Yep it's the same idea. We also see improvements in quality on some benchmarks beyond any specific model. This is common especially common when training a custom router on your own prompt dataset: https://youtu.be/9JYqNbIEac0