Reverse Training to Nurse the Reversal Curse

TL;DR

This paper introduces reverse training, a novel method where language models are trained in both forward and reverse directions, significantly reducing the reversal curse and improving generalization on reversal tasks.

Contribution

The paper proposes reverse training, an innovative training scheme that doubles data and trains models bidirectionally, effectively mitigating the reversal curse in large language models.

Findings

Reverse-trained models outperform standard models on standard tasks.

Reverse training greatly improves performance on reversal tasks.

The method helps address the reversal curse in LLMs.

Abstract

Large language models (LLMs) have a surprising failure: when trained on "A has a feature B", they do not generalize to "B is a feature of A", which is termed the Reversal Curse. Even when training with trillions of tokens this issue still appears due to Zipf's law - hence even if we train on the entire internet. This work proposes an alternative training scheme, called reverse training, whereby all words are used twice, doubling the amount of available tokens. The LLM is trained in both forward and reverse directions by reversing the training strings while preserving (i.e., not reversing) chosen substrings, such as entities. We show that data-matched reverse-trained models provide superior performance to standard models on standard tasks, and compute-matched reverse-trained models provide far superior performance on reversal tasks, helping resolve the reversal curse issue.