UXSim: Towards a Hybrid User Search Simulation
Existing approaches struggle to accurately capture the dynamic and personalized nature of user behavior in interactive search, as static agents or purely large language model (LLM)-based agents often lack grounding in real-world data. This work proposes UXSim, a novel framework that uniquely integrates traditional data-driven user simulators with adaptive LLM agents. By leveraging real user interaction logs to constrain and guide the LLM’s reasoning process, UXSim ensures behavioral fidelity while enhancing the interpretability of the underlying cognitive mechanisms. The resulting simulation not only achieves greater accuracy and dynamism in modeling user search behaviors but also enables verifiable explanations of simulated outcomes, thereby bridging the gap between data realism and generative flexibility in user modeling.