The authors systematically validate LVM's scalability and multi-task capabilities through a series of experiments. The results show that as model size scales from 300M to 3B parameters, training loss (perplexity) consistently decreases, and downstream task performance — including semantic segmentation, depth estimation, surface normal estimation, and edge detection — improves significantly, demonstrating that scaling both model and data size effectively enhances generalization. Through Sequential Prompting, LVM can predict subsequent video frames from preceding ones or perform 3D rotation prediction, with longer context (e.g., 11+ frames) improving prediction accuracy. Analogy Prompting enables the model to generalize to tasks it was never trained on, such as keypoint detection and image segmentation, achieving a PCK score of 81.2% on the Pascal 3D+ dataset — outperforming some specialized models. Furthermore, the model can handle compositional tasks (e.g., combining rotation and keypoint tracking) and unconventional prompts (e.g., object copying or visual reasoning tasks), exhibiting preliminary reasoning capabilities. However, limitations such as prompt ambiguity and insufficient generation quality persist, highlighting the need for further scaling.