Uncertainty-Aware Clarification in LLM Agents with Information Gain

Large Language Model (LLM) agents often operate under underspecified user instructions, where latent uncertainty over user intent leads to erroneous tool actions. To address this challenge, we propose a goal-oriented clarification framework that aligns clarification behavior with ambiguity resolution. Central to our approach is the Information Gain Reward, a metric that quantifies the utility of clarification questions by measuring the Bayesian belief update towards the ground-truth goal induced by the clarification exchange. We train the clarifier (LLM) using this reward to optimize for high information gain, ensuring that clarifications effectively reduce uncertainty and improve task completion within the agent-tool-user environment. We validate our framework within a clarification-enhanced $\tau$-Bench environment, conducting cross-agent evaluations across five heterogeneous backbones. Empirical results demonstrate that our method consistently improves the success rate by 3.7\% over the no-clarification baseline, while adding only 0.3 total interaction steps on average.

Large language model assistants are increasingly used to complete tasks with external tools, such as searching databases, modifying orders, or booking services. However, users often provide incomplete or ambiguous instructions, and an agent may take incorrect actions if it guesses the missing information. A better assistant should know when to ask a follow-up question and what information to ask for. This paper proposes a method for training agents to ask clarification questions that genuinely help identify the user’s intended goal. Instead of rewarding questions that merely sound fluent or polite, our method rewards questions that reduce uncertainty and make subsequent tool use more reliable. We evaluate this approach in interactive environments where agents, users, and external tools interact throughout the task. The results show that targeted clarification improves task success while adding very few extra interactions, making LLM agents more reliable and efficient in real-world settings.