The race to create truly intelligent AI assistants is on, and one of the biggest hurdles is giving them a "theory of mind." Can they understand what other AIs are thinking and adapt accordingly? Recent research suggests that while large language models (LLMs) excel at predicting the *actions* of other AIs (what researchers call "literal theory of mind"), they struggle to use that knowledge effectively. Think of it like this: an LLM might correctly predict that its opponent in Rock, Paper, Scissors will always play "Rock." A human would immediately exploit this and spam "Paper" to win, but surprisingly, LLMs don't. They often fail to adapt their strategy, achieving far less than optimal outcomes. This "functional theory of mind" gap is a major challenge. Even when given explicit instructions about an opponent's behavior *and* the reward structure of the game, LLMs often stumble. Various prompting strategies and even providing the opponent's actual moves as input haven't fully bridged this gap. This highlights the double-edged sword of inductive bias in LLMs: while their pre-existing knowledge can boost performance in simple scenarios, it can also hinder their ability to learn and adapt in the long run. This limitation has implications beyond games, affecting LLM applications that require understanding and responding to different users and agents. Future research needs to move beyond static tests of theory of mind and develop interactive, dynamic evaluations to unlock the full potential of LLMs in a multi-agent world.