A Behavioral Trait Leaks into Preferences: Diagnosing Trait Interference in LLM User Simulators

LLM user simulators suffer from trait interference where high activity levels distort preference boundaries, forcing interactions with mismatched items to sustain browsing. This bias inflates satisfaction scores based on page counts rather than true taste alignment. The proposed PQA method mitigates this by using personalized quality anchors to guide simulation.