Our new paper Rote Learning Considered Useful: Generalizing over Memorized Data in LLMs is on ArXiv now: ArXiv.