The models wouldn't need to adapt their training at all; this is a secondary model trained on their activations. It's electrodes stuck into the model's brain. The original article I posted is built on top of a regular LLM (OpenAI's GPT-OSS)