AXPO Method Improves Multimodal Agent Reasoning in Vision-Language Models