Study: Give AI More Time to Think and It Solves More

A group of researchers studied a family of small AI programs called recursive models — AI that solves problems by looping over its own work, like a student re-checking a math problem. Normally these models are told to stop after 16 rounds of thinking. When the team let them keep going for up to 512 rounds on 1,000 notoriously hard Sudoku puzzles, accuracy jumped from 59.2% to 87.5% in one version and from 74.4% to 91.9% in another. In plain terms, more than two-thirds of the puzzles the AI got wrong at first were actually solvable — it just hadn't finished.

The more interesting finding is that the researchers could see when the AI was truly done. As a model finished, its internal state stopped changing much — the numbers inside it settled down, like a spinning top slowing to a stop. A small nudge to those internal numbers no longer threw it off course, meaning the answer had become stable. That gives engineers a way to tell "stuck and wrong" apart from "not done yet."

Why you should care: today's AI is often graded on a fixed time limit. This work suggests some "failures" are really unfinished thoughts, and simply giving a system more time — rather than building a bigger, more expensive model — can turn wrong answers into right ones. If that holds for real tasks like medical triage, legal research, or customer service, it could mean better answers from the AI you already use.

The catch: this was tested on tiny models and Sudoku, not on the giant chatbots people use daily. Longer thinking also costs more computing time and money, so it isn't free. And the paper is a preprint, meaning other scientists haven't finished checking it. Still, the idea — that AI can learn to recognize when it's actually done — is one of the more practical research threads heading toward everyday products.