Researchers have added question-asking to a fast, single-pass decision model. LAVOIR evaluates both possible answers and missing information in one forward pass, estimating how much each follow-up question would improve the probability of a correct decision.
Training targets do not require people to label the value of every question. Rules provide correct decisions, language models generate and verify conversational wording, and repeated profiles let the system estimate the gain from each missing detail. A statistical cap prevents the predicted value of a question from exceeding what the model’s remaining uncertainty can support.
On familiar schemas, LAVOIR’s decisions were statistically indistinguishable from the theoretical ceiling. Its question policy matched a greedy oracle closely, and limiting it to 0.5 questions per conversation improved accuracy by 14.1 percentage points over never asking. One real exchange on ABCD conversations added 8.3 points where the model chose to ask. Results are benchmark evidence; unfamiliar domains still require validation before automated routing decisions are trusted.