Here is a paper summary:
The paper explores whether large language models (LLMs), despite their impressive capabilities, can genuinely comprehend and be improved by psychological emotional stimuli like humans. Emotional intelligence significantly impacts human behavior and problem-solving.
To study this, the authors propose EmotionPrompt, which combines original prompts with 11 emotional phrases inspired by psychology theories like self-monitoring, social cognitive theory, and cognitive emotion regulation.
The authors conduct extensive experiments on 45 diverse tasks using 6 LLMs, including deterministic tasks like Instruction Induction and BIG-Bench which allow automatic evaluation, and generative tasks like poem writing which require human assessment.
For the deterministic tasks, EmotionPrompt improves performance by 8% on Instruction Induction and 115% on BIG-Bench. The generative task human study with 106 people also demonstrates EmotionPrompt boosts performance, truthfulness and responsibility of LLMs by 10.9% on average.
Analyses reveal emotional stimuli contribute more to model gradients, enhancing original prompt representations. Performance also varies based on emotional stimulus type, model scale, and temperature.
LLMs comprehend and can be augmented by emotional intelligence. This highlights interdisciplinary opportunities between AI and psychology to improve human-LLM interaction.
