Of course it is possible for LLM model to learn this!

The problem was with token embeddings, not with math.