Design and Application Research of Dynamic NPC System Based on Local Large Language Models

Abstract

With the Development of the video game industry, the problem of traditional Non-Player Characters (NPCs) became rigid because they rely on pre-written scripts. Although the Large Language Models (LLMs) offer a new way to generate dynamic conversations, most useful ways nowadays from using cloud-base solutions suffer from high response delay (>500ms) and privacy risks.  Recent research has highlighted the need for on-device AI solutions to address these performance and security concerns. This study explores a local deployment approach by using the DeepSeek model with localized model compression, the paper successfully ran LLM-powered NPC dialogues within the game environment. Tests show that our method reduces response delay to 14% of the cloud-based methods. Meanwhile, it achieves diversity scores of 0.22 and 0.35 based on TF-IDF and BERT, significantly enhancing the use of using LLMs in video games. Our findings demonstrate that this approach significantly improves the use of LLMs in games. It provides a low-latency, privacy-friendly solution and presents a new way to integrate AI into the gaming industry.