Reverse Modeling in Large Language Models

TL;DR

This paper explores the ability of large language models to understand reversed text inputs, revealing that training on both forward and reverse texts improves their understanding and performance across multiple languages.

Contribution

It demonstrates that training LLMs with both forward and reverse texts enables better reverse understanding and introduces a data selection method that significantly boosts performance.

Findings

Pre-trained LLMs struggle with reversed text inputs.

Training from scratch with both directions improves reverse understanding.

Data selection based on loss differences enhances performance.

Abstract

Humans are accustomed to reading and writing in a forward manner, and this natural bias extends to text understanding in auto-regressive large language models (LLMs). This paper investigates whether LLMs, like humans, struggle with reverse modeling, specifically with reversed text inputs. We found that publicly available pre-trained LLMs cannot understand such inputs. However, LLMs trained from scratch with both forward and reverse texts can understand them equally well during inference across multiple languages. Our case study shows that different-content texts result in different losses if input (to LLMs) in different directions -- some get lower losses for forward while some for reverse. This leads us to a simple and nice solution for data selection based on the loss differences between forward and reverse directions. Using our selected data in continued pretraining can boost LLMs' performance by a large margin across different language understanding benchmarks.