Large Language Models (LLMs) have made incredible strides in understanding text and images, but videos? That's a whole other ball game. Videos involve not just *what* is in a scene, but also *how* things change over time, requiring a deeper level of understanding called compositional reasoning. Current Video-LLMs excel at basic tasks like captioning, but struggle with complex questions that involve multiple steps of reasoning about the relationships between objects, actions, and events. Think about a question like, "After the woman fell, what did she do with the red box?" Answering requires understanding the sequence of events, the objects involved, and their relationships. That's where new research on a method called STEP comes in. STEP helps Video-LLMs build up their reasoning skills through a clever self-training process. First, it extracts the essence of a video by creating a Spatio-Temporal Scene Graph (STSG), essentially a map of the video’s objects, actions, and their relationships across time. Then, using this map, STEP generates practice questions and answers, along with step-by-step explanations (called "rationales"), similar to how a teacher might guide a student. By training on these self-generated examples, the Video-LLM learns to reason more effectively about the complex interplay of elements within a video. Experiments show that STEP dramatically improves Video-LLMs' performance on compositional reasoning tasks, especially those requiring multiple steps of inference. Importantly, STEP works across different Video-LLM architectures and requires minimal human input, paving the way for more adaptable and intelligent video understanding AI. While promising, challenges remain. Balancing the complexity of generated questions, ensuring the accuracy of the scene graphs, and scaling the approach to massive video datasets are key areas for future research. But STEP represents an exciting step towards AI that truly understands the world in motion.