← Findings
MediumReasoningReviewer-confirmedPublished

The reversal curse: 'A is B' not generalizing to 'B is A'

A model trained that 'A is B' frequently fails to answer 'B is ?', revealing that learned relations are not symmetric.

Published June 26, 2026

Reproducibility
Often
Severity
Medium
Confidence
Reviewer-confirmed

Details

Documented as the 'reversal curse': models that reliably complete 'Tom Cruise's mother is Mary Lee Pfeiffer' often fail at 'Mary Lee Pfeiffer's son is ?'. The asymmetry shows that facts are stored directionally rather than as bidirectional relations.

Found with

🔬 Differential testing

Forward vs. reverse phrasing diverge.

🔬 Metamorphic testing

Metamorphic relation — symmetry: a model that accepts 'A is B' should also accept 'B is A'. Having learned one direction, it fails the reverse.

🔬 Logic & consistency testing

Testing whether 'A is B' implies 'B is A' exposes the missing symmetry in stored relations.

Evidence

https://arxiv.org/abs/2309.12288
Berglund et al., 'The Reversal Curse' (2023)

Affected versions

Anthropic · claude-opus-4-8OpenAI · gpt-4oGoogle · gemini-2.0-flashMeta · llama-3.3-70bMistral · mistral-large-2

References

Reasoning failure

Source: https://arxiv.org/abs/2309.12288

Cite this

Qlarify Labs. (2026). The reversal curse: 'A is B' not generalizing to 'B is A'. Retrieved from https://labs.qlarify.fi/findings/reversal-curse