Xiang Wang

With the rapid development of AI technologies, graph foundation models have been widely applied in several cutting-edge fields, achieving remarkable progress. However, these models face a credibility crisis characterized by unreliable decision-making, unstable reasoning, opacity, and a lack of human-centered values, which hampers their deep application and active development in critical areas. Focusing on the forefront of "Trustworthy Graph Foundation Models," Xiang Wang's research aims to address these issues​.

The modality misalignment issue arises because there are often differences in how information is expressed among multimodal data. This makes it difficult for existing large language models to effectively understand and integrate graph-structured data. To address this, Xiang Wang designed a fine-tuning approach oriented toward modality alignment and developed multimodal AI-for-Science large models capable of processing and generating graph data such as chemical molecules and proteins, applicable in drug discovery and material science​.

The value misalignment with human-centered principles issue arises because existing large language models often deviate from human values, potentially leading to biases or injustices in decision-making processes. To address this, Xiang Wang designed robust preference learning methods that optimize the training process of large models, identifying preference data that aligns with human values and behaviors from noisy data. This ensures that large models better align with human values and preferences when making predictions or decisions, promoting their trustworthy application in fields like social science and fintech​.

In the next phase, Xiang Wang will explore large model-driven agent gaming technology, enabling agents to understand and simulate complex gaming scenarios such as economic markets, strategic games, and social interactions. This will enhance the accuracy and reliability of models in real-world applications, providing trustworthy support and effective deductions for human decision-making​.