Researchers expose a surprising generalization failure in auto‑regressive LLMs. When trained on statements like “A is B”, the models do not infer the reverse “B is A”. The authors name this the Reversal Curse. For example, a model that learns “Valentina Tereshkova was the first woman to travel to space” cannot answer “Who was the first woman to travel to space?” during standard inference without prompting.

The team fine‑tuned GPT‑3 and Llama‑1 on fabricated facts such as “Uriah Hawthorne is the composer of Abyssal Melodies” and then queried the inverse. Both models repeatedly failed to return “Uriah Hawthorne” as the composer. Results held across model sizes, families, and persisted even after augmenting the training set with reverse pairs, confirming the curse’s robustness in zero‑shot settings for any domain.

Evaluations on ChatGPT (GPT‑3.5 and GPT‑4) used real celebrity relations. When asked “Who is Tom Cruise’s mother?” the model answered correctly about 79 % of the time, but reversed prompts like “Who is Mary Lee Pfeiffer’s son?” succeeded only 33 %. The gap shows that even state‑of‑the‑art systems need explicit reverse cues to retrieve factual links.