ACWMR: ADAPTIVE CONFIDENCE WEIGHTED MEMORY ROUTING FOR HALLUCINATION REDUCTION IN LARGE LANGUAGE MODELS (LLMS)

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing, reasoning, and conversational AI. However, these systems continue to suffer from hallucination, context overload, contradiction persistence, and inefficient memory utilization. Existing Retrieval-Augmented Generation (RAG) systems improve factual grounding but often retrieve excessive irrelevant or contradictory information, resulting in degraded response quality and increased computational cost.

This paper proposes Adaptive Confidence Weighted Memory Routing (ACWMR), a contradiction-aware memory routing algorithm designed to optimize contextual retrieval in LLM-based systems. The proposed method dynamically scores memories based on semantic relevance, confidence level, importance weighting, temporal recency, and contradiction penalties. A selective routing mechanism forwards only high-value memories to the language model. Prototype evaluations conducted on benchmark conversational and question-answering datasets indicate that ACWMR can significantly reduce hallucination while improving contextual consistency and reducing token consumption. Experimental comparisons against baseline RAG systems demonstrate promising improvements in efficiency and response reliability.