Context Reconstruction and Hardware-Aware Memory Optimization for Cross-Lingual Visual OpenQA
Analysis reveals that combining high-resolution memory optimization with a carefully designed QRA prompt template plays a crucial role in stimulating and enhancing the cross-lingual reasoning performance of small models, providing a new perspective for multimodal reasoning tasks.