Uncertainty Decomposition for Clarification Seeking in LLM Agents

The paper presents a novel prompt-based uncertainty decomposition approach for large language model (LLM) agents, aimed at enhancing proactive clarification seeking in scenarios with ambiguous task specifications. This method separates action confidence from request uncertainty, addressing limitations of traditional uncertainty frameworks in interactive settings. Evaluation against five LLM backbones, including GPT-5.1 and GLM-4.7, shows significant improvements in clarification performance, with a 73% increase in F1 score on the ALFWorld-Clarification benchmark compared to existing methods, underscoring its potential for practical deployment in LLM applications.