When and How to Ask: Dynamic Preference Elicitation Strategies for Conversational Recommendation

Conversational Recommender Systems (CRSs) aim to provide personalized recommendations through natural language dialogue. A critical challenge is determining when and how to elicit user preferences. While many systems rely on static, uniform strategies, this paper investigates whether preference elicitation should be dynamic and stage-aware. The authors conduct a preliminary user study to assess the effectiveness of different elicitation strategies and introduce InPE, a new dataset featuring fine-grained annotations for elicitation necessity and strategy selection. Based on this, they propose COPE (Conversational Preference Elicitation via Mixture of Experts), an architecture that dynamically selects strategies based on the current dialogue state.

The authors demonstrate that user preferences evolve from abstract to concrete throughout a conversation, and that no single elicitation strategy is universally optimal. Their empirical analysis reveals a clear pattern: attribute-based inquiries (e.g., asking for genres) are most effective in the early stages of a dialogue, whereas item-based strategies (e.g., asking users to compare specific items) become more effective as the user's preferences refine. The proposed COPE model, which uses a Mixture-of-Experts approach to select the appropriate strategy at each turn, consistently outperforms strong LLM-based baselines in offline evaluations.

This research provides a systematic framework for moving beyond static, one-size-fits-all conversational interfaces. By identifying that the optimal way to ask for user feedback changes as a conversation progresses, the authors offer a roadmap for building more proactive and efficient CRSs. The release of the InPE dataset provides a valuable resource for future research into context-aware dialogue management and strategy-driven recommendation systems.