Towards Fast Domain Adaptation and Fine-Grained User Simulation for Conversational Recommender Systems

Conversational Recommender Systems (CRSs) enhance user experience through multi-turn interactions, yet accurately evaluating their performance remains challenging. Recently, employing Large Language Model (LLM) based user simulators for evaluation becomes an effective approach. However, existing LLM-based user simulators mainly rely on fixed prompts and limited action spaces, which restrict their adaptability to new domains. Additionally, they often lack fine-grained behavior modeling capabilities. To overcome these limitations, we propose AdaptSim, an LLM-based Adaptive domain and automatic prompt tuning User Simulator, featuring capabilities of realistic behavior modeling and diverse style generation. AdaptSim leverages automatic prompt generation and optimization to reduce manual effort in domain adaptation, and introduces an open action generation mechanism to improve flexibility across domains. For response generation, we employ controlled text generation combined with a “think-then-respond” strategy, enabling fine-grained control over user behavior and language style. Extensive experiments across multiple domains demonstrate that AdaptSim generates realistic, diverse, and fine-grained dialogues, thereby facilitating effective evaluation of both the basic capabilities and robustness of CRSs.