Proactive agents decide when to ask users or verify facts
Clarify the User or Verify the World? Uncertainty Routing for Proactive Agents
Artificial IntelligenceComputation and LanguageMachine Learning
Summary
Sometimes AI assistants need more information to do their tasks well. This paper shows how to help AI agents figure out whether to ask the user a clarifying question or check facts in the environment to reduce uncertainty. The authors create a method called PROUR that breaks down uncertainty into parts caused by unclear user goals and parts caused by missing evidence in the world. PROUR chooses which source to query and learns to generate the best questions, improving success rates and reducing interaction steps compared to earlier methods.
What this means in practice
- •For customer support teams: Improve chatbots that decide when to ask customers clarifying questions versus verifying information in databases to resolve requests efficiently.
- •For virtual assistant developers: Build assistants that proactively identify the best source for information to minimize unnecessary user interruptions and speed up task completion.
Authors
Zhaofeng Li, Xuan Zhang, Xiaokui Xiao, Yang Deng
Abstract
Tool-using LLM agents must decide not only whether additional information is needed, but also which source can resolve the uncertainty. Existing proactive approaches often specialize in either user clarification or environment verification, without explicitly determining the appropriate information source for each decision. We formulate this problem as uncertainty routing among ACT, CLARIFY, and VERIFY, and propose PROUR, a proactive uncertainty routing framework. PROUR decomposes action uncertainty into disagreement across plausible user-goal interpretations, which signals user-side ambiguity, and the entropy remaining within each interpretation, which signals missing world-side evidence. To acquire information from the routed source, a query generator is trained with a mode-conditioned information-gain reward, targeting user-goal identification under CLARIFY and next-action identification under VERIFY. On $τ$-bench, PROUR achieves 28.17% average success rate across retail and airline, outperforming the strongest prior method by 4.57% while using 2.17 fewer interaction steps. The learned policy further generalizes to stronger task agents and transactional domains of $τ^3$-bench without retraining, demonstrating the benefit of source-aligned uncertainty resolution for proactive agents.