Language models reveal and simulate psychosis-associated disruptions to semantic foraging in narrative recollections

Retrieval of semantic information from memory may operate as a foraging process, wherebyexploitation within clusters of related information is punctuated by strategic exploratory jumpsbetween clusters. Studies of verbal fluency tasks, in which a sequence of individual concepts(e.g., animals) is retrieved, indicate that semantic foraging processes are disrupted inpsychosis. However, it is unclear if these disruptions extend to natural speech. Furthermore,mechanisms underlying these disruptions are unknown. To explore these questions, in Study 1,we combined the language model BERT with Discourse Atom Topic Modelling to analyze how93 participants with schizophrenia-spectrum disorders semantically forage for topics as theynarrate their life stories. Negative symptom severity was associated with distinct narrativecontent, reduced topical focus, and more frequent switching between clusters. Positivesymptom severity was associated with greater semantic distances between clusters. Cognitivesymptom severity was associated with distinct narrative content, less frequent switching, fewertopics retrieved per cluster, and lower between-cluster semantic distances. To explore possiblemechanisms behind psychosis-related foraging disruptions, in Study 2 we manipulated the largelanguage model (LLM) Llama-3.2-3b by injecting stochastic noise or directional “steeringvectors”—which model the influence of symptoms on narratives—into different layers of the LLMduring life-story generation. Depending upon the targeted layers, these manipulationsreproduced several observed psychosis-related foraging disruptions. By using LLMs within asemantic foraging framework to analyze and simulate narratives, this study extendsobservations from verbal fluency tasks into long-form naturalistic speech and offers mechanisticinsights into psychosis-related speech disturbances.