Does Emotion in Prompts Really Matter? Exploring the Impact of Emotional Framing on Large Language Models

This article explores the impact of emotional prompts on the performance of Large Language Models (LLMs). Key findings include: static emotional prefixes have limited overall effects, but show more noticeable variations in social-related tasks; the adaptive emotion selection framework EmotionRL is more reliable than fixed emotional prompts. Below, we will conduct a detailed analysis from dimensions such as background, methodology, and findings.

Emotion is ubiquitous in human communication and affects the listener's response. Do LLMs have similar reactions to emotional prompts? This question is of great significance for optimizing human-computer interaction: if emotional framing is effective, 'emotional prompt engineering' may become a new optimization dimension; otherwise, other directions need to be explored.

The research team tested the impact of emotional framing in five benchmark domains: mathematical reasoning, medical Q&A, reading comprehension, commonsense reasoning, and social reasoning. A first-person emotional framing was adopted, with prompts presented in ways like 'I feel...' and 'This is very... to me'.

Experiments show that static emotional prefixes (e.g., 'This is really important to me') have a relatively small impact on model accuracy and are not a universally reliable intervention. However, in social-related tasks, the impact of emotional framing is more variable—because in social reasoning scenarios, there is a reasonable interaction between emotional context and interpersonal reasoning.

Stronger emotional wording only brings moderate additional changes (diminishing marginal effect); the effects of emotional prefixes generated by humans and models are similar. Based on this, the study proposes the EmotionRL adaptive framework: dynamically select the most appropriate emotional frame for each query, drawing on reinforcement learning ideas to maximize expected performance.

Theoretical contribution: Emotional tone is a weak and input-dependent signal, neither a dominant factor nor irrelevant noise. Practical recommendations: 1. Do not rely on static emotional prompts; 2. Social reasoning applications can focus on emotional framing; 3. Prioritize methods such as example engineering and reasoning chain design; 4. If using emotional prompts, adaptive selection is more reliable.

Limitations: Only first-person emotional framing was focused on, and other forms were not involved; experiments used general benchmarks, and effects in specific scenarios need to be verified. Future directions: Explore the impact of multimodal emotions, emotional adaptation in long-term interactions, and differences in emotional expression across different cultural backgrounds.