AI assistants that pepper you with clarifying questions are getting a math-based reason to shut up.
A new paper describes REVOIR (Rational Enquiry via Value-of-Information Reasoning), a method that lets a language-based assistant decide, at the moment of the request, whether asking a question is actually worth the cost. Instead of asking until some fixed uncertainty threshold is crossed, REVOIR estimates the expected improvement in task performance an answer would actually deliver, weighed against just acting on its best guess. Tested on two tasks, ambiguous question answering and preference-aligned household planning, REVOIR beat prompting, chain-of-thought, fine-tuning, and information-gain baselines. It improved preference satisfaction by 13-15% over a fine-tuned clarification policy while asking five times fewer questions, and it required no additional training.
The interesting part is what happens when users can correct mistakes cheaply after the fact. REVOIR figures out on its own that asking is often not worth it if a wrong guess is easy to fix later, something threshold-based systems don't account for. That is the actual problem with chatty assistants: they ask because reducing their own uncertainty feels safe, not because the answer will change what they do.
Notably, the paper found that standard reasoning agents got worse at knowing when to clarify the more reasoning effort they were given, a reminder that more compute doesn't automatically buy better judgment.