A manual experiment tested whether a procedural scaffold generated by a large model can transfer planning discipline to smaller models without fine-tuning or revealing the target answer. The results indicate that this approach significantly improves structural readability and composition in small models when applied across different Three.js domains.
- The test compared outputs from DeepSeek V4 Pro, Qwen 27B, and a quantized 35B A3 model on character choreography and mechanical turret tasks.
- For the larger DeepSeek V4 Pro model, the scaffold provided minor polish to lighting and art direction but had limited impact on structure.
- Smaller models (Qwen 27B and 35B Q3_K_M) showed major improvements in scene hierarchy, object separation, and silhouette clarity without the scaffold producing generic or confused outputs.
- The transferred skills were abstract planning steps, such as defining scene contracts, building in layers, and auditing final output, rather than copying specific domain details.
The author concludes that large models can externalize part of their internal planning discipline into a reusable inference-time structure, helping smaller models organize their existing knowledge more reliably.