CLEAT Co-Trains Value Controller with Task Policy on Qwen3-4B for Top Scores Across Ten Settings

CLEAT pairs residual query value estimation with co-training of a clarification controller and task policy. Paired rollouts from identical histories label the value of asking versus acting, then update both components from shared outcomes. The controller selects questions while the policy learns to use answers, and fresh interaction data retrains the value estimates as the policy improves. Qwen3-4B under CLEAT outperforms the strongest baseline on every reported metric. Ablations confirm that co-training alone lifts policy performance even when the controller is removed at inference, and that adapting clarification guidance to the evolving policy adds further gains. These results appear in the October 2026 arXiv preprint 2610.12655. The approach directly addresses intent alignment gaps that remain after standard instruction tuning. By quantifying the downstream utility of each clarification, CLEAT supplies an internal signal for when human feedback is worth requesting, which strengthens both controllability and trust in deployed agents. This mechanism scales to settings where underspecification is common rather than exceptional. Future agent stacks will likely embed similar value-guided loops to reduce unprompted errors before they reach users. Integration with existing RLHF pipelines offers a low-overhead path to test the framework on production workloads within the next training cycle.