LLMs fail at group coordination tasks where humans adapt and improve

Researchers tested large language models and humans on a collaborative game where players had to coordinate their guesses to hit a target number using only group feedback. Humans learned to stabilize their strategy over time, while LLMs mostly failed to improve and switched between guesses erratically, suggesting they don't have the adaptive coordination abilities that humans develop naturally.

This is the first concrete evidence that LLMs struggle with one of the most basic human capacities: learning to work with others toward a shared goal. The gap isn't small — it's behavioral. When humans get richer feedback, they use it to adjust strategy. When LLMs get richer feedback, it barely changes their output. This matters because coordination underpins everything from business teams to emergency response to manufacturing — and if LLMs can't do it even in a simplified lab game, deploying them in real teams will create friction, not fluency.