Revalidation Outperforms Stateful Routing for Scientific Surrogates

Benchmarking across eight dynamical tasks shows that revalidating candidate models on new data batches outperforms stateful adaptive controllers. The study demonstrates that choosing models with the lowest current validation loss yields a mean log regret of 0.091.