EmotionPrompt: Leveraging Psychology for Large Language Models Enhancement via Emotional Stimulus

Large language models (LLMs) have achieved significant performance in many fields such as reasoning, language understanding, and math problem-solving, and are regarded as a crucial step to artificial general intelligence (AGI). However, the sensitivity of LLMs to prompts remains a major bottleneck for their daily adoption. In this paper, we take inspiration from psychology and propose EmotionPrompt to explore emotional intelligence to enhance the performance of LLMs. EmotionPrompt operates on a remarkably straightforward principle: the incorporation of emotional stimulus into prompts. Experimental results demonstrate that our EmotionPrompt, using the same single prompt templates, significantly outperforms original zero-shot prompt and Zero-shot-CoT on 8 tasks with diverse models: ChatGPT, Vicuna-13b, Bloom, and T5. Further, EmotionPrompt was observed to improve both truthfulness and informativeness. We believe that EmotionPrompt heralds a novel avenue for exploring interdisciplinary knowledge for humans-LLMs interaction.