ICLR2026AlgoTask-related Token Compression in Multi-modal Large Language Models from an Explainability Perspective