Learning social norms enhances compatibility in dynamic human-AI coordination

Short summary

Researchers formalize tacit social norms into three quantifiable principles—outcome predictability, value alignment, and advantage awareness—to improve human-AI coordination. Using a pedestrian-vehicle interaction platform with 3,456 human interactions, they show that social-norm-informed LLMs achieve nearly fourfold higher scores than baselines and outperform human-human interactions by 43%. The findings suggest explicit norm modeling can enable more natural AI integration into human society.

• •Three social norm principles identified: outcome predictability, value alignment, advantage awareness
• •Social-norm-informed LLMs achieve 4x higher scores than baseline in human-AI coordination tasks
• •Outperforms human-human interactions by 43% in closed-loop experiments

Generated with AI, which can make mistakes.