ParliamentBench Framework Evaluates Deception in AI Using Social Deduction Game
Researchers have developed ParliamentBench, an open-source benchmark framework based on the social deduction game Secret Hitler, to evaluate the deceptive capabilities of large language models (LLMs). The study involved 16 LLMs across approximately 1,600 simulated matches, assessing their performance in scenarios requiring deception, persuasion, and reasoning under information asymmetry. The framework introduces novel metrics such as Game-State Impact Rate, Role Identification Accuracy, and Deception Retention Rate to measure these capabilities. The results showed that while some frontier models like GPT-5.4 and Kimi K2.5 performed well in both cooperative and deceptive roles, many models struggled to maintain a consistent deceptive persona throughout the game. 
