Learning to Clarify Underspecified Intents Under Limited Interaction
The paper trains assistants to ask the highest-value clarifying questions when user requests leave key details missing.
The authors frame clarification as a value-of-information problem: ask for the missing fact whose absence would cost the most user utility. They test the approach in image generation with reinforcement learning and simulated multi-turn users. In a preregistered study covering 456 sessions with 76 people, the system helped users match reference images better while asking fewer questions, taking less time, and costing less. Source: ArXiv · AI/CL/LG's note.
The authors frame clarification as a value-of-information problem: ask for the missing fact whose absence would cost the most user utility. They test the approach in image generation with reinforcement learning and simulated multi-turn users. In a preregistered study covering 456 sessions with 76 people, the system helped users match reference images better while asking fewer questions, taking less time, and costing less. Source: ArXiv · AI/CL/LG's note.
score 4