After analyzing common computer and board story games, Ing. Macháček used the best language models in the HaLLMark prototype. Great language models are advanced artificial intelligence systems that can process, understand and generate human text. They are based on deep learning techniques and trained on huge amounts of textual data such as books, articles and web pages. The HaLLMark prototype uses a suitable story environment and algorithms to create the story and game characters. These models facilitate loose interactions with NPCs, allowing them to generate immersive and coherent stories in real time.
HaLLMark is aimed at players who like to create their own stories and strategy enthusiasts. It allows players to create their own character - a king who rules a fantasy kingdom and solves the problems of his subjects. Players can freely interact with the characters in the game through text, and their decisions affect the rules of the game world and the metrics of the kingdom (the wealth and power of the kingdom, the popularity of the king, and the patience of the king's court). However, the player must stick to their royal role. If he makes a nonsensical decision or the conversation veers completely off topic, the patience metric of the king's court drops significantly. If either metric drops to zero, the player loses the game. It is therefore in the player's interest to write meaningfully and sensibly. In addition, the game characters are instructed not to stray from the theme in any way and to stick to the role of subjects in a fantasy world. If the player strays, the characters will try to bring the story back to the problems the player must solve.
"Players were happy with the prototype and the generated story and often praised the concept of the game. Plot-twists were added to the game based on the story generation research, which led to many funny moments during testing and improved the game significantly. The test results show previously unseen levels of consistency and interest in the generated story, which was achieved by using vector databases and prompt engineering techniques," says Jiří.
The results show that language models can effectively create consistent narratives, opening up new possibilities in game development. In the future, new models and advanced features could be added to the prototype to improve the player experience. The game could be translated into other languages and new game mechanics added to allow players to take active action, not just react to issues raised by characters.
Jiří has already presented his work and game HaLLMark at two conferences - at the Meaningful Play 2024 conference in Pittsburgh in October and at the AI Hotspot, conference in November, where he won the Young Talent Award for it.
More info about the diploma thesis
More about the game