Table 5.

Standard deviation (σ) of diagnostic accuracy across temperature settings (0.0–1.0) for LLMs

Dataset	Model	Vanilla σ	Cypher RAG σ	Vector RAG σ	RAG σ Reduction (%)
Publication set selective test	GPT-3.5-turbo	2.53	1.47	1.60	39.24%
	GPT-4-turbo	2.81	1.17	1.03	60.83%
	GPT-4o	2.35	0.75	0.75	67.90%
Publication set non-selective test	GPT-3.5-turbo	2.64	1.26	2.50	28.62%
	GPT-4-turbo	4.55	1.41	0.98	73.63%
	GPT-4o	2.26	0.41	0.75	74.29%
GMDB set selective test	GPT-3.5-turbo	2.53	1.86	1.47	34.11%
	GPT-4-turbo	2.35	0.89	0.75	64.88%
	GPT-4o	2.28	1.37	1.03	47.40%
GMDB set non-selective test	GPT-3.5-turbo	2.64	1.26	1.17	53.89%
	GPT-4-turbo	2.26	1.03	0.75	60.47%
	GPT-4o	2.07	1.41	0.98	41.97%

Publication set results show Cypher-RAG achieves the lowest variability (σ = 0.41 for GPT-4o in non-selective), while Vector-RAG demonstrates consistent stability (σ ≤ 1.60 in selective). GMDB set exhibits similar patterns. All RAG LLMs show significantly reduced variance compared to Vanilla LLMs (average RAG σ Reduction: 53.94%), suggesting that they perform with less effect of temperature parameter.