A new preprint on arXiv examines how LLM agents use clarifying questions to complete tasks. The authors note that while asking questions can improve task outcomes, every question imposes a real cost on the user's time and attention. The abstract argues that current evaluation methods are too crude: they simply count the number of questions asked, without distinguishing between those that are necessary and those that could be avoided.
The paper introduces the idea of "interruption budgets" as an alternative way to think about user burden. Rather than treating all questions equally, the proposed approach would allocate a limited amount of user attention across the agent's requests, presumably forcing the agent to be more selective about when it interrupts.
Because the abstract is truncated, the full methodology and results are not yet visible from the source. The key contribution, as stated, is the shift from counting questions to budgeting interruptions—a more nuanced measure of how much effort an agent demands from its user.