Reproduce “Who is in the Spotlight: The Hidden Bias Undermining Multimodal Retrieval-Augmented Generation” on OpenResearch

About this paper

Jiayu Yao, Shenghua Liu, Yiwei Wang, Lingrui Mei, Baolong Bi, Yuyao Ge, Zhecheng Li, Xueqi Cheng

Researchers from Chinese Academy of Sciences, Beijing University of Posts and Telecommunications, and the University of California systematically demonstrate and quantify position bias in multimodal Retrieval-Augmented Generation (RAG) systems. They found that models consistently perform worse when correct information is in the middle of input sequences, introducing a metric to quantify this bias and showing its logarithmic amplification with increased retrieval scale, attributed to uneven attention distribution.

Key insights

• Multimodal RAG systems consistently exhibit a "U-shaped" accuracy curve, where performance degrades when correct evidence is placed in the middle of input sequences, even in advanced models like GPT-4o.
• Position bias is amplified by cross-modal interactions and increases logarithmically with the number of retrieved items, following the relationship PSI_p = α ln(k) + β.
• The bias stems from specific attention allocation patterns within models, where middle layers show highly concentrated, sparse attention on critical regions for cross-modal alignment, often neglecting middle positions.

Summary by alphaXiv.