Colossal Cave Adventure

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 paper proposes free-form typed dialogue with NPCs, where an LLM generates responses under pre-written "directing prompts" that constrain behavior, positioning this against historical text-parser and menu-based dialogue systems.
• Its central thesis is that LLM reasoning, including step-by-step and causal reasoning, lets game state change through qualitative semantic judgments rather than numerical win conditions.
• Rather than seeking deterministic outputs, the authors treat both player input and model output as non-deterministic and achieve predictability through system-level instructions that shape every response.