Learning social norms: LLMs improve human-AI coordination 4x

Learning social norms enhances compatibility in dynamic human-AI coordination, a paper by Yi Yang, Siyuan Liu, Xin Gao, Huamu Sun, Chao Liu, Qing Zhou and Bingbing Nie submitted to arXiv on 8 July 2026, finds that an LLM informed by explicit social-norm principles scored nearly four times higher than a baseline strategy in closed-loop interactions with humans. The study used a simplified pedestrian-vehicle experimental platform and a dataset of 3,456 dynamic human interactions to extract and formalize the norms that drive smooth coordination.

What did the authors do?

They built a simplified experimental platform that models pedestrian-vehicle interaction and collected 3,456 dynamic human interactions to identify underlying social norms. From those interactions the authors distilled three explicit principles: outcome predictability, value alignment, and advantage awareness. They then encoded these principles into AI agents, including an LLM, and evaluated those agents in closed-loop tasks with human partners.

The paper is 44 pages long, includes five figures and supplementary information, and is indexed as arXiv:2607.07021. The authors present the pedestrian-vehicle setup as a representative dynamic interaction that captures key interactive features common to many everyday coordination problems.

How well did the social-norm-informed LLM perform?

In the closed-loop interaction task with humans, the social-norm-informed LLM achieved a nearly fourfold higher total score than the baseline strategy and outperformed human-human interactions by 43%. The baseline is described as a non-social-norm strategy; the social-norm-informed agent explicitly applied the three identified principles during interaction.

The reported gains are presented as aggregate scores from the closed-loop experiments using the collected human interaction data and the experimental platform. The authors argue these improvements show that converting tacit, hard-to-quantify norms into explicit, quantifiable principles can materially change how well AI agents coordinate with humans in dynamic settings.

Why it matters

Formalizing tacit social norms into measurable principles removes an obstacle to predictable, considerate AI behavior in multi-agent settings. The study links a concrete dataset of 3,456 human interactions to three named principles and demonstrates large performance gains when those principles guide agent behavior. That combination of empirical human data plus an explicit, implementable rule set is what produced the nearly fourfold improvement and the 43% advantage over human-human pairings.

These results suggest a practical path for embedding norm-aware behavior into AI systems that interact continuously with people, particularly in environments where split-second coordination and implicit expectations matter, such as pedestrian-vehicle encounters.

What to watch

See whether these three principles — outcome predictability, value alignment, advantage awareness — generalize beyond the pedestrian-vehicle testbed to other dynamic coordination domains and whether independent replications reproduce the reported nearly fourfold improvement. The authors provide supplementary information and figures in the 44-page submission that other teams can use to reproduce or extend the experiments.

Additional details: the paper is listed as arXiv:2607.07021 and carries an arXiv-issued DOI link. The full author list is Yi Yang, Siyuan Liu, Xin Gao, Huamu Sun, Chao Liu, Qing Zhou and Bingbing Nie.