Exploring System 1 and 2 communication for latent reasoning in LLMs