Detecting Miscitation on the Scholarly Web through LLM-Augmented Text-Rich Graph Learning

A research team has introduced LAGMiD (LLM-Augmented Graph Learning-based Miscitation Detector), a novel framework that tackles the growing problem of miscitation in academic literature. The system cleverly combines the semantic reasoning power of large language models (LLMs) with the structural analysis capabilities of graph neural networks (GNNs) to detect when research papers incorrectly cite sources that don't actually support their claims. At its core, LAGMiD employs an evidence-chain reasoning mechanism using chain-of-thought prompting to perform multi-hop citation tracing and assess semantic fidelity between claims and their cited sources.

To overcome the computational expense and hallucination risks of pure LLM approaches, the researchers designed a knowledge distillation method that aligns GNN embeddings with intermediate LLM reasoning states. This allows the system to capture nuanced relationships between citation context and network structure while maintaining efficiency. A collaborative learning strategy further optimizes the framework by routing complex cases to the LLM while training the GNN for structure-based generalization.

Experiments across three real-world benchmarks demonstrate that LAGMiD achieves state-of-the-art performance in miscitation detection while reducing inference costs by approximately 90% compared to using LLMs alone. The framework represents a significant advancement over previous methods that relied primarily on semantic similarity or network anomaly detection, which often missed the complex relationships between citation context and the broader scholarly network.