UMAF: Adapting LLMs Without Fine-Tuning

I'm experimenting with an approach called Universal Model Adapter Fusion to transfer capabilities LLMs without fine-tuning or distillation. The idea's to pull behaviors - like task-specific smarts or reasoning styles -from one model and patch them onto another efficiently. Think cheaper model tweaks, mixing strengths across architectures, or building custom skill sets without big retraining jobs, using an extended lora approach. Take a look at github.com/millionthodin16/umaf. It's early, but the benefits could be big. Anyone see value in this kind of approach for LLMs?