A Self-Healing Architecture for Mitigating Bibliographic Hallucinations in LLM-Generated Academic Texts

Abstract

Abstract Large Language Models (LLMs) have demonstrated remarkable capabilities in academic writing assistance, yet they frequently generate bibliographic hallucinations (i.e., plausible but non-existent citations) that undermine scholarly integrity. Hallucination rates in LLM-generated references can exceed 25%, with some models producing over 90% fabricated citations, posing risks to academic credibility, as evidenced by legal sanctions against professionals who submitted AI-generated documents containing false references. This paper presents a novel self-healing architecture that detects and mitigates bibliographic hallucinations in LLM-generated academic texts. The system implements a multi-layered verification framework, combining citation validation, cross-referencing with authoritative databases, and automated correction, to autonomously identify inconsistencies in bibliographic data (author names, titles, venues, dates) and either correct them using verified sources or flag them for human review. Experimental evaluation shows that the self-healing system significantly reduces hallucination rates while maintaining fluency and coherence. Our baseline evaluation with gpt-4o-mini showed that approximately 40% of generated DOIs were invalid. We tested the approach on more than 1,000 citations across 4 LLM engines, achieving an average identification accuracy of 0.96 and an F1 score of 0.93. Furthermore, optimizing the review (refinement) phase using a sentence-based approach yields significant computational efficiency, saving nearly 67% of tokens and latency compared to full-text refinement. Self-healing LLM thus provides a scalable, transparent, and robust solution for enhancing the trustworthiness of machine-generated scholarly content, contributing to the development of more reliable AI-assisted academic writing tools.