MIT researchers have developed an 'instance-adaptive scaling' method that lets large language models dynamically adjust computational effort based on problem difficulty, improving efficiency without ...