A Novel Metric for Measuring the Robustness of Large Language Models in Non-adversarial Scenarios [2408.01963]