Abstract: Heterogeneous graphs are instrumental in mining complex correlations within graph data, which holds significant importance in real-world applications. Traditional Graph Neural Networks (GNNs), however, are often confined to predefined tasks and rely on clearly labeled data along with fixed training mechanisms. This inherent limitation reduces their flexibility when dealing with open-world tasks. Moreover, existing research on integrating GNNs with Large Language Models (LLMs) predominantly focuses on homogeneous text-attributed graphs, failing to account for node heterogeneity. To address the challenges of misaligned heterogeneous feature representation spaces and open-domain task generalization, we propose a multi-hop graph reasoning mechanism that combines heterogeneous graph learning with contrastive learning. Specifically, the model decouples and reconstructs heterogeneous subgraphs based on meta-path symmetry. It achieves efficient fusion of topological embeddings and semantic representations through a differentiated attention mechanism and a hierarchical feature aggregation algorithm. To tackle modal misalignment, a progressive phase optimization strategy is adopted to train the graph query transformer, while a contrastive learning method is employed to bridge modal differences. Fine-grained feature associations are established via self-supervised image-text matching, and language modeling objectives are incorporated to enable the model to generate accurate answers to queries. Experimental results demonstrate that the proposed model exhibits strong adaptability to both predefined tasks and open-scene generalization. It also shows high-quality reasoning capabilities when addressing unseen questions in heterogeneous network question-answering tasks.