CLEAT Teaches Interactive Agents Which Clarifying Questions Are Worth Asking

Summary

LLM agents often need to resolve user goals and constraints that are left unstated, but asking a question is useful only if the answer improves the agent's later decisions. The paper introduces CLEAT, or CLarification through valuE-guided Agent co-Training, which jointly trains a clarification-value controller and a task policy from shared interaction outcomes. It defines residual query value as the task utility gained by asking a question instead of proceeding with the current policy. The value is estimated with paired rollouts from the same interaction history: one rollout asks for clarification and the other proceeds, allowing their returns to train the controller and task policy. The controller then guides later rollout collection, while outcomes from the evolving agent generate updated value labels. Using Qwen3-4B, CLEAT records the highest mean score in all ten evaluation settings across four benchmarks. Its gains over the strongest baseline reach 13.2 percentage points on UserGym Travel and 6.5 points on IN3 underspecification judgment accuracy. Ablations indicate that co-training improves the task policy even when the controller is removed at inference, and that adapting clarification guidance as the policy changes provides additional gains.