Emotional Language Weakens Math Skills in AI Models—But Neutral Rewrites Help

Researchers have discovered a surprising vulnerability in large language models: when mathematical problems are reframed with emotional language like frustration or urgency, these systems struggle to solve them accurately—despite the underlying numbers and relationships remaining unchanged.

The Problem#

Large language models are typically trained and tested on quantitative reasoning tasks presented in neutral, clear language. In practice, however, users ask questions wrapped in real emotion. The new work tests whether this emotional framing alone undermines a model's ability to reason correctly.

Using a carefully controlled framework that translates problems into emotionally charged versions while preserving all numerical content, researchers created Temper-5400, a benchmark of 5,400 emotion-neutral paired problems drawn from established reasoning datasets (GSM8K, MultiArith, and ARC-Challenge).

The Results#

Testing eighteen models ranging from 1 billion to frontier-scale parameters revealed two key findings:

• Emotional framing costs performance: Adding emotional language reduced accuracy by 2–10 percentage points, a meaningful drop even though nothing mathematical changed
• Neutralization recovers performance: Rewording emotionally charged problems back into neutral language recovered most of the lost accuracy, proving the degradation stems from emotional style rather than corrupted content

The effect was specific to emotion—non-emotional paraphrases that changed wording without emotional content caused no such accuracy loss.

Why It Matters#

The findings highlight a blind spot in how modern AI systems are evaluated and deployed. Real-world users often phrase questions with urgency, frustration, or enthusiasm, yet benchmarks rarely test this mismatch. The study also demonstrates that neutralization can serve as a lightweight fix at inference time, without retraining. More broadly, the benchmark framework offers a tool for studying how stylistic variations affect model robustness across other dimensions.