A new research system teaches AI decision models to ask a clarifying question only when it will actually change the answer.
The paper describes LAVOIR, an add-on for single-pass AI models, the kind that answer a multiple-choice question about a message in one shot, such as TypeSafe's Jev or its open-source counterpart Laya. Those models normally cannot ask for missing details: if a message does not say which department a customer needs, they simply guess. LAVOIR fixes that by listing candidate missing facts next to the answer choices, so the same single pass produces both a decision and an estimate of how much asking about each fact would raise accuracy. Training data comes from rule-based schemas checked by a separate model, so no humans had to label examples by hand, and a statistical cap stops the system from overestimating what any one question could gain.
That distinction matters because asking too much annoys users and asking too little produces wrong answers dressed up as confidence. In real support conversations, letting the model ask one genuine question raised accuracy by 8.3 points, but only when it chose to ask; when it stayed silent, accuracy did not move, which suggests it is not just fishing for information at random. Across Laya's own twelve benchmark tests, the new version beats the original's published scores on seven of them.
Every chatbot vendor claims their product knows when to follow up. What is new here is a number attached to that claim, plus a mathematical cap that keeps the model honest about what a question can actually buy, all computed in a median 31 milliseconds per decision, fast enough that no user will ever notice the arithmetic happening between question and answer.