How it works

Each layer scores tokens and routes only the top-scoring ones through full computation. The rest skip that layer entirely, reducing total operations dynamically.