OmniCapBench: A Deep-Structured Evaluation Framework for Fine-Grained Audio-Visual Captioning
OmniCapBench is a new evaluation framework for AI radio models that moves beyond simple text scores to test audio-visual reasoning. It breaks down captions into atomic units like entity references, audio events, and visual shots. This structured approach lets researchers precisely locate errors in things like identity drift or cross-modal misalignment, providing a clear roadmap for improving multimodal AI.