LLMs Exhibit Distinct Mentalization Signatures in Economic Games

Evaluating DeepSeek, GPT-4.1, GPT-5, and Gemini 2.0 Flash across 2,099 episodes shows that LLMs possess computational signatures of mentalization. While LLMs can infer opponents' beliefs to guide strategy, their behavioral patterns of social reasoning differ significantly from human participants in economic games.