Ying Wang

Knowledge distillation technology shows significant potential in transferring the knowledge of graph neural network (GNN) teacher models to multi-layer perceptron (MLP) student models. However, the existing methods are primarily designed for node classification task and are difficult to directly apply to graph classification. Graph classification task requires knowledge transfer via the generation of graph-level soft labels. Such global supervisory signals are inherently sparse, making it difficult for the student model to capture fine-grained structural information. Additionally, the student MLP model lacks the explicit ability to handle graph structures and has a significantly different representation space from the teacher GNN which further complicates the knowledge transfer process. We address these problems by proposing multigranularity distillation with a dynamic prior knowledge framework named MuDPK. The proposed framework integrates multigranularity structural information with the dynamic prior knowledge mechanism to effectively address the challenges of sparse supervision signals and the representation bottleneck in graph classification. We first design an adaptive graph propagation-based node feature enhancement module to enhance the representation capability of the student model in a limited feature space. We then design a triple distillation loss function that includes graph-level, subgraph-level, and node-level components. We alleviate the problem of sparse supervision signals by hierarchically aligning the structural representations of the teacher-student model at different levels. Finally, we use the intermediate layer features of the teacher as both the target representation and input knowledge, dynamically adjusted by our prior knowledge mechanism. Knowledge transfer is achieved through feature alignment while the student model is regulated by adjusting the proportion of injected knowledge in a controlled manner. Extensive experiments on eight benchmark graph datasets demonstrate the effectiveness and superiority of MuDPK, which achieves an average accuracy gain of +1.37% over the baselines.

Retrieval-augmented generation (RAG) has facilitated large language models (LLMs) by grounding facts in external knowledge. While prior RAG strategies can achieve dynamic control by detecting sentence-wise hallucination, thereby reducing unnecessary computation, they ignore the entity, statement, and their combinations. In this work, we analyze adaptive RAG methods through probing uncertainty to explore real-time entity- and statement-level verification, finding that such information can serve as training-free signals for hallucination detection, and the entropy probes can also guide principled document ranking. Inspired by these insights, we propose SeaRAG, a training-free, adaptive RAG framework that dynamically detects and mitigates hallucinations by probing LLMs at both entity and statement aspects, ranking retrieved passages via entropy-based uncertainty reduction, and regenerating evidence-grounded responses in real-time. Experiments across various Question-Answering benchmarks and multiple LLMs demonstrate consistent accuracy improvements and reduced retrieval rates. SeaRAG with self-correction outperforms Always Retrieve by up to 11.1% on TriviaQA while cutting retrieval frequency from 45.2%, offering an efficient and real-time hallucination control.