CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction
Researchers have introduced CMI-RewardBench, a new benchmark designed to evaluate music reward models that can handle complex multimodal instructions. This benchmark is accompanied by two datasets, CMI-Pref-Pseudo and CMI-Pref, to facilitate fine-grained alignment tasks. The team also developed CMI reward models (CMI-RMs), a parameter-efficient model family that demonstrates strong correlation with human judgments on musicality and alignment, and can be effectively scaled using top-k filtering. AI
IMPACT Enhances evaluation capabilities for multimodal music generation, potentially leading to more sophisticated AI music tools.