Unveiling the Latent Directions of Reflection in Large Language Models