Beyond Hallucinations: A Composite Score for Measuring Reliability in Open-Source Large Language Models

TL;DR

Open-source LLMs often exhibit overconfidence, brittleness to distribution shifts, and uncertain outputs, complicating deployment in critical domains. This work introduces the Composite Reliability Score (CRS), a unified metric that combines Calibration, Robustness, and Uncertainty Quantification into a single interpretable score, with $CRS = \alpha C + \beta R + \gamma U$ and pillar definitions $C = \max(0, 1 - \frac{\text{ECE}_{\text{model}}}{\text{ECE}_{\text{max}}})$, $R = 1 - \frac{\text{Avg. Accuracy Drop}}{\text{Avg. Acc}_{\text{clean}}}$, and $U = \frac{\text{AUROC} - 0.5}{0.5}$. The authors evaluate ten open-source LLMs across five QA datasets, finding that CRS yields stable rankings and reveals reliability weaknesses masked by accuracy alone. The results show that the most dependable systems balance accuracy, robustness, and calibrated uncertainty, with Mistral-8×22B often leading in overall reliability and single-dimension metrics being insufficient for deployment decisions.

Abstract

Large Language Models (LLMs) like LLaMA, Mistral, and Gemma are increasingly used in decision-critical domains such as healthcare, law, and finance, yet their reliability remains uncertain. They often make overconfident errors, degrade under input shifts, and lack clear uncertainty estimates. Existing evaluations are fragmented, addressing only isolated aspects. We introduce the Composite Reliability Score (CRS), a unified framework that integrates calibration, robustness, and uncertainty quantification into a single interpretable metric. Through experiments on ten leading open-source LLMs across five QA datasets, we assess performance under baselines, perturbations, and calibration methods. CRS delivers stable model rankings, uncovers hidden failure modes missed by single metrics, and highlights that the most dependable systems balance accuracy, robustness, and calibrated uncertainty.