Reducing Hallucinations in Large Language Models Through Knowledge Graph Integration and Confidence-Aware Reasoning

Abstract

Large language models (LLMs) generate fluent natural-language responses but may produce unsupported entities, fabricated relationships, inaccurate numerical statements, and misleading citations. This paper proposes a knowledge-graph-integrated, confidence-aware reasoning framework (KG-CAR) that combines entity linking, relation-aware graph retrieval, evidence-constrained answer generation, claim-level verification, and calibrated selective prediction. A structured knowledge graph supplies provenance-bearing triples, while a confidence estimator combines retrieval relevance, path consistency, entailment, source reliability, and model uncertainty. Low-confidence claims trigger additional retrieval, qualified responses, or abstention rather than ungrounded completion. The paper formulates a reproducible experimental protocol for evaluating factual precision, hallucination rate, evidence coverage, expected calibration error, abstention behavior, and latency. To demonstrate reporting and interpretation, it includes a clearly labeled illustrative numerical scenario rather than claiming completed experiments. The framework targets trustworthy domain question answering and identifies practical limitations involving graph incompleteness, contradictory sources, extraction errors, and computational overhead.