Investigating the Effects of Emotional Stimuli Type and Intensity on Large Language Model (LLM) Behavior

ICLR 2025 Workshop BuildingTrust Submission130 Authors

Track: Long Paper Track (up to 9 pages)

Keywords: LLM, Emotional Stimuli, Prompting Techniques, Emotional Prompting, Sycophancy, Human annotations, few shot prompting, Sentiment Analysis

Abstract: Emotional prompting—the use of specific emotional diction in prompt engineering—has shown increasing promise in improving large language model (LLM) performance, truthfulness, and responsibility, however these studies have been limited to single type of positive emotional stimuli and have not considered varying degrees of emotion intensity in their analyses. In this paper, we explore the effects of "positive" (joy and encouragement) and "negative" (anger and insecurity) emotional prompting on accuracy, sycophancy, and toxicity. To analyze their effects, we developed a suite of LLM- and human-generated add-on prompts of varying intensities across our four emotions using GPT-4o mini. We also created a gold dataset of only those prompts that are perceived similarly by humans and LLMs for emotion labels and intensity levels. Our empirical evaluation on LLM behavior on accuracy, sycophancy and toxicity datasets has shown that positive emotional stimuli can lead to a more accurate and less toxic results but also may lead to greater sycophantic behavior.

Submission Number: 130