An arXiv preprint finds that adaptive routing reduced an AI model’s reasoning-token use on maths benchmarks while preserving near-base accuracy on MATH-500.