MultiDems will also host an associated shared task on Text Summarization for Online Citizen Consultations. Online consultations enable large-scale democratic participation by collecting citizens' proposals and their votes on public issues (e.g., “We should incorporate the right to clean air into our constitution.”). Turning these contributions into concise, accessible summaries is essential for informing public deliberation and decision-making. Although large language models offer promising opportunities for automating this process, they also raise important questions: Whose proposals are represented? Are minority viewpoints adequately reflected? Does the summary accurately convey both the meaning of proposals and their level of public support?
to participate in the shared task: Register using this form https://forms.gle/FHe4SGvofhYWAtwZA, you’ll be invited to join our google group to stay informed on the Competition launch.
Participants will develop systems that produce concise, coherent summaries of collections of citizen proposals and associated voting data. Each consultation will contain proposals expressed in natural language, a citizen–proposal voting matrix, and the language of the consultation. Systems must generate natural-language summaries in which statements are accompanied by references to the original proposal IDs supporting them. This source-grounded format enables evaluation of both the selection of proposals and their representation in the resulting text, while remaining agnostic to the summarization method.
The task introduces a multilingual benchmark based on real-world citizen consultation data across multiple European countries. Beyond conventional summary quality, generated summaries will be evaluated using democratically grounded criteria, including proportional representation of citizens’ preferences, preservation of consensus and disagreement, semantic fidelity to the original proposals, balance in the representation of arguments, and coverage of the opinion space.
Systems will be evaluated separately on each dimension, alongside an aggregate score. Importantly, trade-offs between metrics are expected and constitute a central motivation for this shared task: different principles of democratic representation may conflict, and optimizing one criterion may compromise another. The task therefore aims not only to identify effective summarization methods, but also to investigate how alternative approaches navigate these trade-offs.
The competition will be hosted on CodaBench. Participation is open to researchers and practitioners from NLP/ML/AI, computational social choice, political science, and related fields. Teams may participate in the competition without submitting a paper. Participants are also encouraged to submit archival system-description papers presenting their approaches, experimental results, and analyses of democratic evaluation criteria. Selected teams, including those performing particularly well across different evaluation dimensions, will be invited to present their work at the workshop.
Data release and Opening of the Codabench competition: early November 2026
Codabench competition submission deadline: December 3, 2026
System-description paper submission deadline: December 23, 2026
Notification of acceptance: January 22, 2027