Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs | Read Paper on Bytez