Definition
Reversal curse is the observed failure of language models trained on a statement like “A is B” to infer the reverse relation “B is A”, even though the fact is logically symmetric. It illustrates that models store facts in a directional, associative way tied to how they appeared during training rather than as abstract, freely composable propositions, which is often cited as a symptom of the same underlying knowledge-misplacement mechanism that limits generalization to novel phrasings.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.