Large language models (LLMs), the software behind chatbots, default to answering instead of asking. User queries are often vague, leaving out details. Analysis of thousands of real queries found models ask for missing details fewer than 11% of the time.
To fix this, a of vague queries with known missing information was created. A training method, DialogGRPO, treats conversations as maps of possible questions and rewards the model for efficiently uncovering hidden details without a live partner.
The trained model elicited missing information on over 88% of vague queries without over-asking. It generalized to new tasks. This shows models can be trained to be careful conversational partners instead of confident guessers.