"exploring the ability of LLMs to understand emotional stimuli... automatic experiments on 45 tasks using various LLMs... span deterministic and generative applications ... show that LLMs have a grasp of emotional intelligence, and their performance can be improved with ... “EmotionPrompt” ... [which] requires only the addition of emotional stimuli to the initial prompts... to design effective emotional stimuli ... we take inspiration from three types of well-established psychological phenomena":
- self-monitoring: "the process by which individuals regulate and control their behavior in response to social situations and the reactions of others"
- social cognitive theory: "stresses that learning can be closely linked to watching others in social settings, personal experiences, and exposure to information... that individuals seek to develop a sense of agency for exerting a large degree of control over important events in their lives"
- cognitive emotion regulation: "people lacking emotion regulation skills are more likely to engage in compulsive behavior and use poor coping strategies... Techniques from this theory, such as reappraisal, can help".
Some key findings - EmotionPrompt:
- improves more in few-shot setting than zero-shot
- "enhanced capacity for generating ethically responsible responses"
- "enriched supporting evidence and superior linguistic articulation"
- "stimulates the creative faculties and overarching cognizance of LLMs"
- may elicit responses with "more definitive terms, such as “completely” and “will not”, while the narrative produced by the original prompt adopts a more tempered tone"
The paper goes on to examine why this works, and which emotional stimuli are most effective.