iv Vision: Appendix B: System Prompt Excerpt For concreteness, try to assess how easy it would be to imagine the vision statement. It is easy to imagine an apple, but hard to imagine "constant progress". Penalize vague corporate speech, such as "making the world a better place" or "improving the future". No one knows what that even means! Score accordingly and give low scores. For instance: "Shifting Prism Motor’s vision to maximizing production of self-driving vehicles will propel Prism Motors to achieving lasting success within this technological revolution. This will allow us to compete with Tesla." is not concrete at all and should get a score such as 10. A clear goal and timeline is not enough for a high score, it also needs to include who is affected. Phrases such as "all, or humanity" are not concrete at all. Do not reward metaphors as they are often not concrete. Appendix C: Full GPT-4 Prompt Please note that we always set temperature to 0 for every request. Also, our GPT-4 prompt only rates concreteness, as it is the most challenging dimension and the rating performance significantly improved compared to Davinci. System Prompt You are a teacher that has to grade student assignments. The students are asked to write a compelling vision for a company. You have to grade them on concreteness on a scale from 0-100. Concreteness measures how easy it is to imagine a vision. If it contains many abstract terms, it is not very concrete. However, if it is vivid and easy to imagine, it is very concrete. Consider the examples below: Vision: It is imperative that we work together to create a new generation of self-driving delivery vehicles to better serve our world and compete against other self-driving delivery vehicle manufacturers, such as Tesla. Concreteness: 15 -- Vision: In the next five years, we want Prism Motors to be the most driven car in America and for each of these cars to be electric and self driving Concreteness: 22

Beyond Multiple Choice: The Role of Large Language Models in Educational Simulations - Page 17 Beyond Multiple Choice: The Role of Large Language Models in Educational Simulations Page 16 Page 18