DeepSeek V4 Pro scaffolding improves small model's procedural skills on different tasks

A viral experiment from a researcher tested procedural skill transfer between AI models using a novel scaffolding technique. The setup used DeepSeek V4 Pro as the larger model and examined two very different Three.js tasks: a complex character dance scene (Thriller-style with multiple figures, lighting, and staging) and a low-poly mechanical object (BMPT-72 turret with recognizable silhouette). The protocol involved first generating baseline outputs from both models on both tasks. Then, without seeing the second task, the larger model analyzed the smaller model's weakness on the first task and created a procedural scaffold—a set of step-by-step guidelines emphasizing planning, hierarchy, composition, and visual structure. The smaller model then re-attempted both tasks using the scaffold.

The results were clear: for a weaker small model, the scaffold dramatically improved its output on the completely unrelated second task. Unlike the larger model, which already had strong internal planning depth (so the scaffold only added polish), the smaller model gained structural integrity—proper geometry proportions, better camera placement, and coherent scene organization. The renders showed visibly improved silhouette and composition for the tank turret, and far more believable character staging. The test was manual and not a formal benchmark, but the effect was strong enough to warrant automated testing. This approach could reduce the need for expensive fine-tuning, allowing smaller models to leverage the planning habits of larger ones via a simple prompt-time guide.