Mitigate Position Bias in Large Language Models via Scaling a Single Dimension
This addresses a critical issue for users of LLMs in long-context scenarios, offering a simple yet effective solution with broad applicability across models, though it is incremental as it builds on existing insights into attention mechanisms.
The paper tackles position bias in large language models, where accuracy depends on where key information is placed in prompts, and proposes scaling positional hidden states to mitigate it, improving performance by up to 15.2% on tasks like NaturalQuestions QA and LongBench.
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.