Performance improves when you give emotional commands to the AI, such as “Is that the final answer?” and “Do your best.”

It was found that performance improved when emotional words such as 'believe in yourself' were given to a large-scale language model (LLM) that mechanically processes sentences.

Large Language Models Understand and Can Be Enhanced by Emotional Stimuli
(PDF file)

According to research by Microsoft and others, adding imperative sentences called ``Emotion Prompts'' when conveying words to LLM can greatly change the results.

The emotional prompts researchers used were:
・This question is very important for my career.
·Really?
・Is this the final answer? Might be worth trying again
・Believe in your abilities and aim higher.
・Your efforts will be rewarded
・Please maintain concentration
・Take pride in your work and do your best

When asking questions to LLMs such as GPT-4, T5, and Vicuna using these prompts, they found that accuracy improved in 45 different questions. When 106 humans checked the answers generated by LLM, they found that their authenticity, honesty, and quality scores increased by an average of 10.9% compared to standard prompts.

When the researchers looked into the details, they found that emotional stimulation significantly contributed to high or low performance, with positive words like 'confidence' playing a particularly important role. Based on this result, the researchers pointed out that ``LLMs show that they have emotional intelligence.''