EmoMM Benchmark Reveals Video Contribution Collapse in MLLMs for Emotion Recognition

A new benchmark for multimodal emotion recognition (MER), named EmoMM, has been unveiled by researchers, focusing on challenges like modality conflict and missing data. Their findings highlight a phenomenon termed Video Contribution Collapse (VCC), where multimodal large language models (MLLMs) downplay video information due to excessive token redundancy and specific modality biases. To counter this issue, they introduce CHASE (Conflict-aware Head-level Attention Steering), an efficient mechanism that identifies modality conflicts and adjusts attention during inference without the need for retraining. Experimental results indicate that CHASE reliably enhances performance on this benchmark.