Which form of communication should Large Language Models (LLMs) use when collaborating on a problem, e.g., one or two models processing the context and another model answering a question without seeing the context itself: plain text, hidden state, or some other way? (1/5)