LLM 오케스트레이션 기반 TRPG 서사 제어 엔진 설계 연구 = Design of a TRPG Narrative Control Engine Using LLM Orchestration

국문 초록 (Abstract)

최근 대규모 언어 모델(LLM; Large Language Model) 기반 상호작용 서사 시스템이 등장하고 있으나, 확률적 토큰 생성 특성으로 인해 캐릭 터와 세계 반응의 일관성이 유지되지 않아 서사적 신뢰가 쉽게 붕괴된 다. 이는 TRPG처럼 서사가 장기적으로 누적되는 환경에서 두드러진다. 본 연구에서는 이러한 문제를 상호작용 서사의 핵심 조건인 핍진성 (Verisimilitude)의 붕괴로 진단하고, 핍진성을 일관성(Consistency), 반응성(Reactivity), 연쇄성(Cascading)의 세 축으로 구조화한다. 세 축은 각 각 단일 진실 원본(SSOT; Single Source of Truth), Dungeon World의 Moves 철학 기반 액션, 상태 기반 서술(State-Based Narration)으로 구 현되며, 이를 통합한 결정론적 LLM 서사 오케스트레이션 구조인 아르마 기스 엔진(Armagis Engine)을 제안한다. 제안한 시스템은 LLM의 역할 을 자연어 행동을 액션으로 접지하는 계층과 상태 변화를 문장으로 표현 하는 계층으로 분리하고, 상태 변화는 규칙 제어 모듈이 산출한 Δstate 를 단일 진실 원본에 적용하도록 설계하였다. 프로토타입 실험에서 상태 및 서사 일관성, 재현 가능한 결정론적 반응성, 전역적 연쇄성이 재현됨 을 확인하였으며, 제안된 구조가 LLM 기반 상호작용 서사에서 일관성과 인과성을 결정론적으로 유지하는 대안으로 기능할 수 있음을 보였다.

다국어 초록 (Multilingual Abstract)

Although LLM-based interactive narrative systems have recently emerged, their probabilistic token generation often fails to maintain consistent character and world responses, which can quickly undermine narrative trust. This issue is especially pronounced in settings such as tabletop role-playing games (TRPGs), where narratives accumulate over long horizons. We diagnose this problem as a breakdown of verisimilitude, a core requirement of interactive narrative, and conceptualize verisimilitude along three axes: consistency, reactivity, and cascading. These axes are realized through a Single Source of Truth (SSOT), a Dungeon World– inspired Moves philosophy–based action space, and state-based narration, respectively. We then propose the Armagis Engine, a deterministic LLM narrative orchestration architecture that integrates these components. The proposed system separates the role of the LLM into two layers: a grounding layer that maps natural-language inputs to actions, and an expression layer that verbalizes state changes into sentences. State changes are produced as Δstate by a RuleResolver and deterministically applied to the SSOT. Prototype experiments confirm consistent and reproducible state and narrative behavior, deterministic reactivity to actions, and world-scale cascading effects, demonstrating that the proposed architecture can function as a deterministic alternative for preserving consistency and causality in LLM-based interactive narratives.

목차 (Table of Contents)

