VISTA: Enhancing Visual Conditioning via Track-Following Preference Optimization in Vision-Language-Action Models