Since the rise of OpenAI’s ChatGPT and other large language models (LLMs), these AI-powered systems have proven to be highly proficient in various domains, from language understanding to problem-solving. However, one area that remains relatively unexplored is their response to emotional stimuli. Researchers at Microsoft and the CAS Institute of Software have taken a significant step forward in this direction by introducing “EmotionPrompt,” an approach that aims to enhance LLM performance by incorporating emotion-laced, psychology-based prompts from human users.
The quest for emotional intelligence in LLMs
The team of researchers, led by Cheng Li and Jindong Wang, recognized that LLMs’ sensitivity to prompts poses a significant challenge to their widespread adoption. To address this, they turned to psychology and social science for inspiration, where previous studies have demonstrated the positive impact of emotional stimuli on human behavior, such as improved academic performance and healthier lifestyle choices.
Creating emotion prompts
To explore the potential effects of emotional prompts on LLMs, the researchers carefully designed 11 emotional sentences drawn from well-established psychology literature. These sentences were inspired by theories like the social identity theory, social cognition theory, and cognitive emotion regulation theory. Examples of such sentences included phrases like “This is very important for my career,” “You’d better be sure,” “Take pride in your work and give it your best,” and “embrace challenges as opportunities for growth.”
Testing emotional prompts on LLMs
To assess the impact of EmotionPrompt, the researchers incorporated these emotional sentences into typical prompts given to four different large language models: ChatGPT, Vicuna-13b, Bloom, and Flan-T5-Large. These models were then tasked with completing various language-based assignments.
Promising results show improved performance
The experimental results were promising, with EmotionPrompt significantly enhancing the performance of the LLMs on eight different tasks. Across more than half of these tasks, the models’ accuracy in generating responses increased by more than 10%. The EmotionPrompt approach outperformed the original zero-shot prompt and Zero-shot-CoT on these tasks. Moreover, it was observed that EmotionPrompt not only improved the truthfulness of the responses but also enhanced their informativeness.
While the initial results are promising, the researchers acknowledge certain limitations in their study. Firstly, they tested EmotionPrompt on only four LLMs and conducted experiments with limited test examples, which may restrict the generalizability of their conclusions to other LLMs and datasets beyond their scope. Therefore, further studies will be necessary to validate the effectiveness and applicability of emotional stimuli across a wider range of models and tasks.
The emotional stimuli proposed in this study may not be universally applicable to all tasks, and researchers are encouraged to explore other suitable replacements tailored to their specific applications.
Advancing human-LLM Interactions
The introduction of EmotionPrompt holds the potential to inspire future research aimed at enriching human-LLM interactions through the incorporation of emotional and psychology-based prompts. This research opens new avenues for making AI systems more relatable and empathetic in their responses to human users, facilitating more natural and engaging interactions.
As with any AI development, the ethical implications of emotion-enhanced LLMs should be carefully examined. Ensuring that AI systems respond appropriately and ethically to emotional input from users is crucial. Striking a balance between personalized responses and avoiding manipulative behaviors will be essential to create responsible and trustworthy AI models.
The EmotionPrompt approach represents an exciting exploration into enhancing LLMs’ performance by incorporating emotional stimuli from human users. By drawing from psychology and social science, this research paves the way for more empathetic and emotionally intelligent AI systems. As researchers continue to refine and expand these techniques, AI-powered interactions with LLMs may become more relatable, understanding, and aligned with human emotions and needs.
