[Purpose/Significance] Faced with emergent online rumors, large language models are restricted by static knowledge boundaries and fail to produce accurate judgments. Traditional static recognition schemes lack real-time fact-checking capability and cannot adapt to emerging and diversely expressed new rumors. This study designs and implements an integrated retrieval-learning dynamic rumor recognition paradigm to improve the model’s recognition accuracy for newly emerging rumors and facilitate automatic governance of online public opinions. [Method/Process] Taking the fine-tuned Qwen3-8B as the base model, this study constructed a three-category discrimination mechanism including rumor, non-rumor and unknown. If the input information is judged as unknown by the model, the system automatically activated the Web search engine to capture evidence from authoritative sources, and took the retrieved results as real-time context for the model to conduct secondary accurate evaluation. The system structurally stored every successful query, supporting evidence and discrimination record to build an incremental sample set. Lightweight low-rank adaptive fine-tuning technology was adopted to periodically update the base model, so as to realize the integrated coordination of retrieval augmentation and continual learning, and form a complete closed loop of real-time verification, sample accumulation and model iteration. [Result/Conclusion] Experimental results show that the proposed method achieves excellent performance in the rumor recognition task, with an F1-score of 0.96 and an FTCA index of 94.8, which is significantly better than many mainstream baseline models. When dealing with sudden and novel rumors, it can effectively reduce the proportion of unknown judgments and presents favorable stability and practicability. It can provide technical references for the intelligent identification of dynamic online disinformation, and also offer new ideas for optimizing retrieval-augmented rumor detection models.