LatentCRS: A Variational EM Framework for Bridging Semantics and Behavior in LLM-based Conversational Recommendation

The article presents LatentCRS, a model-agnostic Variational EM Framework designed to enhance Conversational Recommender Systems (CRS) by integrating semantic understanding from Large Language Models (LLMs) with user behavioral patterns. By employing a variational expectation-maximization procedure, LatentCRS connects user intent with both semantic and behavioral representations, addressing the existing representation gap and improving recommendation accuracy. Experimental results on real-world datasets indicate that LatentCRS significantly outperforms baseline models, highlighting its potential for practitioners seeking to optimize LLM-based recommendation systems.