Analysis of Structural Flaws in the Attention Mechanism

Attention sinks, Representation collapse, and Norm stratification, which are pointed to as major bottlenecks in Transformers, were found to be not flaws in the Attention mechanism itself, but rather geometric characteristics of Content-based routing that uses a norm-blind similarity measurement method.

According to the research, standard Dot-product attention uses a norm-blind metric that does not take into account the magnitude of the Key. As a result, the model produces the following compensatory effects, forming structural bottlenecks.

• Routing concentration: Routing becomes concentrated on a small number of positions
• Representation collapse: Representations collapse into a low-rank subspace
• Norm stratification: Key norms become stratified, with magnitude information that is formally ignored ending up serving as a back-channel

This phenomenon was confirmed to be a universal characteristic commonly observed across various routing structures, from various Transformer models ranging from GPT-2 to Pythia 2.8B, to Graph Attention Networks (GAT).