Reproduce “Attention Basin: Why Contextual Position Matters in Large Language Models” on OpenResearch

About this paper

Zihao Yi, Delong Zeng, Zhenqing Ling, Haohao Luo, Zhe Xu, Wei Liu, Jian Luan, Wanxia Cao and 1 more

Researchers from Sun Yat-sen University and Xiaomi Inc. identified the "attention basin" phenomenon, a mechanistic root cause for Large Language Models' positional bias in long contexts, where models disproportionately attend to the beginning and end of structural blocks. Leveraging this insight, they developed Attention-Driven Reranking (AttnRank), a training-free method that improves multi-hop QA accuracy to 44.72% on HotpotQA and increases Joint Goal Accuracy by 1.58% in few-shot in-context learning by strategically reordering documents.

Key insights

• Large Language Models consistently exhibit an "attention basin" phenomenon, a U-shaped attention distribution where they systematically focus on the beginning and end of *structural blocks* within long input contexts.
• There is a direct mathematical and empirical relationship between the attention allocated to a correct document and the likelihood of the LLM generating the correct answer, providing a mechanistic explanation for positional bias.
• Shallow Transformer layers are the primary drivers of this positional bias, operating in a "position-dominated regime," making their attention patterns key indicators for reranking.

Summary by alphaXiv.