MQAScore: fine-grained text-music alignment eval across 5 dimensions via LALM
affige_yang · x · 2026-10-28
A NTU GICE team (Wei-Jaw Lee, Yi-Hsuan Yang et al.) introduces MQAScore, submitted to ICASSP 2027, for fine-grained text-to-music alignment evaluation.
- Global similarity scores (CLAP, MuQ) collapse all attributes into one number, hiding failure modes like "right mood, wrong instrument"
- MQAScore decomposes captions into atomic musical concepts and scores each along five dimensions (genre, instrument, mood/theme, vocal, usage) via the probability that an LALM answers a yes/no question affirmatively
- Verification labels are built from presence/absence annotations and LLM-guided concept substitutions, with a Sensitivity-Invariance Ratio measuring evaluator selectivity
- Experiments show concept-level scoring improves both QA- and similarity-based evaluators
Paper, demo, and code are public.
More from Research
- Rollback: World Models Rewind Robot Decisions to Explore Counterfactual Actions — agibot-world · 2026-10-28
- SPOT Speeds Up Diffusion Language Models Up to 6.8x via Commit-Order Prediction — Eliezer Shlomi · 2026-10-28
- RSI Arena: 10 AI Agents Autonomously Post-Train Models, Best Closes 76% of Gap — RSIArena · 2026-10-28
- HaMo Combines Experience Cards with Model Adaptation for Long-Horizon Web Agents — Zhen Bi · 2026-10-28
- ChronoReplay Fixes Interface Lock-in in Spatial VLMs, +13.41 on OST-Bench — LINs-lab · 2026-10-28
- Higher Reasoning Effort Costs Up to 42x and Can Break Tasks Lower Effort Solves — togethercomputer · 2026-10-28