Light-MER: Smaller MLLMs Achieve State-of-the-Art Emotion Recognition.

Recent advancements in multimodal large language models (MLLMs) have significantly boosted performance in multimodal emotion recognition (MER), enabling detailed descriptions by integrating video, audio, and language. However, these performance gains typically come with a substantial increase in model size, often exceeding 7 billion parameters, leading to high computational costs and reduced inference efficiency. This poses a significant challenge for real-time deployment on devices with limited resources, such as robots and mobile phones. This research directly questions the assumption that larger models are inherently necessary for high-quality MER. It introduces Light-MER, a lightweight MER framework that achieves superior and faster multimodal sentiment understanding and recognition through an innovative knowledge distillation process. Light-MER transfers complex knowledge from a powerful, large-scale teacher model to a much smaller, sub-billion-parameter student model. The framework incorporates two key optimization strategies to enhance knowledge transfer: an optimal transport loss combining Sliced Wasserstein Distance with hidden-state alignment, and a multi-reward optimization strategy based on GRPO to balance MER performance and efficiency. Extensive experiments across nine benchmark datasets confirm that Light-MER not only achieves state-of-the-art performance but also drastically improves inference efficiency, demonstrating the strong potential of smaller MLLMs for future applications.