Adaptive Tools in Societies of Strangers: An Agent-Based Model of Abstract Category Emergence

Humans routinely predict each other's behavior using abstract category labels, such as "honest" or "reliable." However, the level of abstraction a population employs is not fixed; cross-cultural evidence indicates that individuals in larger, more anonymous societies rely on coarser, more generalized categories. How do the structural properties of a society shape the cognitive representations of its members? I propose that abstract psychological categories emerge as an adaptive tool, jointly determined by the structural properties of society and the cognitive learning dynamics of individual agents, which I implement through reinforce- ment learning in this paper. To test this, I develop an agent-based model in which agents must predict target behaviors. I formalize a society where three distinct constraints scale naturally with population size: data sparsity (exploration costs), cognitive load (computation costs), and common-ground building (coordination costs). By tracking the evolution of the agents' language, encoded as bitstrings, I find that small populations sustain highly specific, context-dependent vocabularies. In contrast, as population scale increases, the compounding pressure of these scale-induced costs forces agents to shift from a model-free mode of reasoning, which relies on instance-by-instance tracking, toward model-based gen- eralization. My findings demonstrate that abstract categories function as adaptive tools dictated by the structural imperatives of "societies of strangers." Moreover, distinct from centralized institutions, language can also function as a decentralized tool for navigating the cognitive and logistical burden of sociality.