Researchers Identify Three Social Norms That Help AI Agents Coordinate Better With Humans

A new study from arXiv researchers analyzed 3,456 human interactions to identify three principles underlying social norms—outcome predictability, value alignment, and advantage awareness—and incorporated them into AI agents. When tested in pedestrian-vehicle interaction scenarios, the social-norm-informed LLM achieved nearly four times higher scores than baseline approaches and outperformed human-human coordination by 43%.

Why it matters: As LLMs and AI agents become embedded in daily life, understanding how to formalize tacit social norms into explicit principles is critical for achieving natural, mutually beneficial human-AI coordination at scale.