lalms

MusTBench is a benchmark for evaluating temporal grounding in Large Audio-Language Models (LALMs) for music understanding. The authors propose MusT, a four-stage training recipe that significantly improves temporal grounding performance over existing models.