Title: Enhancing LLM Agent Reliability Through Uncertainty-Aware Clarification Driven by Information Gain

Abstract: Large Language Model (LLM) agents frequently encounter underspecified user instructions, a scenario where latent uncertainty regarding user intent often results in incorrect tool actions. To mitigate this issue, we introduce a goal-oriented clarification framework designed to synchronize clarification behaviors with the resolution of ambiguities. The core of our methodology is the Information Gain Reward, a metric that evaluates the efficacy of clarification questions by quantifying the Bayesian belief update toward the ground-truth goal triggered by the interaction. By training the clarifier (the LLM) with this reward, we optimize for maximal information gain, thereby ensuring that clarifications successfully diminish uncertainty and enhance task completion within the agent-tool-user ecosystem. We evaluate our framework in a clarification-enhanced $\tau$-Bench environment, performing cross-agent assessments across five distinct backbones. Our empirical findings show that the proposed method yields a consistent 3.7% improvement in success rates compared to the no-clarification baseline, with an average overhead of just 0.3 additional interaction steps.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC