically coupled in a context-dependent manner. As recursive depth increases, maintaining and updating multiply nested mental states generally demands more working memory and executive control, making System 2 engagement more likely. Nevertheless, lower-order reasoning may also require controlled processing when a situation is novel or violates expectations, whereas complex strategies may become increasingly automatic through repeated practice. ToM performance therefore depends not only on an individual’s upper capacity limit but also on whether the anticipated marginal benefit justifies further resource investment. This distinction separates relatively stable capacity limits from context-sensitive strategy selection and provides a unified interface for explaining within-person variation in ToM performance across tasks.
Dynamic games provide experimental paradigms for testing this perspective. Unlike static tasks, which primarily assess whether a person possesses a particular form of mental-state understanding, dynamic games incorporate continuous feedback, changing opponent strategies, uncertainty, and benefit-cost trade-offs. They allow cognitive resources, task demands, and opponent complexity to be manipulated, while choices, response times, belief reports, and model parameters track changes in processing mode and recursive depth. On this basis, we formulate three falsifiable propositions.
First, the resource-threshold proposition states that System 2 engagement and sustainable recursive depth are jointly constrained by task demands, available resources, and expected benefits. When additional reasoning is likely to improve outcomes and resources are sufficient, greater task demands may increase both. When computational demands exceed capacity, or the marginal benefit falls below the cognitive cost, both should plateau or decline. Manipulating cognitive load and opponent complexity would permit comparisons among a processing-mode mediation model, a separate resource-constraint model, and a recursive-demand-driven model.
Second, the nonlinear benefit-depth proposition states that recursive depth and decision benefits are not necessarily monotonically related. When recursive computation is costly, opponents are limited in complexity, or environmental noise is high, shallow reasoning may fail to anticipate an opponent, whereas excessive depth may produce cognitive depletion, overfitting, or reduced strategic efficiency. The relationship may therefore be inverted U-shaped, with the optimum varying with opponent complexity, environmental predictability, feedback quality, and incentives. Hierarchical opponent, Bayesian, and dynamic belief-updating models can estimate recursive states and test the quadratic effect of depth and its interactions with environmental structure.
Third, the interaction-coordination observation proposition states that, with continuous feedback and some strategic predictability, partners use each other’s outcomes to update resource investment and reasoning strategies. Their processing-mode shifts, recursive-depth choices, and updating rhythms may consequently become coordinated. Behaviorally, this can be tested through temporal coupling in choices, strategy switches, and response times; computationally, through similarities in latent strategy-state sequences and transition rules. Neurally, after controlling for shared stimuli, joint attention, common task structure, and motor synchrony, inter-brain synchrony (IBS) may serve as an auxiliary indicator of coordination under different latent reasoning states. Behavioral dynamics and computational modeling should provide the primary evidence; IBS offers state-related neural evidence but cannot alone establish a shared processing mode or recursive level.
In summary, the resource-adaptive ToM reasoning perspective integrates processing mode, recursive depth, decision benefits, and interaction coordination within a single explanatory framework. It shifts the central question from whether individuals possess ToM to how they dynamically recruit and regulate ToM under resource constraints. By distinguishing capacity limits, strategic choices, and interpersonal coordination, this perspective also provides an operationalizable, comparable, and falsifiable route for combining behavioral manipulations, computational modeling, and hyperscanning to investigate resource regulation in social reasoning.

Key words: bounded rationality, dynamic games, dual-process accounts of theory of mind, recursive theory of mind, interaction coordination