Crisp Attention: Regularizing Transformers via Structured Sparsity

Abstract: The quadratic computational cost of the self-attention mechanism is a primary challenge in scaling Transformer models. While attention sparsity is widely studied as a technique to improve computational efficiency, it is almost universally assumed to come at the cost of model accuracy. In this paper, we report a surprising counter-example to this common wisdom. By introducing structured, post-hoc sparsity to the attention mechanism of a DistilBERT model during fine-tuning on the SST-2 sentiment analysis task, we find that model accuracy improves significantly. Our model with 80\% attention sparsity achieves a validation accuracy of 91.59\%, a 0.97\% absolute improvement over the dense baseline. We hypothesize that this phenomenon is due to sparsity acting as a powerful implicit regularizer, preventing the model from overfitting by forcing it to make predictions with a more constrained and robust set of features. Our work recasts attention sparsity not just as a tool for computational efficiency, but as a potential method for improving the generalization and performance of Transformer models.