Evaluating Large Language Models in a Complex Hidden Role Game

Niklas Bauer

arXiv:2605.22826 · 2026-05-25 공개 · arXiv · PDF

chain-of-thought ai-safety llama-3-1-70b llm-deception role-identification alignment-research reasoning-enhancement game-state-impact

Abstract

Quantifying the deceptive potential of Large Language Models (LLMs) is critical for AI safety, yet difficult to achieve in uncontrolled environments. This work investigates the reasoning, persuasion, and deceptive capabilities of LLMs within the social deduction game Secret Hitler. I introduce an open-source framework and novel metrics to measure performance: Role Identification Accuracy, Deception Retention Rate, and Game State Impact Rate. By benchmarking models against rule-based algorithms and human games, I identify a gap between conversational ability and strategic depth. The study also analyzes the impact of reasoning-enhancement techniques on win rates and strategic reasoning. Neither Chain-of-Thought prompting nor internal memory bring improvements in performance, with up to 23.2% worse win rates for fascist roles. While rule-based agents align with expert human voting decisions 86.7% of the time, models like Llama 3.1 70B achieve only a 59.7% accuracy. Models playing as Fascists consistently yield negative impact scores and fail to sustain deception, resulting in roughly 40% shorter games compared to humans. These findings suggest that current architectures remain ineffective at complex, multi-turn manipulation. As capabilities advance, detecting when models begin to master these deceptive behaviors is crucial. The developed framework serves as a reproducible testbed for future alignment research.

한국어 요약

📋 한 줄 요약

**[AI 안전 / 사회적 추론 게임]** Secret Hitler 게임으로 LLM의 추론·설득·기만 능력 평가 프레임워크와 신규 metric(역할 식별 정확도·기만 유지율·게임 상태 영향도) 제시, frontier 모델도 rule-based 86.7% 대비 Llama 3.1 70B 59.7%로 격차 노출.

🎯 핵심 기여도

💡 핵심 아이디어

LLM의 기만·설득·전략 능력 평가에는 multi-turn 사회적 추론 게임이 통제된 testbed로 적합하며, 단순 win rate가 아닌 역할 식별·기만 유지·게임 상태 영향을 분리 측정해야 conversational 능력과 strategic 깊이의 격차가 정량 노출된다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: LLM 기만 능력 평가의 reproducible testbed 제공, conversational vs strategic 격차의 정량 노출, 향후 alignment 연구에 대비한 monitoring 인프라. **한계**: 단일 게임(Secret Hitler) 중심으로 다른 사회적 deception 시나리오 일반화 검증 필요, 평가된 모델 수 제한, fascist 역할의 부진은 모델별·prompt별 변동 가능.

🚀 실용적 활용