Researchers led by the University of California, Davis, have developed a method called Dynamic Blocked Attention Sparsity via Softmax Thresholding, or BLASST, that allows AI models to skip low-value computations, making models up to 50% faster without needing to retrain models or introduce new hardware. The method, detailed in a paper published on arXiv, works by reusing information the model has already computed to identify which attention scores contribute little value, reducing the quadratic growth of computation as prompts lengthen. The work received the Best Paper Award during MLSys 2026, and NVIDIA is already integrating the method into its AI software. Cameron Shinn, an electrical and computer engineering Ph.D. student working in John Owens’ research group, explained that the approach focuses on tweaking model operations to fit more naturally with hardware rather than just making existing mathematical operations faster.