Rational Clarification by Assistive Agents via Value-of-Information Reasoning

Users of language-based assistive agents often make ambiguous requests. In response, an assistant can either directly act on its interpretation of the request, or ask a clarifying question. Which option is the most helpful? A common approach is to ask questions that minimize the assistant's uncertainty about the user's intent until a threshold or budget is reached. However, this neglects the impact of the uncertainty reduction on downstream task performance, the costs of asking a question versus acting immediately, and the possibility that users may provide corrections without being asked. To rationally navigate these trade-offs, we introduce Rational Enquiry via Value-of-Information Reasoning (REVOIR). REVOIR makes clarification decisions via inference-time reasoning about the value-of-information (VoI) of a question, which captures the expected improvement in task reward due to the answer received. In two assistive tasks --- ambiguous question answering (CondAmbigQA) and preference-aligned household task planning (ADAPT) --- we show that REVOIR achieves greater success with fewer questions than approaches based on prompting, fine-tuning, or maximizing information gain, improving preference satisfaction on ADAPT by 13-15% over a fine-tuned clarification policy while requiring no training and asking five times fewer questions. Furthermore, when the assistant can receive low-cost user corrections without asking questions, REVOIR naturally infers that explicit clarification is not always warranted, demonstrating the adaptivity of our decision-theoretic approach. We also observe that scaling reasoning efforts at inference time does not inherently translate to better performance in interactive environments, while leading to prompt-scaffolded assistants asking fewer questions unilaterally.