LLM-Guided Distillation Enhances Multimodal Fake-News Detection

Researchers propose LLM-MRD, a teacher-student framework for multimodal fake-news detection, submitted March 10, 2026. The teacher generates multi-view reasoning chains from an LLM across textual, visual and cross-modal perspectives; a calibration distillation transfers reasoning to an efficient student. Experiments report average improvements of 5.19% accuracy and 6.33% F1-Fake across benchmarks, and the authors released code.