A comprehensive research framework for analyzing whether averaging adjacent tokens in a LLM can reduce the compute compared to standard model.