Title: Targeted Inquiry: Enhancing Group Elicitation Through Adaptive Multi-Turn LLM Dialogues

Abstract:

Extracting data to clarify unknown group-level characteristics through surveys and collective evaluations necessitates the strategic distribution of finite questioning resources, particularly when facing real-world constraints and incomplete data. While large language models (LLMs) facilitate adaptive, natural language conversations across multiple turns, current elicitation techniques typically fix the respondent pool. Consequently, they fail to adjust participant selection or utilize population structures when dealing with partial or missing responses.

To bridge this gap, we investigate adaptive group elicitation within a multi-round framework. In this setting, an agent dynamically chooses both the questions to pose and the respondents to approach, operating under strict budgets for queries and participation. We introduce a framework grounded in theory that integrates two key components: (i) an LLM-driven expected information gain metric to evaluate potential questions, and (ii) heterogeneous graph neural network propagation. The latter aggregates observed answers and participant traits to estimate missing data and inform respondent selection in each round.

This iterative process identifies a compact, high-value subset of individuals to query, using structured similarity to deduce population-level responses. Evaluations across three real-world opinion datasets demonstrate that our approach consistently enhances the accuracy of population-level response predictions under tight budget constraints. Notably, the method achieved a relative improvement of over 12% on the CES dataset when operating with only a 10% respondent budget.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC