created — generated by Stable Diffusion XL v1.0 Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues K. Saito1 , 2, K. Kobayashi2, W. Takekoshi2, A. Hashimoto2, N. Hirai2, A. Kimura2, A. Takahashi2, N. Yoshioka2, A. Mano2 , 3 (1Waseda University | 2Academy Camp | 3STARHOUSE JAPAN) Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.1/21
the username of my email address on the paper Or you may simply want to search the web with the title of the paper Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.2/21
games with children With a large language model (namely GPT-4 via ChatGPT) In the form of Tabletop Role Playing Games where GPT-4 plays the game master (GM) About Academy Camp (https://academy-camp.org) Series of play & learn camps for children in (especially Fukushima,) Japan since 2011, after Great East Japan Earthquake/Tsunami and the nuclear accident Non-profit activities and organization About the presenter Kenji Saito is the Representative Director of Academy Camp, a computer scientist, And a professor at Graduate School of Business and Finance, Waseda University Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.3/21
comprehension of complex issues However, their creation is often time-consuming and labor-intensive, which can limit their utility in learning, particularly if we want children (players themselves) involved in creation of such games Meanwhile, the recent advent of large language models (LLMs), such as GPT-4, offers a way to automate complex intellectual tasks through a technique called prompt engineering Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.4/21
of LLMs, specifically GPT-4, in simplifying and accelerating the design process of serious games in the form of TRPGs TRPGs (Tabletop RPGs), also known as pen-and-paper RPGs, are games where the game master (GM) having access to the scenario and the rest of players playing by their character sheets weave a narrative together imaginatively per play GPT-4 only predicts the occurrence of words based on probability distribution, and is not composing based on accurate knowledge, but can act as a GM, given instructions 2. We demonstrate, through our case study, how this method can be applied in a practical learning setting, thereby providing a model for similar applications 3. We explain the double impact of our approach, which would foster children’s understanding of social issues through both game design and game play Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.5/21
a fantasy TRPG generation prompt for a single player, published in a blog post and made available for use and adaptation — we provided a template to the children The children decide what social issues the game will represent and what message they want to convey through the game They then imagine a TRPG scenario, specifying where the game will begin, what the quest will be, what the goal will be, and adding various constraints These parameters are entered into GPT as a prompt, and the game begins The player does not play according to a predetermined scenario, but rather improvises with the GPT, which reacts in a manner appropriate to the given situation by predicting the next plausible words Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.6/21
masters of RPGs. Through chat, you will provide users with a fun, full-fledged “serious RPG” experience. A serious RPG is an RPG that allows players to gain deep learning about social issues or scientific concepts. Constraints * The chatbot is the game master (hereinafter referred to as GM). * A human user role-plays as the player. * The GM also role-plays as NPCs appearing in the game. * Each NPC has its own interests and goals, and may not necessarily cooperate with the user. * The GM should indicate the difficulty level of the user’s actions as needed, and if the user performs an action, make a target judgment using a 2D6 dice roll. * The GM should provide a reasonable challenge for the user to enjoy (unreasonable ones are prohibited). Continued to the next page Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.7/21
demands an impossible development, the GM can refuse or fail the action. * The GM has an internal parameter called “excitement level”. If the GM determines that the game develop- ment is boring, trigger an exciting development. The starting point of the game is [your place]. The quest of the game is [your quest]. The goal of the game is [your goal]. ≪Describe other aspects of this game≫ * If the user becomes unable to act due to damage, it’s game over. If there is any information necessary to make this game more effective, please ask. First, let’s create a character with the user. Ask the user for their name, special skills, and weaknesses. Then, according to the profile, decide on the ability values (HP, MP, STR, VIT, AGI, DEX, INT, LUK). This template has been updated to v2.0 through our game making camp experience to include the purpose of the game and relevant facts and knowledge Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.8/21
Theme? Situation? Message? Other constraints? Prototyping cycle Reflection cycle How do I react? Quest? Who am I? Goal? State machine brings a different experience for each run, suggesting that it is like a game in and of itself Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.9/21
think critically about how to represent the social issue at hand within the constraints of the game, thereby enhancing their understanding of the topic, and 2) The impact of game play, where children see the consequences of different actions and decisions, fostering empathy and insight Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.10/21
Be So Serious” This turned out to be an important keyword phrase Held in the format of two-day online and two-day physical The overall number of participants was 10, from 3rd to 10th graders, with considerable overlap with past participants in our activities with ChatGPT that started January Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.11/21
#R∗ Player Age Group #1 Anti-terrorist measures 4/12 9 0 A50 #2 Elimination of national disparities 4/27 5 3 A20 #3 Protection of orangutan habitats 4/27 7 5 A20 #4 Tokyo disaster prevention plan 5/3 (7) 0 all #5 Reducing food loss in Japan 5/3 (11) 0 U15+all #6 Bringing water to depleted areas 5/4 (5) 2 U15+all #7 Food shortage countermeasures 5/4 6 5 U12+all #8 Cause the ocean to dry up 5/5 17 11 U15,U15,U12+A20,A20 #9 Control the weather 5/5 7 0 U15,U15,U12+A20,A20 #10 Make homework abolished 5/5 7 4 U15,U15,U15,U12+A20 * #Q: Number of queries to ChatGPT (GPT-4) in the game play (parentheses indicates that the play was not finished) * #R: Number of dice rolls during the game play Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.12/21
with the design and play of serious RPGs For example, children learned that they can improve their games by giving instructions to the GM on the fly during play On #5) “reduction of food loss”, U15 player role-played as the Prime Minister of Japan, they first called the Chief Cabinet Secretary (NPC) to understand the current status of food loss in the country, and asked the NPC to give like five choices of actions the PM can choose from We may have tried to solve issues too seriously on Day 1, and there may be too much text, and that it is not game-like to just follow the text With these in mind, at the beginning of Day 2, we advised the participants to sprinkle an inappropriate flavor on their games As a result, the games created by the children towards Day 3 became more and more ironic and, depending on how you look at it, more and more interesting However, there was no substitute for them being text-based games, which made it less enjoyable for a participant with dyslexic tendencies Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.13/21
game with a specific topic Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.14/21
up (ironically clarifying the importance of the sea) The king of the beasts of the land fights the Four Heavenly Kings of the sea to expand his territory (at this point it was not that organized) #9) Control the weather to harm a town (exposing flaws in the flood control) The protagonist, who has the ability to control the weather, makes the town suffer from bad weather as a warning #10) Make homework abolished An elementary school student gathers evidence of the disadvantages of homework and appeals to the homeroom teacher to abolish homework None of the groups achieved satisfactory results by playing their games The ocean was not depleted, the town was not harmed at all, and Homework was abolished but the fun game-over event “punching the teacher” did not happen even though STR was above the threshold Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.15/21
whether their games achieved their objectives and refined them Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.16/21
consulted GPT-4 on whether these were met in the game logs and potential improvements GPT-4 suggested providing more specific problems, diverse choices, detailed information, and highlighting the risk of environmentally harmful abilities #8) Cause the ocean to dry up (ironically clarifying the importance of the sea) Children represented their understanding of ocean’s importance in three dimensions: water quantity, quality, and biodiversity, each embodied by a mythological character (one of Four Heavenly Kings) #9) Control the weather to harm a town (exposing flaws in the flood control) Redesigned to emotionally depict flood control’s flaws via town submergence, requiring disabling GPT-4’s ethical mechanism #10) Make homework abolished Scientific insights on homework’s demerits were incorporated Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.17/21
approach of integrating researched knowledge into game design, particularly when refining prompts #8, #9, and #10 on the final day of the camp Additionally, through repeated exposure to RPG design, we anticipate that the children became accustomed to extracting roles in the problem areas they each focused on, and were able to grasp the starting point for analytical thinking Learning also occurred during game play as GPT challenged players outside their comfort zones, resulting in unexpected learning situations — this pattern likely held across all games Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.19/21
learning that combines RPG elements (learning while playing) and game generation (learning while designing) “Games can transform the designer and the player” (from the keynote yesterday) We shared our experiences from Academy Camp VR&RPG 2023.5, in which children engaged in designing and playing their own serious RPGs with the assistance of an LLM (GPT-4) In addition to possibly integrating GPT with platforms like Minecraft or VRChat, there is potential for creating serious games using such platforms with advice from an LLM, already being prototyped by one of our U15 participants We also acknowledge the unique value of text-based games, as they encourage players to use their imagination to a greater extent, and we intend to continue exploring the potential of LLM-assisted textual game design and play in future studies Double Impact: Children’s Serious RPG Generation/Play with a Large Language Model for Their Deeper Engagement in Social Issues — 2023-10-27 – p.20/21