A novel causal framework that simultaneously reduces language and vision biases in VQA systems. Our method achieves 2x accuracy improvement on numerical questions while maintaining robust multimodal reasoning.