The work compares three search styles across simulated semantic spaces that vary in how tightly words cluster and how spread-out clusters are. The findings show that different search strategies succeed under different structural conditions. A strategy that narrows attention to a local area and adapts as it goes keeps producing human-like patterns across many realistic semantic layouts. By contrast, search driven by long-range random jumps or by uniform random exploration falters when the organization of meaning changes, producing response patterns that depart from what people do.
For anyone curious about how memory, language, and decision rules interact, this paper points to flexible, adaptive control as central to human-like retrieval. The next step is to ask how these adaptive rules develop, how they vary between individuals, and how they might be supported to boost learning and inclusion across diverse speakers and learners. Read the full article to see the models, simulations, and what this implies for designing tools that align with how our minds actually navigate meaning.
Abstract
Verbal fluency tasks reveal clustering and switching patterns traditionally explained by strategic search or stochastic processes like Lévy or random walks. However, previous comparisons ignored how search processes interact with semantic structure, leaving unclear whether model performance reflects strategic mechanisms or fortuitous alignment with semantic organization. This study developed and validated a novel Area Restricted Search (ARS) agent-based model of semantic memory retrieval, then systematically compared it against Lévy Walk (LW) and Random Walk (RW) models to investigate when different search mechanisms succeed under varying structural conditions. The model implements incremental decision-making based on local information, without predetermined switching points or complete semantic space access. Semantic structure parameters were treated as free variables during optimization, allowing examination of process–structure interactions across diverse configurations. Performance was evaluated against 50 participants across three semantic categories using clustering, switching, and temporal variables. Two simulations examined model fit and adaptability to varying semantic structures. Different mechanisms require distinct semantic configurations: ARS performed well in moderate clustering, LW in sparse arrangements, and RW under dense clustering, but RW generated response distributions different from participants. However, when semantic density was constrained while varying cluster dispersion, ARS maintained human-like performance across multiple configurations, while LW showed limited flexibility, and RW consistently failed to get close to participants’ response distributions. These findings show that human-like semantic memory retrieval across diverse contexts requires strategic mechanisms capable of dynamic adaptation to varying semantic organizations, rather than universal superiority of any single approach or of models based on context-independent stochastic processes.
