A Dual-Path Approach to Fake News Detection Based on Ensemble Learning and Fine-Tuned Large Language Models

The rapid proliferation of misinformation in today’s digital ecosystem poses threats to public’s trust on institutions and informed decision-making. To mitigate these challenges, the development of efficient and automated fake news detection systems has become increasingly essential. This study presents a dual-path framework, FIND, an ensembled model, combines Random Forest classifiers, fine-tuned BERT models, and prompt-engineered LLMs to address mainstream detection tasks with high accuracy. Additionally, we fine-tune LLMs to adapt to long-tail news (ie news with uncommon topics) distributions, enhancing model generalizability and scalability. Experimental evaluations on benchmark datasets confirm the robustness and effectiveness of the proposed method in handling complex fake news, proving its applicability in real-world scenarios.