Emotional Framing Degrades LLM Quantitative Reasoning by Up to 10%

A new study from arXiv (2604.07801) investigates how emotional language affects large language models' performance on quantitative reasoning tasks. The researchers developed a controlled emotion translation framework that rewrites neutral math problems into emotional variants while preserving all numerical content and relationships. Using this framework, they constructed Temper-5400, a dataset of 5,400 semantically verified emotion-neutral pairs derived from GSM8K, MultiArith, and ARC-Challenge. They evaluated eighteen models ranging from 1 billion to frontier scale. The study found that emotional framing reduces accuracy by 2-10 percentage points even when all numerical content is preserved. Furthermore, neutralizing the emotional variants recovered most of the lost performance, indicating that the degradation is directly tied to emotional framing rather than changes in content. This research highlights a potential vulnerability in LLMs when applied to real-world queries that often carry emotional tone, such as frustration, urgency, or enthusiasm. The findings have implications for the deployment of AI in fields like education, customer service, and decision support, where emotional context is common. The study was announced as a replace-cross update on arXiv.