“The MRKL neuro-symbolic architecture uses a general-purpose LLM as a router to send inquiries to the most suitable 'expert' module, which can be neural or symbolic.”