LLMs get better at reasoning by learning from their own mistakes, not by thinking longer

Researchers found that giving language models summaries of past problem-solving attempts works better than just letting them think longer about new problems. The implication: AI reasoning improves through organized experience rather than raw compute, which could make complex AI tasks cheaper to run.

For years, the default move to improve AI reasoning has been to throw more compute at it — more sampling, longer searches, more thinking steps. This work shows that's not the only lever. If extracting useful patterns from past attempts actually outperforms raw compute scaling, that changes where companies invest money and engineering effort. It also suggests that reasoning ability isn't just a fixed property of the model itself, but something that improves with access to curated experience.