COPE framework optimizes when and how conversational AI asks questions

Conversational Recommender Systems (CRSs) rely on multi-turn dialogue to understand evolving user preferences, but the timing and selection of preference elicitation strategies have been underexplored. Existing studies often use static attribute-based questions, ignoring the potential of item-based elicitation and how strategies should shift across dialogue stages. In this work, the authors conduct a systematic investigation with a stage-aware perspective, providing empirical evidence that optimal strategies are both stage-dependent and context-sensitive. Attribute-based questions work best in early stages to broadly narrow preferences, while item-based strategies (e.g., asking about specific movies) become more effective as preferences refine later in the conversation.

To support this paradigm, the team introduces InPE, a dataset enriched with fine-grained annotations for elicitation necessity and strategy selection, and proposes COPE (COnversational Preference Elicitation via Mixture of Experts), a novel architecture that dynamically selects the best strategy per turn using a mixture-of-experts approach. Extensive offline evaluation on InPE shows that context-aware strategy selection significantly improves recommendation quality. The predicted strategies also reveal consistent stage-wise tendencies, confirming common interaction patterns. The dataset is publicly available. The work was accepted at SIGIR 2026 and highlights a shift toward more adaptive, dynamic conversational recommendations.