💡 To address the issue that diffusion-based LLMs cannot reuse standard KV caches due to bidirectional attention, we propose a training-free dual adaptive caching framework that selects tokens in a two-s