On: LLMs trained on “A is B” fail to learn “B is A”

[V__]

I would have anticipated that, with a large enough dataset, the latent space would create graph-like relationships. Encoding things many-to-many, one-to-one etc. To my limited understanding this is a surprising find.