This study reveals and quantifies key factors in visual token pruning from a geometric covering perspective, and proposes Multi-Objective Balanced Covering, which significantly accelerates MLLM reasoning with negligible performance loss.