QA-ViT is a Question Aware Vision Transformer designed to enhance multimodal reasoning by integrating question awareness directly into the vision encoder. This approach enables dynamic visual features that focus on relevant image aspects based on the posed question, improving alignment with user queries. QA-ViT is model-agnostic and can be efficiently integrated into any Vision-Language (VL) architecture. Experiments show that QA-ViT consistently improves performance across various tasks and benchmarks, demonstrating its effectiveness in enhancing visual and scene-text understanding. The method addresses the limitation of existing VL architectures where vision encoding is decoupled from user queries, leading to suboptimal alignment with query-specific elements. By conditioning the vision encoder on textual prompts, QA-ViT achieves more accurate predictions, as illustrated by GradCAM analysis showing focused attention on relevant image regions. The method is implemented by encoding the question into features, fusing them with the vision encoder, and projecting the results into an LLM for task completion. QA-ViT is tested on multiple VL architectures, including BLIP2, InstructBLIP, and LLaVA-1.5, showing consistent improvements. The approach is also effective in zero-shot settings and across different model sizes, demonstrating its versatility. Ablation studies confirm the method's effectiveness, showing that question-aware visual features significantly contribute to performance gains. QA-ViT's integration into the vision encoder allows for better alignment with queries, leading to improved performance in both general and scene-text tasks. The method is compatible with various LLM architectures and has shown superior performance compared to existing generalist models. Overall, QA-ViT represents a significant advancement in question-aware vision modeling, enhancing the ability of VL models to understand and respond to visual and textual information more effectively.QA-ViT is a Question Aware Vision Transformer designed to enhance multimodal reasoning by integrating question awareness directly into the vision encoder. This approach enables dynamic visual features that focus on relevant image aspects based on the posed question, improving alignment with user queries. QA-ViT is model-agnostic and can be efficiently integrated into any Vision-Language (VL) architecture. Experiments show that QA-ViT consistently improves performance across various tasks and benchmarks, demonstrating its effectiveness in enhancing visual and scene-text understanding. The method addresses the limitation of existing VL architectures where vision encoding is decoupled from user queries, leading to suboptimal alignment with query-specific elements. By conditioning the vision encoder on textual prompts, QA-ViT achieves more accurate predictions, as illustrated by GradCAM analysis showing focused attention on relevant image regions. The method is implemented by encoding the question into features, fusing them with the vision encoder, and projecting the results into an LLM for task completion. QA-ViT is tested on multiple VL architectures, including BLIP2, InstructBLIP, and LLaVA-1.5, showing consistent improvements. The approach is also effective in zero-shot settings and across different model sizes, demonstrating its versatility. Ablation studies confirm the method's effectiveness, showing that question-aware visual features significantly contribute to performance gains. QA-ViT's integration into the vision encoder allows for better alignment with queries, leading to improved performance in both general and scene-text tasks. The method is compatible with various LLM architectures and has shown superior performance compared to existing generalist models. Overall, QA-ViT represents a significant advancement in question-aware vision modeling, enhancing the ability of VL models to understand and respond to visual and textual information more effectively.