LLM-Powered NPC Agents

Sammanfattning

Non-Playing Characters (NPCs) are fundamental to video games, taking on the role of each character not played by any human that inhabits the game world. Instead they’re actions and behaviours are controlled by the game’s computer system. However, a persistent challenge in NPC design is achieving believability, ensuring that NPCs meet players’ expectations in the roles they’re meant to portray. Traditional NPC behaviors, interactions, and dialogues have relied on scripted content, which, despite advancements, remains inherently rigid. The recent breakthroughs in the field of artificial intelligence (AI), particularly large language models (LLMs), presents an opportunity to overcome some limitations, especially in the realm of NPCs’ social interactions with the player. LLMs can dynamically generate contextually appropriate dialogue and behaviors, which may greatly enhance the NPCs believability as well as aligning with the principles of emergent narrative, an approach in Game Design that describes the occurrence of character-driven storytelling over embedded narratives which are the plots already put in place by the developers. While emergent narrative offers compelling possibilities, it also introduces challenges, such as distinguishing meaningful events from mundane or contradictory occurrences and effectively presenting dynamic stories in real-time. This study examines the potential of LLM-powered NPCs in addressing these challenges, particularly in enhancing believability and emergent storytelling. By integrating LLMs into NPC design, this research aims to evaluate and compare their impact on player experience and interactive storytelling, juxtaposed to scripted NPCs, providing insights into the future of AI-integrated video games. Eight participants with prior experience of interacting with traditional NPCs in role-playing games (RPGs) such as Skyrim, were recruited for this study. Each participant underwent an initial structured interview, followed by the case study which had them interacting with LLM-powered NPCs in the RPG Skyrim using the Mantella Mod, a modification for Skyrim which allows for LLM integration. Lastly, a final semi-structured interview was held after the case study. With the use of a Thematic Analysis, the data collected from the interviews and gameplay recordings showed that use of LLM-powered NPCs have various impacts on NPC design and were identified and categorized in four distinct themes; “Immersion & Believability,” “Narrative Emergence,” “Comparative Perception” and “Technical & Design Constraints”. The results showed that the participants experienced a greater sense of immersion by communicating with LLM-powered NPCs as it felt more natural and unrestricted. However, the findings also indicated that this technology is not without its current technical flaws, and by mitigating the problem of restricted dialogue trees of traditional NPC design, it introduces new problems which relates to the player’s real life communication skills or the lack thereof. Moreover, the freedom provided by the usage of LLMs may not be easily implemented in a game with embedded narratives as they are in direct contradiction with each other.