Semantic-Guided Graph Reasoning with Large Language Models for Social Bot Detection

摘要

Detecting social bots in large-scale social networks has become increasingly challenging due to the emergence of human-like textual behaviors and complex, heterogeneous interaction patterns. Existing graph-based approaches often rely on structural homophily or purely learned attention mechanisms, which can be unreliable in realistic scenarios where both textual and structural signals are noisy and weakly aligned. In this work, we propose a dual-view framework that integrates large language models (LLMs) and relational graph neural networks (GNNs) for robust social bot detection. Rather than treating LLMs as generic feature extractors, we leverage their semantic reasoning capability to derive high-level node-level priors from user profiles and historical content, capturing latent bot-related cues such as automation style, identity consistency, and coordination intent. These semantic priors are further extended to the relational level, where LLMs estimate the reliability of social connections to guide relational message passing in GNNs. By jointly incorporating node-level semantic signals and edge-level reliability guidance, our framework enables GNNs to selectively aggregate information from trustworthy neighbors and to validate individual behavioral cues through collective structural evidence. Extensive experiments demonstrate that the proposed approach consistently outperforms strong multimodal and graph-based baselines, achieving notable improvements in detection accuracy and robustness. Further analyses show that our framework provides interpretable insights into both individual bot behaviors and social interaction patterns.