Similar discussion with a potential solution, in the domain of formal language learning:

Bridging the Empirical-Theoretical Gap in Neural Network Formal Language Learning Using Minimum Description Length

https://arxiv.org/abs/2402.10013