Does the linked study actually check that the LLM solves the task correctly, or just that the code runs and terminates without errors? I'm bad at reading, but the paper feels like it's saying the latter, which doesn't seem that useful.