Researchers have found that large language models (LLMs) are highly sensitive to minor changes in prompts. They analyzed a dataset of 132,000 prompt variants and discovered a 'Scaling Law of Prompt Performance Stability', which shows that higher task performance is associated with lower variance and greater robustness across prompt perturbations. The team identified two key factors contributing to this stability: domain-specific terminology and explicit action directives. Based on these insights, they developed an automated prompt-refining agent that reduces performance variance by 40.7% in code generation tasks while preserving or improving mean performance. --- Why it matters: This study matters to AI engineers because it provides a statistically grounded framework for achieving robust prompt engineering, which is crucial for developing reliable and efficient language models. The findings can help researchers and developers improve the stability and reliability of their models by understanding how minor changes in prompts affect performance. Source: https://arxiv.org/abs/2608.20349