MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models