Researchers propose a method called Triage that can predict and prune audio tokens in large audio language models (LALMs) before the language model runs, achieving significant compression and improving performance.