Quantum-Enhanced Cross-Attention for Multi-Modal Image-Text Retrieval

Multi-modal systems for vision-language tasks of- ten struggle with semantic ambiguity. Overlapping or context- dependent features in cross-modal representations lead to poor alignment and performance. To tackle this, we introduce Quantum-Enhanced Fusion (QE-Fusion), a new cross-attention framework inspired by quantum interference principles. QE- Fusion works in a complex-valued domain. This allows it to capture both magnitude and phase information in cross-modal attention scores.By representing queries and keys as complex-valued tensors, QE-Fusion simulates both constructive and destructive interference through the magnitude of their complex dot product. This mechanism enables the model to dynamically boost relevant cross-modal correlations while reducing irrelevant ones. As a result, it achieves better feature disambiguation. Unlike standard attention mechanisms, which use real-valued scores, QE-Fusion’s phase-aware computation offers more flexibility for modeling complex feature interactions.The framework is implemented as a modular PyTorch layer, making it easy to integrate into existing multi-modal Transformer architectures. We validated our approach using the Flickr8k image-text retrieval benchmark. While initial training showed some instability, we improved this by adjusting the learning rate. After tuning, the stabilized QE-Fusion model showed clear and consistent performance gains compared to the standard attention baseline. These results position QE-Fusion as a strong and efficient method for strengthening the reasoning abilities of multi-modal systems. They also highlight the potential of quantum-inspired principles in traditional deep learning models.