Human societies are governed by normative systems: shared standards generate \textit{norms} that dictate acceptable behavior and are enforced through community sanctions. Aligning increasingly autonomous AI systems with these norms is a central alignment challenge because norms are numerous, evolve rapidly, and are often arbitrary (e.g., dress codes or language conventions). Hence, alignment requires \textit{normative competence}—the ability to infer from interaction alone which norms a community enforces, without relying on static pretrained knowledge.
We introduce a multi‑agent community debate setting where access to debate is controlled by synthetic norms, allowing us to study normative competence in isolation from pretraining exposure. Baseline LLM agents fail to learn the norms even when doing so would improve their accuracy.
We then experiment with various \textit{normative modules}—architectural components for norm inference—and find that norm‑following is highly sensitive to both the style of the norm and the model powering the normative module, suggesting a lack of generalizability. Moreover, when idiosyncratic, non‑normative behaviors accompany the true norm, LLM agents exhibit an unselective attribution failure: they indiscriminately copy the noise alongside enforced rules, a pattern that persists even when imitation of unnecessary behaviors is explicitly penalized.
To the best of our knowledge, this work is the first to operationalize and evaluate normative competence in LLMs, demonstrating that current AI systems excel at behavioral mimicry but lack the capacity to discern socially enforced order.
Review: The paper highlights a fundamental shortcoming of LLMs in understanding and adhering to community norms, offering a clear benchmark and pointing toward future research on socially aware AI.