A new study reveals that large language models are highly sensitive to small changes in prompt wording, leading to significant shifts in performance quality. This research moves beyond simple template approaches to analyze the precise impact of lexical variations on model outputs.