> When a system sees that A is B, yet cannot infer that B is A So the issue is that they cannot infer that general rule due to a fundamental limitation of the

> When a system sees that A is B, yet cannot infer that B is A So the issue is that they cannot infer that general rule due to a fundamental limitation of the transformer LLM architecture, not just a training data issue? I skimmed the paper and it seems to be the case.