Application and Analysis of Large Language Models in Game NPCs

Abstract

In recent years, the breakthroughs of large language models (LLMs) in capabilities such as natural language understanding and multi-turn dialogue generation have not only promoted the technical implementation in fields like natural language processing (NLP) and computer vision (CV), but also provided core support for the integrated exploration of "LLMs + intelligent agents". However, most tests rely on simple and static virtual scenarios, which are difficult to simulate the complex environments with "multi-variable interaction and dynamic target changes" in real situations. Games, however, can exactly fill this gap: they can not only construct interactive scenarios requiring advanced cognitive abilities (e.g., task decision-making in open worlds, emotional dialogue of non-player characters (NPCs)), but also achieve low-cost data acquisition and controllable variables through procedural generation. Therefore, games have become an ideal carrier for evaluating the comprehensive capabilities of intelligent agents. With the steady growth of the game market share in recent years, various game manufacturers have introduced LLMs into in-game NPCs and achieved positive results. Research on the application of LLMs in game NPCs is of great significance for the intelligent advancement of NPCs. This review mainly summarizes and analyzes the model frameworks of LLM applications in NPCs, which can provide assistance for the construction of more reasonable and in-depth frameworks. It is conducive to the cross-optimization of models and provides reference significance for model development. In the future, the research applied to game NPCs is also expected to be extended to other fields.