关键词: 刻板印象, 社会泛化, 强化学习, 贝叶斯理论, 计算建模

Abstract: Stereotypes—generalized beliefs that members of a social group tend to possess certain traits—are central to human social cognition. Traditional accounts have attributed stereotyping to motivational needs or cognitive resource constraints. However, these perspectives remain largely descriptive, and offer limited insight into the computational mechanisms governing how stereotypes are acquired, updated, and applied. Adopting a computational cognitive neuroscience perspective, this review integrates two dominant formal frameworks, reinforcement learning (RL) and Bayesian theory, to provide a unified, process-oriented account of stereotype formation and generalization.
With respect to social learning, we first examine how Bayesian structure learning enables the brain to infer latent group categories from observable social data, estimating the posterior probability of group assignments. Building on this, we delineate two functionally distinct pathways through which group-trait associations are established. The experiential pathway encompasses Pavlovian conditioning, instrumental learning, and observational learning. These operate primarily through RL mechanisms—particularly value updating, formalized as V(group) ← V(group) + αδ—to form associative representations. These automatic links between group labels and affective valence that influence behavior without requiring propositional evaluation. The linguistic pathway, by contrast, supports the acquisition and transmission of propositional representations (truth-apt beliefs, e.g., “group G possesses trait T”) through symbolic communication. The Bayesian framework provides a natural computational account for this type of representation, formalizing stereotypic beliefs as conditional probabilities, P(trait | group), that are updated in accordance with Bayes’ rule. Recent empirical evidence confirms that human stereotype judgments closely approximate Bayesian posteriors, and methodological advances further enable the direct quantification of propositional stereotype content in natural language corpora.
We further analyze how these associations are maintained or revised. Prediction error (PE) constitutes the core learning signal driving stereotype updating: counter-stereotypical information generates heightened PE, accelerating the revision of existing associations. However, pre-existing stereotypes resist updating through biased priors and asymmetric learning rates; individuals tend to learn more rapidly from evidence that confirms their stereotypic expectations than from disconfirming evidence. We identify the explore-exploit dilemma as a critical convergence point of the two frameworks: Bayesian priors shape RL-based exploration strategies, producing biased information sampling that in turn reinforces existing beliefs, creating a self-perpetuating cycle. This computational mechanism offers a principled explanation for why stereotypes resistant to change, even in the absence of motivational bias or cognitive limitations.
At the level of social generalization, we focus on the computational challenge of group identification: how the brain categorizes a novel individual into a known group to deploy learned associations. We distinguish two routes of similarity-based matching. Perceptual generalization operates on directly observable physical features in early visual processing, whereas functional generalization relies on abstract role and category information mediated by latent variable inference. The dynamic interaction between bottom-up perceptual signals and top-down conceptual knowledge, coordinated by the dorsomedial prefrontal cortex (dmPFC), means that group identification is not a serial process but a continuous integration of multiple cue types. Once group membership is established, both associative and propositional pathways are engaged during retrieval, with the former enabling rapid, resource-efficient behavioral responses and the latter supporting more deliberate trait inference.
Finally, we identify three promising directions for future research: (1) investigating how relational cues, such an individual’s position within a social network, contribute to stereotype generalization beyond feature-based similarity; (2) leveraging the social cognitive map framework, supported by recent evidence of distance and grid-like coding in the hippocampal-entorhinal system, to model the integrated representation of multi-dimensional social information; and (3) employing large language models (LLMs) to simulate linguistic transmission, in order to examine how minor initial biases become amplified and consolidated into shared cultural stereotypes through iterative communication.

Key words: stereotype, social generalization, reinforcement learning, Bayesian theory, computational modeling