REVISE: A Framework for Paragraph-Level Misinformation Correction in Large Language Models
Shivangi Tripathi, Teancy Jennifer, Henry Griffith, Heena Rathore · 2025
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language generation but remain prone to hallucinations—producing factually incorrect yet plausible content. This paper presents a novel, structured framework for systematically identifying and revising misinformation in LLM-generated multi-sentence paragraphs. Unlike prior approaches that focus on isolated sentence-level corrections, our method ensures paragraph-level coherence through sequential sentence decomposition, fact-checking question generation, evidence retrieval via web APIs, and context-aware revision. Leveraging a semantic agreement gate and factual consistency scoring, our approach minimizes edits while preserving the original style and meaning. Empirical results demonstrate improved factual accuracy and coherence, validating the effectiveness of our framework for misinformation mitigation in generative AI systems.