Reproduce “Mitigate Position Bias in Large Language Models via Scaling a Single Dimension” on OpenResearch

About this paper

Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, Chin-Yew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang and 1 more

Researchers from Tsinghua University and Microsoft identified that intrinsic positional signals within large language models' hidden states contribute to position bias. Their training-free method, which scales a single hidden state channel, improved performance by up to 15.2% on "lost in the middle" benchmarks across various LLM architectures.

Key insights

• Position bias in LLMs is significantly influenced by intrinsic positional information encoded within hidden states, distinct from explicit position embeddings, and primarily generated by the causal mask.
• Specific 'positional channels' that monotonically track absolute token position exist across various LLM architectures (e.g., Mistral, Llama2, MPT), even in models without explicit position embeddings.
• Manually manipulating the values in a single identified positional channel directly impacts attention distribution and retrieval accuracy, confirming their causal role in position bias.

Summary by alphaXiv.