In machine learning, boosting is a way to combine weak learners into a strong one. Perhaps something similar can be done with language models?