On June 29, 2026, researchers Irena Saracay, Ludwig Schmidt, and Carlos Guestrin published a paper titled "Beyond expert users: agents should help users construct preferences, not just elicit them." The study critiques the conventional assumption that users possess well-defined preferences and emphasizes the need for agents to assist users in developing their preferences through interactive dialogue.
Understanding User Preferences
Agents in recommender systems often operate under the misconception that users have clear and articulate preferences. However, as the authors of the study point out, this assumption is frequently inaccurate. Many users lack the domain-specific knowledge necessary to define their preferences fully. When prompted about their preferences, users may struggle to respond without guidance. This highlights a critical gap in the interaction between users and agents.
The study introduces the CoPref model, which illustrates how users can develop their preferences through interactions with agents. By leveraging examples and explanations, agents can facilitate a deeper understanding of user needs. This approach contrasts sharply with traditional methods that merely seek to clarify existing preferences.
CoShop: A Benchmark for Evaluating Agents
To explore these concepts, the researchers developed CoShop, an interactive benchmark designed to assess how well agents can assist users in constructing their preferences. Within this framework, agents engage in conversations with users who are operating under the CoPref model.
