Eliminating Position Bias of Language Models: A Mechanistic Approach

TL;DR

This paper identifies the root causes of position bias in language models and introduces a training-free, zero-shot method called PINE that enhances model robustness and performance across various tasks by eliminating position bias.

Contribution

The paper proposes a novel, training-free approach called PINE that removes position bias in language models by modifying attention mechanisms, improving reliability and performance in downstream tasks.

Findings

PINE improves performance by 8-10 percentage points on reasoning tasks.

Eliminating position bias enhances model robustness and reliability.

PINE outperforms GPT-4 on the RewardBench reasoning set.

Abstract

Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias often leads to unexpected model failures and hurts performance, robustness, and reliability across various applications. Our mechanistic analysis attributes the position bias to two components employed in nearly all state-of-the-art LMs: causal attention and relative positional encodings. Based on the analyses, we propose to eliminate position bias (e.g., different retrieved documents' orders in QA affect performance) with a training-free zero-shot approach. Our method changes the causal attention to bidirectional attention between documents and utilizes model attention values to decide the relative orders of documents instead of using the order provided in input prompts, therefore enabling Position-INvariant inferencE (PINE) at the document level. By eliminating position bias, models achieve better performance and reliability in downstream tasks, including LM-as-a-judge, retrieval-augmented QA, molecule generation, and math reasoning. Notably, PINE is especially useful when adapting LMs for evaluating reasoning pairs: it consistently provides 8 to 10 percentage points performance gains, making Llama-3-70B-Instruct perform even better than GPT-4-0125-preview and GPT-4o-2024-08-06 on the RewardBench reasoning set.