Given the lower energy requirements and the one-shot learning capabilities of functor models the LLM Offloading Pattern is quite useful. It suggests an alternative to scaling for the LLM in certain cases. Here agents work with functor micro models to discover new data. The data processor handles curation: cleansing, dedups, canonicalization, learning preparation, etc.

​

Then the learning router submits the learning unit to the appropriate SLM. The LLM can invoke inference on any SLM to gain knowledge, generate code, etc. The Graph Neural Networks are contained in a new type of ensemble that allows messaging between nodes in different models using an orchestrator. See this document and the diagram below for further details.