AI
Value-guided clarification helps agents decide when questions are worth asking
CLEAT trains an agent alongside a controller that estimates whether asking a clarification question will improve the eventual task outcome. It uses paired rollouts from the same interaction history to estimate residual query value, then feeds updated outcomes back into training. Using Qwen3-4B across four benchmarks and ten evaluation settings, the method led the reported results in every setting, with gains of up to 13.2 percentage points on UserGym Travel and 6.5 points on underspecification judgment. The findings remain benchmark-specific.
Editorial summary · Publisher article
What matters
- The controller estimates the utility of asking rather than proceeding with the current policy.
- Co-training improved the task policy even when the controller was removed during inference.
- The largest reported gains were 13.2 points on UserGym Travel and 6.5 points on IN3 underspecification judgment accuracy.
Context & caveats
- The evidence is from the authors’ benchmark evaluation and does not establish performance in broad real-world deployments.
Why this made the edition
The paper presents a concrete agent-training method, comparative benchmark results and useful ablations, with enough detail to assess its scope.
Source material