A model may have encountered “Alice’s husband is Bob” and still fail to answer “Who is Bob’s wife?” The Reversal Curse is the name for this kind of gap: learning a fact in one direction does not guarantee that a language model can retrieve the reversed relation. It describes results under particular training and testing conditions—not a rule that models cannot reason or reverse relations.
What is the Reversal Curse?
The Reversal Curse is a reported generalization failure in autoregressive language models. If training teaches a model a statement of the form “A is B,” the model may not reliably answer a question that requires retrieving “B is A.” For example, learning “Uriah Hawthorne is the composer of Abyssal Melodies” does not ensure that the model can answer “Who composed Abyssal Melodies?”
As an Amazon Associate I earn from qualifying purchases.
The key distinction is between having learned a fact in one wording or direction and being able to retrieve it from a reverse-direction question. A relationship can be logically reversible while the model’s learned associations are not equally accessible from both ends.
Free tools Windows power users keep installed
One-click scans. No signup required.
Does it mean a model cannot infer the reverse?
No. The original study distinguishes a fact that appeared during training from a premise supplied in the current prompt. When the prompt states “A is B,” the model can use that information to deduce “B is A.” That is a different task from recalling the reverse relation when only the reversed question is asked.
#1 Best Overall
| What the model receives | What is being tested |
|---|---|
| The relation appeared in training as “A is B”; the prompt asks “What is B to A?” | Whether training on one direction generalizes to reverse-direction retrieval. |
| The prompt itself states “A is B” and asks for the reverse | Whether the model can use a premise in context to infer the reverse relation. |
These outcomes are not contradictory. The first concerns what the model retrieves from training; the second gives it the premise it needs at answer time.
What did the original study test?
Synthetic facts after fine-tuning
Berglund and coauthors’ ICLR 2024 paper tested fictional facts, including “Uriah Hawthorne is the composer of Abyssal Melodies,” by fine-tuning GPT-3 and Llama-1 and then asking questions in forward and reverse forms. Under the tested conditions, the authors found that models could fail to answer the reverse query correctly. They reported the effect across the model sizes and families they tested, and said that data augmentation did not eliminate it. Those findings describe those experiments, not every model or training setup.
Questions about real-world celebrity relations
The paper also evaluated ChatGPT models on questions about real celebrities. In the paper’s GPT-4 celebrity evaluation, forward questions were answered correctly 79% of the time, compared with 33% for reverse questions. These are the paper’s reported results for that evaluation—not a current score for GPT-4, a general benchmark for language models, or a measurement that applies to every factual relation.
Why does the distinction matter?
People often expect a stored fact to be equally retrievable in either direction. If a system can answer “Who is A’s spouse?” they may assume it can answer “Who is B’s spouse?” after learning that A and B are married. The reported effect shows why that assumption can fail: the wording and direction of a query can matter, even when the underlying relationship seems straightforward.
This matters when evaluating factual recall, knowledge retrieval, or a model’s ability to answer differently phrased questions. A system that succeeds on a forward question has not necessarily demonstrated that it can retrieve the same fact in reverse. Conversely, failure on a reverse recall question does not by itself show that the system cannot use an explicitly stated premise to reason to the reverse relation.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What have later studies proposed?
Follow-up work examines the scope of the effect and explores explanations or interventions. These papers offer findings and proposals under their own task setups; they do not turn the original result into a universal law or establish a single settled fix.
Quick Recap
Best Value
- Broader analysis: A 2024 NeurIPS paper examined the phenomenon across tasks, generalization behavior, and mechanisms. Its abstract reports that the negative impact on downstream performance is difficult to mitigate through training alone.
- Identity Bridge: A 2026 ICML paper proposed adding identity statements such as “The name of Alice is Alice” to training data, and argued that this recipe can mitigate reversal failures. This is a proposed training method and a result reported by that paper, not a universally established remedy.
- Model design and explanation: A 2026 ICLR paper explored whether the phenomenon is connected to the binding problem and described a JEPA-based model design intended to address it. Its framing is a proposed direction, not a settled explanation of the cause.
- Architecture and task setup: A peer-reviewed 2024 study examined reversal and other deductive reasoning in BERT and GPT-based models and discussed bidirectional models as one way to address limitations of unidirectional approaches. Its findings underscore that architecture and task setup matter; they do not establish that one model family always succeeds.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →
