We benchmarked five large language models (GPT‑5.1, Gemini 2.0 Pro, Llama 3, Qwen, BioGPT) across five heterogeneous knowledge graphs: DrugBank, NeuroSynth, Cognitive Atlas, Brain Map, and GDPKG. A multitier query set was created to assess reasoning depth and factual accuracy under graph‑grounded (KG‑aware) and graph‑agnostic conditions. Inter‑annotator agreement was strong, with Cohen's κ = 0.781 (range 0.753‑0.807) and Fleiss' κ = 0.77. Results show that general‑purpose models, especially GPT‑5.1 and Gemini 2.0 Pro, generally outperform domain‑specific models on multi‑hop and high‑level reasoning tasks, although performance degrades as query complexity rises. Graph grounding consistently reduces hallucination rates for all models. Contrary to our initial hypothesis, domain‑specific models do not consistently beat general models on specialized tasks without grounding. Based on these findings we offer practical guidance for selecting LLM‑KG integration strategies given task complexity and computational constraints, and discuss hybrid neurosymbolic architectures for personalized medicine, cognitive disorder diagnosis, and explainable AI.
Review: The work delivers a rigorous evaluation pipeline and demonstrates that grounding LLMs in structured graphs is essential for mitigating hallucinations and enhancing multi‑hop reasoning reliability, offering actionable insights for real‑world deployments.