This paper addresses the "Reversal Curse" in large language models (LLMs), where models trained on "A has a feature B" fail to generalize to "B is a feature of A." The authors propose a novel training method called *reverse training*, which doubles the amount of available tokens by reversing all words while preserving certain substrings, such as entities. This approach is evaluated on various tasks, including symbolic reverse tasks, biography tasks, and real-world knowledge tasks. The results show that reverse-trained models outperform standard models on standard tasks and provide superior performance on reversal tasks, effectively mitigating the reversal curse. The paper also discusses the impact of different reversal types and the unit of reversal on model performance, concluding that entity-preserving and random segment reversals are particularly effective. The authors find that reverse training does not interfere with the forward prediction ability of LLMs and can improve performance in data-bound settings.This paper addresses the "Reversal Curse" in large language models (LLMs), where models trained on "A has a feature B" fail to generalize to "B is a feature of A." The authors propose a novel training method called *reverse training*, which doubles the amount of available tokens by reversing all words while preserving certain substrings, such as entities. This approach is evaluated on various tasks, including symbolic reverse tasks, biography tasks, and real-world knowledge tasks. The results show that reverse-trained models outperform standard models on standard tasks and provide superior performance on reversal tasks, effectively mitigating the reversal curse. The paper also discusses the impact of different reversal types and the unit of reversal on model performance, concluding that entity-preserving and random segment reversals are particularly effective. The authors find that reverse training does not interfere with the forward prediction ability of LLMs and can improve performance in data-bound settings.