A new study finds large language models can reason strategically, but only up to a point.
Researchers built a framework that separates three parts of strategic thinking: forming beliefs about what an opponent will do, evaluating possible moves against those beliefs, and picking a best response. They tested frontier LLMs across a series of non-cooperative games, including a new context-free game designed specifically to stop models from just recalling memorized textbook examples. Comparing each model's stated choices with its written reasoning, the researchers found models can execute a best response when given an assumption about the opponent, and will form their own opponent-specific conjectures when left to their own devices.
As the games got more complex, that careful step-by-step reasoning broke down. Models swapped explicit recursive logic for heuristics and shortcuts that differed from model to model, and those shortcuts did not match the predictable biases seen in human strategic behavior. That combination makes model missteps in complex, high-stakes settings harder to anticipate than errors from a human negotiator.
Firms are already floating LLMs for negotiation, policy design, and market simulation, exactly the complex settings where this study says the reasoning is thinnest. Passing a clean game-theory test in a lab is a long way from holding your own across a bargaining table.