New preprint on how well do LLMs capture code functional equivalence from a semantics perspective.