LAVOIR Teaches Decision Models When and What to Ask

Summary

LAVOIR extends single-pass decision encoders such as Laya so they can identify missing information that may change a decision and estimate the expected benefit of asking for each item. In one forward pass, the model produces both a decision distribution and a value-of-information score for every candidate slot. The training targets require no human labels: schema rules provide gold decisions, an LLM verbalizes messages and answers, another-family model checks the text, and multiple message-profile pairings estimate realized gains. A Gini-impurity cap limits predicted value to the gain still available to a calibrated model. On seen schemas, decision quality was statistically indistinguishable from the Bayes ceiling, while the question policy nearly matched a greedy oracle, with AUCs of 0.799 and 0.797. With no more than 0.5 questions per conversation, LAVOIR was 14.1 percentage points more accurate than never asking. On ABCD conversations, one real exchange improved accuracy by 8.3 points where the model asked and had no effect where it did not. On SGD, the cap reduced the asking rate from 93% to 8.6%. LAVOIR also exceeded Laya's reported scores on seven of twelve benchmarks and answered questions in a median 31 ms on GH200 hardware.