This Dagstuhl Seminar aims to bring together researchers from economics, theoretical computer science, and artificial intelligence working on collective decision making. Much of computational social choice (COMSOC) assumes that agents’ preferences are fully known and focuses on how these preferences should be aggregated into a good collective decision. This seminar starts one step earlier: preferences must first be elicited from agents, and the possibilities and limitations of elicitation fundamentally shape how collective decisions can be made.

Preference elicitation has been studied in many areas, but often within individual problem domains and largely in isolation. The seminar aims to bring these different strands together and to establish preference elicitation and discovery as a central research frontier for computational social choice. While elicitation has always been foundational to the field, recent developments have highlighted its importance as a research topic in its own right. Indeed, the integration of generative AI is making practical collective decision-making processes increasingly open-ended, allowing final outcomes to be chosen from a vast and often unstructured set of alternatives, e.g., natural language outputs. Similarly, when collective input is used to guide the alignment of AI systems, the space of alternatives effectively becomes the set of all possible model configurations.

The seminar will be structured around three central questions:

• How should we elicit preferences? Agents may be asked to rank alternatives, provide pairwise comparisons, assign numerical scores, or select subsets of alternatives. Which elicitation formats are most informative? In combinatorial domains, which structural assumptions make elicitation tractable while preserving sufficient expressiveness?
• How much preference information do we need? In many settings, eliciting complete preferences is costly or outright impossible. How much information is sufficient to make high-quality collective decisions? How can we make elicitation adaptive so that we elicit only the most relevant information?
• What are we eliciting? Following its roots in economics, COMSOC often models preferences as fixed and exogenously given. In practice, however, preferences may be uncertain, context-dependent, shaped by the elicitation process itself, or evolving over time. How should we elicit and aggregate such preferences, and what does it even mean for a decision to reflect preferences that can change?

By examining these questions across the breadth of computational social choice, while placing particular emphasis on emerging interactions with AI, the seminar aims to connect previously separate lines of research, identify common conceptual and methodological challenges, and develop a research agenda for collective decision making under limited, uncertain, and evolving preference information.

Creative Commons BY 4.0

Niclas Boehmer, Paul Gölz, Ayumi Igarashi, and Marcus Pivato