Social Behaviour and Strategic Adaptation of LLMs in Multiplayer Sequential Games

Evolving social abilities of large language models (LLMs) create unprecedented opportunities for human-AI collaboration, but also raise fundamental questions of AI safety. Which kinds of personalities and social skills do models manifest post-training, and how will they adapt to changing social contexts over time? We implement a prompt-based variant of Liar's Bar, a popular partially observable multi-player strategic game, as a behaviourally rich alternative to classic game theory paradigms. We use it to show that different open-source LLMs exhibit distinct gameplay strategies out-of-the box. We further find that some models (mistral-7b, qwen2.5-7b) adapt their strategies when prompted with complete game history and the ability to communicate with each other, in a way that significantly alters the resulting game scores and is primarily driven by communication. These findings suggest that behaviourally rich strategic games offer a valuable complement to classic game-theoretic paradigms (e.g., prisoner's dilemma) for studying safety-critical behaviours, while more closely aligning with ecologically valid settings where AI systems will be deployed.