New Method Enables LLM Agents to Seek Clarification by Decomposing Uncertainty

The classical framework of aleatoric and epistemic uncertainty is often insufficient for interactive Large Language Model (LLM) agents, especially when they need to proactively seek clarification or build shared mental models. This paper responds to the call for more nuanced, decomposed, and communicable uncertainty representations that can unlock advanced agent capabilities. The practical constraints of black-box APIs, latency budgets, and lack of labeled data often limit viable methods to prompt-based estimation. The authors propose a straightforward prompt-based decomposition that distinctly separates an agent's confidence in its intended action from its uncertainty regarding the task request itself. This distinction is crucial because it allows the LLM agent to identify when a task specification is ambiguous and, consequently, to ask for clarification rather than proceeding with an uncertain understanding. To evaluate this new approach, two clarification-augmented benchmarks, WebShop-Clarification and ALFWorld-Clarification, were developed, with 50% of tasks deliberately underspecified. The proposed decomposition was systematically compared against existing methods like ReAct+UE and Uncertainty-Aware Memory (UAM) across five different LLM backbones. The results show significant improvements: the new decomposition boosts clarification F1 on ALFWorld-Clarification by 73% over ReAct+UE and 36% over UAM, and leads on most backbones across both new benchmarks, demonstrating its generalizability and effectiveness in enabling agents to seek clarification.