Researchers develop statistical framework showing misinformation in prompts significantly impairs LLM performance on arithmetic tasks

New regression-based framework helps explain how specific prompt components impact LLM performance by fitting models that relate prompt portions to evaluation results

Testing on Mistral-7B and GPT-OSS-20B models reveals regression models explain 72-77% of variation in arithmetic problem-solving performance

Study finds incorrect example query-answer pairs substantially hurt both models' ability to solve arithmetic queries, while positive examples show minimal variability in impact

Approach extends existing explainable AI (XAI) methods to provide transparency into how LLMs utilize prompts for task completion