A priority map is all you need: Exploring the roots of neural mechanisms underlying transformer-based large language models.