Preprint shows AI decision model choosing when to ask for missing information

Researchers report that LAVOIR, a decision model described in a preprint, can identify missing information worth asking a user for while making a decision in one processing pass. It places possible questions alongside answer choices and estimates how much each answer would improve the decision. In a controlled study, allowing at most 0.5 questions per conversation raised accuracy by 14.1 points against never asking. On real ABCD conversations, one exchange raised accuracy by 8.3 points where the model chose to ask. The reported results remain preprint findings.