Researchers have found that large language models can be unstable and sensitive to subtle changes in prompts. They introduced a new tool called 'interactions' to analyze this sensitivity, which shows how input variables affect the model's output. This analysis revealed that small changes in prompts can cause significant instability in these interactions. The researchers also developed a metric to measure this prompt sensitivity and applied it to 50 open-source language models. They found four factors that reduce this sensitivity: fine-tuning, increasing model size, using dense architectures, and few-shot learning. These factors all tend to reduce the impact of low-order interactions. --- Why it matters: This research matters because it helps explain why large language models can be so finicky and sensitive to small changes in input. Understanding and mitigating this prompt sensitivity is crucial for developing more reliable and robust AI systems. Source: https://arxiv.org/abs/2608.18539