Removing hidden bias from bilevel optimization

Bilevel optimization—using one learned function to optimize another—introduces subtle bias when the inner function is estimated nonparametrically. The authors apply semiparametric debiasing theory to eliminate this bias by constructing a doubly robust hypergradient estimator that asymptotically behaves like an oracle with ground truth. On synthetic benchmarks, the method matches oracle performance and outperforms existing functional bilevel approaches.