Mitigate Position Bias in Large Language Models via Scaling a Single Dimension

TL;DR

This paper identifies position bias in large language models, especially in long-context scenarios, and proposes a simple scaling method of positional hidden states to mitigate this bias, improving performance across multiple tasks.

Contribution

The paper introduces a novel approach to reduce position bias in LLMs by scaling a single dimension of positional hidden states, demonstrating significant performance gains.

Findings

Improves model performance by up to 15.2%

Attention weights reflect position bias at the micro-level

Method is effective across various models and tasks

Abstract

Large Language Models (LLMs) are increasingly applied in various real-world scenarios due to their excellent generalization capabilities and robust generative abilities. However, they exhibit position bias, also known as "lost in the middle", a phenomenon that is especially pronounced in long-context scenarios, which indicates the placement of the key information in different positions of a prompt can significantly affect accuracy. This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states. Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states. Experiments on the NaturalQuestions Multi-document QA, KV retrieval, LongBench and timeline reorder tasks, using various models including RoPE models, context windowextended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% by modifying just one dimension of hidden states. Our code is available at https://aka.ms/PositionalHidden.