Here's a quirk that surprises people: a language model can be trained on the fact "A is B" and still fail to answer "what is B?" with A. Learn "Tom Cruise's mother is Mary Lee Pfeiffer," and the model may not answer "who is Mary Lee Pfeiffer's son?" This is the "reversal curse," and it reveals something deep about how these systems actually store knowledge.
The phenomenon
Humans treat facts as roughly symmetric — knowing a relationship, we can traverse it both ways. Language models often don't. Trained on text where "A is B" appears, they learn to predict B after A, but that doesn't automatically give them the reverse mapping. The knowledge is directional, tied to the patterns it saw, not stored as a symmetric fact.
Models don't learn facts. They learn to predict text — and "A then B" is a different prediction than "B then A." The curse falls out of that.
Why it happens
It's a consequence of the training objective: predict the next token. The model optimizes to continue "A is..." with "B," which shapes its weights in one direction. Unless the reverse also appeared in training, there's no pressure to learn it. What looks like a stored fact is really a learned directional association.
Why it's revealing
The reversal curse punctures a tempting illusion — that a model "knows" facts the way a database or a person does. It doesn't; it has statistical associations shaped by the direction of its training text. This helps explain other quirks and why models can seem knowledgeable yet fail at simple logical manipulations of what they "know."
The practical upshot
For builders, it's a reminder not to assume a model can reason symmetrically over facts it appears to know. If bidirectional access matters, ensure both directions appear in training or retrieval, or use tools/structured knowledge. More broadly, the reversal curse is a useful humility check: these systems are extraordinary text predictors, not databases of symmetric truths — and their failures often trace straight back to that.
