Reading Notes of DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference