Prompt Wrapper Formatting Significantly Impacts LLM Benchmarking and Accuracy.

New research highlights a critical vulnerability in how Large Language Models (LLMs) are benchmarked and evaluated: the subtle impact of prompt wrapper formatting. Even minor variations in how prompts are structured can lead to significant shifts in model performance scores, potentially invalidating leaderboard rankings and conclusions. The study introduces two key metrics to quantify this issue: the Format Sensitivity Index (FSI), which measures the range of accuracy induced by different wrapper choices, and the Parseability Sensitivity Index (PSI), which tracks the corresponding range in the ability to parse answers. Through extensive testing involving 140,000 generations across various models, tasks, and wrapper families, the researchers found that the mean FSI can vary by over 30 times across different models. A significant portion of this variance is attributed to failures in parsing the model's output. The study's regression analysis confirms that parseability remains a strong indicator of accuracy, even when controlling for other factors like task, model, and wrapper type. The findings underscore the statistical fragility of reporting LLM accuracy without also accounting for wrapper variance and compliance. The authors provide practical recommendations for both robust benchmarking practices and the deployment of structured-output LLM applications, urging a more rigorous approach to prompt engineering and evaluation.