MusTBENCH: Benchmarking and Advancing Temporal Grounding in Music LLMs

Researchers introduce MusTBENCH, a benchmark for evaluating temporal grounding capabilities in Large Audio-Language Models (LALMs) for music understanding, and propose MusT, an optimization framework that significantly improves model performance on time-sensitive musical tasks like instrument entries and rhythmic transitions.

The research identifies a critical gap in current music understanding AI systems: while Large Audio-Language Models have made strides in general music comprehension, they struggle to accurately pinpoint when specific musical events occur within audio files. This temporal grounding challenge represents a fundamental limitation for any AI system intended to provide precise, actionable musical analysis—a requirement for music production, DJing, musicology research, and content creation applications.

The development of MusTBENCH addresses a real need in AI evaluation frameworks. Most existing benchmarks focus on what models understand about music rather than when they understand it, leaving a blind spot in capability assessment. The introduction of a music-expert-validated benchmark sets a higher bar for model evaluation and creates accountability for developers claiming advanced music understanding capabilities.

MusT's four-stage optimization approach—combining music encoder adaptation, LLM fine-tuning, supervised learning, and reinforcement learning—demonstrates that temporal grounding can be systematically improved rather than treated as an inherent architectural limitation. This has practical implications for music-related AI applications, from music recommendation systems requiring event-level precision to AI-assisted music composition tools.

For the broader AI industry, this research highlights how specialized benchmarks reveal hidden weaknesses in general-purpose models. As LALMs become more prevalent, domain-specific evaluation frameworks like MusTBENCH become increasingly important for understanding where these systems actually fall short versus where they merely appear capable. The work establishes temporal reasoning as a key research direction for audio-language models.

• →Existing Large Audio-Language Models lack precise temporal grounding, failing to accurately identify when musical events occur within audio
• →MusTBENCH provides the first expert-validated benchmark specifically designed to measure temporal grounding capabilities in music AI systems
• →The proposed MusT optimization framework achieves significant performance improvements through a four-stage training and fine-tuning process
• →Temporal grounding is critical for music applications like production, DJing, and content creation where event timing matters
• →This research reveals how general-purpose large language models struggle with domain-specific timing requirements in audio understanding