Why diffusion LLMs' performance cannot beat Auto Regressive LLM now?