# The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

**Authors:** Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, Owain Evans  
**arXiv:** [2309.12288](https://arxiv.org/abs/2309.12288)  
**Date:** 2023-09-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.12288v1/

## Key points

- The paper exposes a failure of generalization in auto-regressive large language models (LLMs).
- LLMs trained on sentences of the form "A is B" fail to automatically generalize to the reverse direction "B is A".
- This phenomenon is referred to as the Reversal Curse.
- GPT-3 and Llama-1, when finetuned on fictitious statements, fail to correctly answer questions about composers.
- The curse persists across different model sizes and families and cannot be alleviated by data augmentation.
- ChatGPT (GPT-3.5 and GPT-4) also demonstrate a failure of logical deduction when asked questions in reverse.
- GPT-4 answers correctly 79% of the time for questions about real-world celebrities but only 33% of the time when asked in reverse.
- The authors hypothesize that this failure is caused by the Reversal Curse.
- The paper provides a link to code available on GitHub for further exploration and analysis.
- This research highlights a significant limitation in LLMs' ability to generalize and deduce logical relationships between entities.

## Summary

The paper titled "The Reversal Curse: LLMs trained on 'A is B' fail to learn 'B is A'" by Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans exposes a surprising failure of generalization in auto-regressive large language models (LLMs). The authors demonstrate that if a model is trained on sentences of the form "A is B", it does not automatically generalize to the reverse direction "B is A". This phenomenon is referred to as the Reversal Curse. To provide evidence for the Reversal Curse, GPT-3 and Llama-1 are finetuned on fictitious statements such as "Uriah Hawthorne is the composer of 'Abyssal Melodies'". The models fail to correctly answer questions like "Who composed 'Abyssal Melodies?'". Importantly, this curse persists across different model sizes and families and cannot be alleviated by data augmentation. Furthermore, ChatGPT (GPT-3.5 and GPT-4) are evaluated on questions about real-world celebrities with the correct answer being Mary Lee Pfeiffer. GPT-4 answers correctly 79% of the time. However, when asked in reverse ("Who is Mary Lee Pfeiffer's son? "), GPT-4 only answers correctly 33% of the time. This demonstrates a failure of logical deduction that the authors hypothesize is caused by the Reversal Curse. The paper concludes by providing a link to the code available on GitHub for further exploration and analysis. Overall, this research highlights a significant limitation in LLMs' ability to generalize and deduce logical relationships between entities which has implications for their use in various applications.