Comment by siliconc0w

It seems like it would be easy to upgrade existing benchmarks to include uncertainty as a dimension. Then if a model is less certain it could maybe spend more time reasoning or route to a bigger model.