Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding

TL;DR

This paper reveals that massive values in self-attention modules are crucial for understanding contextual knowledge in large language models, influenced by Rotary Positional Encoding, and impacts model interpretability and design.

Contribution

It uncovers the emergence and significance of massive attention values in Q and K, linked to RoPE, and their role in contextual understanding rather than parametric knowledge retrieval.

Findings

Massive values consistently emerge in specific attention regions.

Ignoring massive values reduces performance on contextual tasks.

Rotary Positional Encoding causes the concentration of massive values.

Abstract

Large language models (LLMs) have achieved remarkable success in contextual knowledge understanding. In this paper, we show that these concentrated massive values consistently emerge in specific regions of attention queries (Q) and keys (K) while not having such patterns in values (V) in various modern transformer-based LLMs (Q, K, and V mean the representations output by the query, key, and value layers respectively). Through extensive experiments, we further demonstrate that these massive values play a critical role in interpreting contextual knowledge (knowledge obtained from the current context window) rather than in retrieving parametric knowledge stored within the model's parameters. Our further investigation of quantization strategies reveals that ignoring these massive values leads to a pronounced drop in performance on tasks requiring rich contextual understanding, aligning with our analysis. Finally, we trace the emergence of concentrated massive values and find that such concentration is caused by Rotary Positional Encoding (RoPE), which has appeared since the first layers. These findings shed new light on how Q and K operate in LLMs and offer practical insights for model design and optimization. The Code is Available at https://github.com/MingyuJ666/Rope_with_LLM.