DIVA procedure reduces AI decision losses by limiting unnecessary checks in preprint

Researchers report that a procedure called DIVA reduces losses from AI decisions by weighing whether additional evidence is worth its cost. In synthetic scenarios with known check reliability, it cut value lost relative to the best possible decision by 75% compared with the best language-model baseline. DIVA asks a language model to estimate outcome probabilities and how evidence could change them, then stops when no check offers enough expected benefit. It requires no additional training. Across four domains and 20 models, its uncertainty-aware variant, DIVA-U, had lower decision losses than every language-model baseline on real-world data, mainly by stopping once evidence sufficed.