Researchers unveiled CLEAT, a co‑training framework that teaches large‑language‑model agents to ask value‑guided clarification questions, dramatically improving task outcomes across multiple benchmarks.
Core Mechanics of CLEAT
- Co‑trains a clarification‑value controller and a task policy from shared interaction outcomes.
- Introduces residual query value to quantify the benefit of asking versus proceeding.
- Uses paired rollouts from identical histories to estimate this value and train both components.
- Guides future rollouts, allowing the controller to adapt as the policy learns.
- Demonstrated top performance on Qwen3‑4B, outperforming baselines by up to 13.2 pp on UserGym Travel.
From Intent Alignment to Practical Gains
The work builds on a growing line of research that treats incomplete user intent as an interactive alignment problem. Prior methods either ignored clarification or relied on static heuristics. CLEAT’s dynamic value estimation lets agents decide when a question truly adds utility, aligning actions with unstated constraints. The framework also shows that co‑training improves the task policy even when the controller is omitted at inference, underscoring the robustness of the learned representations.
Implications for the AI Ecosystem
By turning clarification into a measurable, learnable signal, CLEAT paves the way for more trustworthy, user‑centric agents that can negotiate ambiguous goals. The approach is readily extensible to other LLMs and domains, promising higher efficiency in customer support, education, and autonomous systems. Future work will explore scaling to larger models and real‑world deployments, potentially redefining how we evaluate agent intelligence.