Mitigating LLM Hallucinations in Mobile AR Tutoring Environments using an SD-DGR Algorithm

Abstract

Mobile Augmented Reality (MAR) is used to deliver immersive and interactive educational contents to learners. It is the second stage in the intelligent tutoring process with pedagogical diagnosis being the first stage. Generative Large Language Models (LLMs) such as OpenAI's GPT-4o are used as the standard in this paper. The challenge faced in mobile pedagogical automation is delivering accurate, hallucination-free instructional content under severe network constraints. The previous RAG algorithms which work well on high-bandwidth networks cannot be efficient under volatile latency and vice-versa. This paper proposes a state-driven deterministic grounding and retrieval algorithm which produces a scalable local data stream which has a number of fine layers of pedagogical validation. This modified form of the retrieval-augmented generation algorithm is used to generate an adaptive pedagogical companion which is scalable and uses the state-driven and format-enforcement stages which are responsible for the syntactic and semantic terminations that are actually detached as practically all the data detached during the local retrieval phases at the client side is supplemented towards the prompt at cloud inference stage. Therefore, clearly the retrieval phase which is modified to produce a local data stream which is scalable. This modified algorithm works well at both lower and higher hardware specifications. Quantitative evaluations were done using three heterogeneous Android devices under high network jitter conditions and the mean normalized processing latencies, memory footprints, and syntax accuracies at various trials was noted and compared with the previous RAG algorithms.