Large Audio Language Models (LALMs) are designed to reason over speech, environmental sounds, music, and other acoustic information. But their audio encoders generate dense token sequences, and that density makes long-form audio costly to process. The computational burden grows quickly as audio length increases, limiting practical use.
AudioScorer, described in a new arXiv preprint, tackles this by learning question-aware token relevance. Instead of treating all audio tokens equally, the method scores tokens based on how relevant they are to the specific question or task at hand. This allows the model to focus on the informative parts of the audio and reduce the number of tokens that need to be processed.
The paper's abstract does not provide full experimental details, so the reported efficiency gains are not yet quantified here. But the core idea is clear: by making token selection depend on the query, AudioScorer aims to cut the cost of long-form audio reasoning while preserving the model's ability to answer correctly.