Evaluating the Impact of Multimodal Cognitive and Informational Interventions on Rumor Detection in Large Language Models: A DeepSeek Case Study

With the development of social media, social media rumors can be spread more rapidly. It may disrupt personal lives and threaten social stability. Traditional rumor detection methods rely on large-scale labeled data and face challenges such as poor generalization and high transfer costs. Recently, large language models (LLMs) with strong zero-shot capabilities have provided a novel approach for rumor detection. This study uses the Liar dataset and DeepSeek-V3 model to explore how DeepSeek's deep reasoning and web search functionalities impact rumor detection accuracy. A multi-level evaluation system combining Kappa, weighted Kappa, and Spearman's rank correlation is developed to reduce incidental fluctuations from relying on a single metric. The results show that the model performs best in the “NoIntervention” setting, and external information or reasoning interventions did not significantly improve accuracy. This innovative evaluation framework with fine-grained analysis highlights its importance in improving evaluation precision and revealing model performance, offering significant methodological contributions.