Processing input sequences much longer than a model's training context window while maintaining accuracy and efficiency.