A new study casts doubt on whether AI agents actually reason about other players' strategies, or just crib patterns from recent history.
Researchers built a public goods game that requires recursive belief reasoning - each agent guessing what the others expect, who are in turn guessing back. They scored LLM agents' decisions against a history-independent rational expectations equilibrium, a benchmark for how a genuinely reasoning agent should behave regardless of what happened before. Then they deliberately scrambled the statistical structure of the agents' interaction history while leaving its length untouched. When those patterns were disrupted, the benefit of a longer context largely vanished, and decision quality collapsed back to the no-context baseline - and the collapse got sharper the more the game depended on players anticipating each other's moves.
That matters because a lot of current excitement about multi-agent LLM systems - agents negotiating, bidding, or coordinating on a user's behalf - assumes in-context learning is a stand-in for real strategic reasoning. This study suggests that, at least in this setting, some of that apparent improvement is closer to statistical extrapolation than belief-based reasoning, which is a bigger problem in adversarial or fast-changing situations than in the stable ones typically shown in demos.
It's an echo of the earlier debate over whether chain-of-thought output reflects real reasoning or plausible-looking narration - except this time the test is a game where you can't fake your way to the right answer twice.