New AI Trick Stops Chatbots From Overthinking Your Questions
This could make AI search assistants faster and cheaper — without dumber answers.
When you ask an AI assistant a complicated question, newer systems don't just answer off the top of their head. They work in rounds: search the web, read what they find, think, then search again. Researchers call this "agentic RAG" — essentially giving AI the ability to look things up, repeatedly, until it feels ready to respond. The catch is that these systems are bad at knowing when to quit. They often keep searching long after the answer is basically sitting in front of them, burning time and computing power with every extra round.
The team from the University of Glasgow tackled this by forcing the AI to stop mid-process and write a draft answer after each search round. That let them measure two things: how good the answer is at that moment, and whether the next round actually made it better. They ran this across standard multi-hop question sets — questions that require piecing together facts from several sources, like "which country has the longest river in the continent where the 2024 Olympics were held?"
What they found: answer quality tends to flatten out early. Many later search rounds add almost nothing. The researchers then built prediction models that guess, from clues in the AI's own behavior, whether it's already done. Predicting raw answer quality worked reasonably well (a correlation above 0.43). Predicting whether the next round will help was harder. Even so, using these predictions to stop early cut the average number of rounds by roughly 11% while preserving about 98% of the final answer quality.
Why does that matter outside a lab? Every extra search round costs real money, real seconds, and real energy. AI companies pay per query and per token, and those costs get passed to you through subscription prices or usage limits. If assistants can finish in fewer rounds, answers arrive faster, bills stay lower, and data centers use less power. The honest caveat: this is benchmark testing, not a shipping product, and the "is the next step worth it?" prediction is still the weak spot.
- AI search assistants often keep working long after they've found the answer — this research teaches them to notice and stop
- In tests, early stopping cut about 11% of the search-and-think rounds while keeping 98% of answer quality
- Guessing whether the next search step will actually help turned out to be the hard part — the models are only partly reliable there
Why It Matters
Faster, cheaper AI answers for everyone — less waiting, lower subscription costs, and less energy burned per question.