How does recurrent neural network implement model based RL system purely in its activation dynamics (in blackbox meta-rl setting)?

Question

I have read these papers "learning to reinforcement learn" and "PFC as meta RL system". The authors claim that when RNN is trained on multiple tasks from a task distribution using a model free RL algorithm, another model based RL algorithm emerges within the activation dynamics of RNN. The RNN with resulting activations acts as a standalone model based RL system on a new task(from the same task distribution) even after freezing the weights of outer loop model free algorithm of that. I couldn't understand how an RNN with only fixed activations act as RL? Can someone help?