• Ⅰ. 서론 ···············································································································1
• Ⅱ. 상호작용 서사와 핍진성의 이론적 배경 ········································· 4
• 1. 캐릭터 일관성과 몰입의 심리적 메커니즘 ···········································4
• 가. TRPG의 몰입 : 애착과 주체성 ························································5
• 나. 애니메이션의 몰입 : 캐릭터 페르소나 일관성 ····························· 6
• 2. 서사적 세계의 존립 조건 : 핍진성(Verisimilitude) ··························· 7
• 가. 일관성의 붕괴 사례 :『THE iDOLM@STER』 ···························· 8
• 나. 반응성의 재현과 한계 :『The Elder Scrolls V: Skyrim』 ······· 10
• 다. 연쇄성의 확장과 제약 :『Baldur’s Gate 3』 ································ 11
• 3. TRPG의 상호작용 규칙과 서사 구조 ·················································12
• 가. 캐릭터 시트와 룰 기반 자연어 상호작용 ·····································12
• 나.『Dungeon World』의 상호작용 규칙 ··········································· 13
• 다. 자동화의 필요성과 본 연구의 위치 ···············································14
• Ⅲ. 관련 연구 및 기존 접근의 한계 ·························································16
• 1. 기호적 통제와 환각 억제 기법 ·····························································16
• 가. 기호적 접지 ·························································································16
• 나. 스키마 기반 형식 제어 ·····································································17
• 다. 외부 메모리 기반 접근의 한계 ·······················································18
• 2. 멀티에이전트 오케스트레이션과 역할 분리 ·······································18
• 가. 역할 기반 에이전트 분리 ·································································18
• 나. 협력 구조와 상호 검증 ·····································································19
• 3. LLM 중심 제어 구조의 한계 ································································20
• 4. 기존 접근의 한계와 본 연구의 기여 ···················································21
• Ⅳ. 제안 시스템의 설계 ··············································································23
• 1. 문제 정의 ···································································································23
• 2. 설계 목표와 원칙 ·····················································································24
• 3. 서사 제어 파이프라인 :『Armagis Engine』 ····································· 25
• 4. 일관성 계층 (Consistency Layer) ·······················································27
• 가. WorldState : 게임 세계의 단일 진실 원본 ································· 27
• 나. WorldLog : 엔진 내부의 전체 이력 저장 ··································· 28
• 5. 반응성 계층 (Reactivity Layer) ···························································29
• 가. ActionParser [LLM] : 자연어 문장을 액션으로 매핑 ·············· 30
• 나. RuleResolver : 액션의 월드 변화량 생성 ····································31
• 6. 연쇄성 계층 (Cascading Layer) ·························································· 32
• 가. WorldUpdater : 월드 변화량 적용 및 파급 생성 ······················ 33
• 나. Narrator [LLM] : 월드 변화에 기반한 서사 재구성 ················ 34
• Ⅴ. 구현 및 실험 설계 ·················································································36
• 1. 구현 환경 및 기술 스택 ·········································································36
• 2. 전체 시스템 아키텍처 ·············································································38
• 3. 파이프라인 구현 개요 ·············································································38
• 4. 최소 데이터 계약 ·····················································································40
• 5. 결정론/로그 정책 ·····················································································40
• 6. 시스템 구성 및 실험 시나리오 정의 ···················································41
• 가. 액션 시스템 : 플레이어 페르소나 설정 ······································· 41
• 나. 반응 시스템 : 판정 기반 NPC 반응성 설계 ······························· 42
• 다. 업적 시스템 : 칭호 기반 연쇄성 트리거 설계 ····························· 43
• 라. 평판 시스템 : Faction 기반 연쇄성 반응 설계 ·························· 43
• 마. 서사 시스템 : 호감도 기반 페르소나 일관성 설계 ··················· 45
• 바. 회귀 시스템 : 스냅샷 기반 결정론적 사망 회귀 ······················· 46
• 사. 안내 시스템 : 입력 검증 및 시스템 피드백 기능 ······················47
• Ⅵ. 실험 및 평가 ····························································································48
• 1. 실험 목표 ···································································································48
• 2. 반응성 검증 ·······························································································48
• 가. 결정론적 반응성 검증 ·······································································48
• 나. 재현 가능한 반응성 검증 ·································································50
• 3. 연쇄성 검증 ·······························································································51
• 가. 국소적 연쇄성 검증 ···········································································51
• 나. 전역적 연쇄성 검증 ···········································································52
• 4. 일관성 검증 ·······························································································53
• 가. 상태 일관성 검증 ···············································································53
• 나. 서사 일관성 검증 ···············································································55
• 5. 실험 결과 요약 ·························································································57
• Ⅶ. 결론 ·············································································································58
• 1. 연구 요약 및 의의 ···················································································58
• 2. 연구의 한계 및 향후 과제 ·····································································59
• 참고문헌 ··········································································································60
• 부록 ··················································································································65
• Abstract ··········································································································69