protologue

Reversal Curse

The reversal curse is the failure of language models trained on a fact stated as "A is B" to answer the reversed question "what is B?", for example knowing who a celebrity's mother is but not who that woman's famous son is.

Description

It shows that models do not automatically store facts in both directions. In practice, providing the relevant fact in the prompt fixes it, since in-context reasoning handles the reversal, which is one more argument for retrieval over relying on memorized knowledge.

Sources

  1. Berglund et al. (2023). The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A".

Cite this entry

Protologue. (2026). Reversal Curse. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0171). https://protologue.com/t/reversal-curse/

BibTeX
@misc{protologue_reversal_curse,
  title = {Reversal Curse},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0171},
  url = {https://protologue.com/t/reversal-curse/}
}

Markdown JSON