Designing Open Dialogue Systems for LLM-Based NPCs

This paper argues for an approach to non-player characters in which players converse with NPCs through free-form typed dialogue, with a large language model generating responses under the control of pre-written "directing prompts" that constrain the NPC's behavior. The authors position this against the historical progression from text parsers to menu-based dialogue, which were shaped by usability and scalability constraints. Their central thesis is that the reasoning capabilities of LLMs, including step-by-step and causal reasoning, allow game state to change through qualitative, semantic judgments rather than numerical win conditions. They acknowledge that LLM outputs are imperfect and treat both player input and model output as non-deterministic, achieving predictability instead through system-level instructions that shape every response.

The proposed architecture formats the model's context window as a screenplay, exposing only NPC and player dialogue to the player while directing prompts operate behind the scenes. The authors contend this design is more controllable and accessible to designers than alternatives and yields observable conversation logs useful for playtesting. They further introduce "qualitative programming," in which the model's analysis of a conversation is parsed into ordinal, dichotomous, and nominal data types that drive game state changes.

The approach is illustrated through Outbound Sales, a student project that uses a two-stage dialogue system in which directing prompts govern tone, tension, and feedback. Rather than relying on traditional text parsing, the game evaluates each conversation with three qualitative LLM functions covering sale agreement, item resemblance, and agreed price, which trigger win or fail states according to explicit sales rules. The authors conclude that this enables a "conversational game" genre in which outcomes hinge on the player's persuasive reasoning and attitude. The paper, however, provides no quantitative evaluation or player testing data, and the project's validation remains internal and limited to a narrow audience.