Complete contemporary multimodal-language-model evaluation

Complete a comparison of contemporary multimodal language models on the identity-masked Music Understanding Benchmark, including evaluation of tempo, key, time signature, and structural segmentation performance.

Background

The benchmark provides local DSP and answer-distribution baselines but does not report results from contemporary multimodal LLMs. Consequently, the benchmark has not yet demonstrated how such models perform on its four musical-understanding tasks. Completing this comparison is necessary to establish the empirical usefulness of the benchmark for evaluating current multimodal systems.