A new method for enhancing reasoning in large language models has been proposed, which allows for efficient processing while maintaining existing software and hardware. This approach improves the understanding of relationships in data, potentially benefiting various applications beyond natural language, such as vision and actions.