Researchers challenge the assumption that large models are necessary for high-quality multimodal emotion recognition (MER) by proposing Light-MER, a lightweight framework that transfers knowledge from a large teacher to a sub-billion-parameter student via distillation.
The method introduces two optimization strategies: an optimal transport loss combining Sliced Wasserstein Distance with hidden-state alignment, and a multi-reward strategy based on GRPO to balance performance and efficiency. Extensive experiments on nine benchmark datasets show Light-MER achieves state-of-the-art performance while significantly improving inference efficiency for resource-constrained platforms.
This work highlights the strong potential of small multimodal emotion language models for future research, enabling interpretable description generation without the high computational costs of larger models.