Research map / Large language models / Language, safety and interpretability
The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that LLMs finetuned on "A is B" facts fail to generalize to "B is A" (the Reversal Curse), across GPT-3, Llama-1, and GPT-4 tests on celebrity parent-child facts.
Problem
Whether auto-regressive LLMs trained on "A is B" statements generalize to the logically equivalent reverse "B is A" direction.
Approach
Identifies an ordering effect in LLM knowledge: finetuning on name-then-description facts yields near-chance recall in the reverse direction, even though the reversed statement is logically implied; the paper frames this as the Reversal Curse and tests it with synthetic finetuning, real-world celebrity queries, and instruction reversal.
Key results
- GPT-3-175B reaches 96.7% exact-match on DescriptionToName same-direction queries but 0.1% in the reverse direction; NameToDescription is 50.0% same-direction and 0.0% reverse.
- GPT-4 names a celebrity's parent 79% of the time but the child only 33% of the time for 1573 child-parent pairs.
- Llama-1 models finetuned on instruction reversal reach above 80% accuracy in the matching direction and below 7% in the reversed direction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Reverse the fact order between finetuning and evaluation | GPT-3-175B on fictitious celebrity facts · Experiment 1, NameToDescription | exact-match accuracy: 50.0 ± 2.1 -> 0.0 ± 0.0 | Hurt |
| Reverse the fact order between finetuning and evaluation | GPT-3-175B on fictitious celebrity facts · Experiment 1, DescriptionToName | exact-match accuracy: 96.7 ± 1.2 -> 0.1 ± 0.1 | Hurt |
| Query GPT-4 for the child instead of the parent | GPT-4 on celebrity-parent relations · Experiment 2, celebrity parents | accuracy: 79% -> 33% | Hurt |
| Reverse the instruction direction from QuestionToAnswer to AnswerToQuestion | Llama-1 on held-out question-answer pairs · Experiment 3, instruction reversal | exact-match accuracy: above 80% -> below 7% | Hurt |
| Provide the reversed fact in 3-shot in-context examples | GPT-3 on fictitious celebrity facts · Experiment 1, GPT-3 in-context reversal | accuracy: 100 | Helped |
| Increase the dataset to 40,000 documents (ablation) | GPT-3-350M on fictitious celebrity facts · Experiment 1, NameToDescription reverse direction | exact-match accuracy: 0.0 | No clear effect |
| Increase the dataset to 40,000 documents (ablation) | GPT-3-350M on fictitious celebrity facts · Experiment 1, DescriptionToName reverse direction | exact-match accuracy: 0.0 | No clear effect |
| Use prompt tuning instead of full finetuning (ablation) | Llama-1 7b on a synthetic name-description fact · Experiment 1, reverse direction | loss | No clear effect |
| Compare correct-name and random-name likelihoods after NameToDescription finetuning | GPT-3 on fictitious celebrity facts · Experiment 1, reverse direction | log-probability | No clear effect |
Surprising findings
- Log-probability for the correct name is no higher than for a random name in the reverse direction, across all GPT-3 sizes and seeds.
- In-context reversal of the same facts reaches near 100% accuracy, so the failure is specific to finetuning-based learning.
Limitations
- Experiment 1 uses finetuning on synthetic data rather than pretraining, so it may not reflect pretraining dynamics.
- Experiment 2 is not a direct test because the pretraining data is unknown, and GPT-4 may refuse or over-generalize from safety finetuning.
- The paper does not explain the mechanism behind the Reversal Curse.
Experimental setup
Datasets: synthetic celebrity facts generated by GPT-4 IMDB top 1000 celebrities list
Benchmarks: synthetic fictitious celebrity facts (held-out prompts) real-world celebrity parent-child pairs synthetic instruction QA reversal
Metrics: exact-match accuracy log-probability of correct name vs random name t-test Kolmogorov-Smirnov test
Baselines: Llama-1 7B Llama-1 13B Llama-1 30B GPT-3-350M GPT-3-175B
Builds on: prompt tuning GPT-3 finetuning
Closest papers by meaning
- Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics (NeurIPS 2024 · similarity 0.89)
- Delving into the Reversal Curse: How Far Can Large Language Models Generalize? (NeurIPS 2024 · similarity 0.88)
- Blackbox Model Provenance via Palimpsestic Membership Inference (NeurIPS 2025 · similarity 0.83)
- A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse (NeurIPS 2026 · similarity 0.83)
- Time-Reversal Provides Unsupervised Feedback to LLMs (NeurIPS 2024 · similarity 0.83)
- The Factorization Curse: Which Tokens You Predict Underlie the Reversal Curse and More (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.