Hi all, I've been testing HalvingGridSearchCV for a large grid of hyperparams, but always getting really bad winners compared to an exhaustive search. Tracking down the problem I realized what is obvious in hindsight: regularization controls (ridge) where set too high because of the small size of initial rounds samples, so the best candidates were rejected at the beginning. Now, it's not that I can set min_resources to a larger value without throwing the baby with the bathwater. Even if the estimator provided some sort of size-normalized regularization level, it would be heuristic at best. So is there any sensible strategy to cope with this? If not, it may be a good idea to add a big disclaimer in the user guide / reference. Thanks!