Imagine teaching a computer to understand human actions just from stick figures. That’s the challenge of skeleton-based action recognition. While efficient, skeletons lack the rich detail of full videos, limiting their accuracy. Researchers have tried using multiple data types like RGB images and depth maps, but this makes the system bulky and slow. A new approach, Multi-Modality Co-Learning (MMCL), uses the strengths of multiple data sources during training while keeping the speed of skeletons for actual use. The secret? Leveraging the power of large language models (LLMs). During training, MMCL uses a clever two-step process. First, it matches features from RGB images with skeleton data, helping the system understand how the stick figures relate to real movements. Second, it feeds images and instructions into powerful multimodal LLMs. These LLMs analyze the image and generate descriptions, offering valuable insights to refine the system’s understanding. This is like giving the computer a coach that can explain the nuances of each action. This method achieves higher accuracy than previous skeleton-based methods, even rivaling more complex, multi-data approaches. Impressively, it maintains this performance even with limited data, highlighting the LLMs’ robust generalization ability. This innovation opens doors for more efficient and adaptable action recognition systems. Imagine lightweight AI that can understand actions on your phone or in robots, all thanks to the clever combination of simple skeletons and powerful LLMs.