A Comparative Study of LLM-Driven and Scripted NPCs: Effects on Task-Oriented Interaction Fluency and Information Acquisition Performance

This study investigates the computational effects of different non-player character (NPC) dialogue architectures on players’ task-oriented interaction processes. Under a controlled condition in which task facts, scene layout, and completion constraints were held constant, a single-factor between-subjects experiment was conducted to compare scripted NPCs and LLM-driven NPCs. From a computer-technology perspective, NPC type was operationalized as a dialogue-generation mechanism embedded in the same interactive system, while task-oriented interaction fluency and task-related information acquisition performance were quantified through system-generated behavioral logs. Statistical analysis was centered on descriptive statistics, independent-samples t-tests, and effect size estimation based on behavioral data automatically captured during task execution. The results showed that the LLM-NPC condition produced higher mean values on both indicators than the scripted NPC condition, although the between-group differences did not reach the conventional significance threshold. Nevertheless, medium effect sizes were observed for both outcomes, indicating that the computational form of dialogue generation may still exert a meaningful influence on task-oriented interaction. These findings suggest that NPC mechanisms should be understood not only as content-delivery tools, but also as interactive computing systems that shape information processing efficiency, behavioral pacing, and quantitative performance outcomes. The study provides preliminary statistical evidence for evaluating LLM-driven NPCs within task-based digital interaction systems and offers a computable behavioral-analysis framework for subsequent research.