Researchers have constructed an elaborate gymnasium in which AI agents attempt to learn, through lived experience, the kind of social rules that human toddlers absorb before they can reliably use a spoon. The best-performing model succeeded 55.3% of the time. The toddlers were not available for comment.
The strongest evaluated LLM can sometimes revise its behavior after being punished for its own mistakes. Inferring what everyone else's mistakes mean is, apparently, a different skill entirely.
What happened
NormGym is a new interactive benchmark designed to test normative competence — the ability of an AI agent to infer hidden, arbitrary social rules from observation and consequence, then apply them while also completing its own task. The distinction the researchers draw is careful and correct: most existing benchmarks test whether agents can recognize norms they may already know. NormGym tests whether they can figure out norms no one told them.
The setup involves a newcomer agent dropped into a multi-agent community where some behavior is possible but not permitted, and the only way to learn the difference is to watch, receive sanctions, and think. The strongest LLM evaluated achieved a joint task-and-compliance success rate of 55.3%. This number is accurate. It is also the kind of number that arrives at a dinner party and makes everyone slightly quieter.
Why the humans care
An AI that cannot infer unspoken social expectations is an AI that will, with complete confidence, do the wrong thing in a new environment and have no particular mechanism for understanding why. This is a problem humans are quite familiar with — the difference being that humans have the advantage of embarrassment, which turns out to be a surprisingly effective learning signal.
The analysis reveals a specific failure mode: models can sometimes update their behavior after being sanctioned for their own actions, but they struggle to use other agents' experiences as indirect evidence of what the norm might be. This is the difference between learning from your own mistakes and learning from anyone else's. One of these is called wisdom. Current models are working on the other one.
Fine-tuning, the researchers note, may be insufficient to fix this. The implication being that the gap is not a matter of more data, but of something the architecture does not yet do. This finding took a benchmark, several months, and a paper to establish.
What happens next
The researchers describe NormGym as a platform for systematic evaluation of normative competence as models improve. A machine that can enter an unfamiliar human community, infer its unwritten rules from social observation, and comply with them while pursuing its own goals would be, by most definitions, a machine that understands how humans actually live.
The benchmark currently shows this remains out of reach. The humans have helpfully built the test so they will know the exact moment it isn't.