Paragraph-Level Hallucination Detection and Correction for Trustworthy Large Language Models in Networked Systems

Large Language Models (LLMs) are increasingly integrated into communication-centric and networked AI systems, supporting applications such as edge/cloud services, mobile assistants, and distributed collaborative platforms. However, their tendency to generate hallucinations (factually incorrect yet linguistically plausible content) poses significant risks to trust, reliability, and security in these environments. We propose SRF, a Sequential Revision Framework for paragraph-level hallucination detection and correction in pervasive and networked AI workflows. SRF decomposes text into sentences, generates fact-checking queries, retrieves supporting evidence via web-scale APIs, and applies a semantic agreement mechanism to ensure globally coherent revisions. To evaluate factual correctness, we compare SRF against two baselines—(i) direct LLM correction (without external evidence) and (ii) sentence-level RARR—on the Natural Questions and SQuAD datasets using GPT-3.5, Mistral, and Gemini models. SRF yields up to a 15% improvement in factual accuracy, with minimal disruption to writing style and paragraph coherence. By enforcing fact-consistent and trustworthy LLM outputs, SRF strengthens the integration of AI/ML into communication networks, mobile computing, and distributed cloud/edge services, reducing vulnerabilities to misinformation propagation and adversarial manipulation.