A New Interpretation of the Self-Attention Mechanism in Transformers  
Author

Maoyuan Zhou

 

Co-Author(s)

Xinyu Hu;  Siqi Liu; Sheng Fu; Yonggang Ji

 

Abstract This study presents a novel perspective on interpreting the query and key matrices in Transformers. Traditional interpretations often treat these matrices as independently defined, lacking a robust mathematical foundation. This limitation introduces ambiguity in understanding feature relationships and complicates the model’s decision-making process. In contrast, we propose an interpretation based on a bilinear model that directly links the definitions of the query and key matrices to the correlations between input features. This framework clarifies the mutual influence of features, making the weight computation process more intuitive. By enhancing the model’s interpretability, our approach offers a clearer insight into its behavior and decisionmaking logic.

 

Keywords bilinear model, query matrix, key matrix
   
    Article #:  RQD2026-172
 

Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design
August 5-7, 2026