Skip to content
Book Open access

DART: Difficulty-Aware Adaptive Reasoning for Music Information Retrieval

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 23 references

Abstract

Large audio-language models (LALMs) have advanced music understanding and driven progress in music information retrieval (MIR), which requires integrating low-level acoustic perception and high-level semantic interpretation. These two levels impose markedly different reasoning demands: over-deliberation on perceptual tasks can induce hallucinations, whereas under-reasoning on semantic tasks leads to shallow predictions. Adaptive reasoning is therefore crucial for MIR, yet remains largely unexplored. Pervasive annotation noise in standard benchmarks further compromises training and evaluation. Moreover, existing approaches typically treat the allocation of reasoning effort as a black-box optimization problem, providing little insight into how models assess difficulty internally. To address these limitations, we propose DART, a difficulty-aware adaptive thinking framework for music tagging that further shows promising results when transferred zero-shot to other MIR tasks. First, we construct CLEAR-MusicTags, an LLM-assisted consensus-based re-annotation of existing benchmarks that covers both training and evaluation splits, along with an expert-annotated test subset (ExpertEval-MusicTags) for gold-standard evaluation. Second, we introduce a Rubric-guided Adaptive Thinking mechanism via Group Relative Policy Optimization (GRPO). Leveraging a teacher-aligned rubric, our approach trains the model to dynamically calibrate whether to invoke reasoning. Third, we further identify internal difficulty directions within model activations and show that steering along these vectors at inference time enables fine-grained control over reasoning behavior without retraining. Extensive experiments show that DART achieves state-of-the-art performance on refined benchmarks while providing an interpretable approach for complex music understanding.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.