Delusions Emerge and Persist due to Generative Model Reorganisation: Insights from Hybrid Predictive Coding

Hybrid predictive coding simulations show that contextually inadequate belief initialization plus excessive certainty reorganizes the generative model, enabling delusional beliefs to persist and resist contradiction without increased global prediction error.

• Why it matters: Existing predictive-coding accounts of delusions are largely untested with formal models, leaving gaps in how thematic recurrence and imperviousness to contradictory evidence arise computationally.
• What they did: The study used simulation experiments with hybrid predictive coding to model delusional beliefs, varying belief initialization and certainty to test effects on generative-model structure and prediction error.
• The result: The reorganization maintains thematically stable, internally consistent delusional beliefs that remain impervious to contradictory evidence even after certainty returns to normal, suggesting generative models learned under atypical conditions.