Metadata-Free Rumor Detection via LLM-Augmented Reasoning

Abstract

Rumor detection in social media is critical for countering the rapid spread of misinformation, especially in dynamic and low-resource settings. In this work, we propose a unified framework that leverages large language models (LLMs) not only for classification but also for semantically enriching the input data through lightweight, class-agnostic augmentation. Unlike prior methods that rely on user metadata, graph-based propagation structures, or external knowledge sources, our approach (MFRD) prompts LLMs to generate useful augments. Two high-level summaries for each event are generated: a discussion narrative distilled from stance-rich user replies, and a propagation analysis inferred from repost patterns and temporal dynamics. These augmentations are integrated into structured event documents that combine source posts, user feedback, and contextual reasoning cues. We evaluate the effectiveness of this enriched input across five modeling paradigms: 1) TF-IDF with a multi-layer perceptron (MLP), 2) frozen transformer encoders with pooling, 3) fine-tuned transformer encoders, 4) prompt-based zero- and few-shot generative LLMs, and 5) an instruction-tuned LLM, which forms the core of our proposed method. Our results show that the integration of LLM-generated augmentations with instruction tuning yields strong and interpretable performance, bridging content, discourse, and temporal signals without reliance on user profiles or external resources.