Researchers are drawing on ideas from game theory to improve large language models and make them more correct, efficient, and consistent.