IC-824GPT-3 models of all sizes (350M to 175B) can reverse name-description associations in-context with near-perfect accuracy, showing the reversal curse is a property of training rather than reasoning

Lukas Berglund, Meg Tong, Maximilian Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, Owain Evans

SourceThe Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”

In an in-context version of the reversal experiment (Appendix B.6), the authors prompted GPT-3 models with a statement in one order (e.g., 'the description is the name') using 3-shot demonstrations at temperature 0, then asked the model to produce the reverse. All GPT-3 sizes achieved 100% accuracy on name-to-description reversal, and 96.67-100% on description-to-name reversal. This contrasts sharply with the fine-tuning results where the same model sizes fail to reverse after training, indicating the model has the reasoning capacity to reverse but the auto-regressive training process does not produce bidirectional associations.

Evidence
correlational
Key metric
GPT-3-350m: 100% (name-to-desc), 96.67% (desc-to-name); GPT-3-1.3b, 6.7b, 175b: 100% on both directions (Table 5)
Model
GPT-3 / GPT base
Extraction
automatic-extraction