Can LLMs discover and apply social norms dynamically in open systems?

This position paper argues that LLM-based agents should be built with explicit normative capabilities — norm discovery, norm reasoning, and norm conformance — rather than relying on the "domain knowledge expressed in symbolic logical representations" that traditional normative multi-agent systems use. The authors write that such a priori symbolic norm knowledge "is unrealistic for agents participating in open systems where norms change," and propose instead that "the MAS community should actively engage in developing normative LLM agents" by extending the generic LLM-agent architecture that Wang et al. describe — a profile module (roles, personalities, relationships), a memory module, a planning module, and an action module.

The mechanism they propose maps each normative capability onto an existing module. Norm discovery would prompt a general-purpose LLM to surface norms "implicitly captured in its world knowledge," with fine-tuning on human-annotated data where generic norms don't fit a domain or culture; because fine-tuning freezes norms into a "static set," the authors propose extending the planning module with Theory-of-Mind reasoning to identify norms emerging from interaction (for example, inferring a norm from being "told off"), and extending vector-database memory to store norms discovered this way without retraining. Norm reasoning — identifying which norm applies, detecting violations, identifying sanctions, spreading norms, enforcing compliance — would run through the same planning module, prompted either with situational information alone (asking the LLM to supply the applicable norm) or with a candidate norm attached (asking the LLM to confirm it applies).

This is a design-space proposal, not a demonstration — it sits upstream of the kind of empirical question the vault already holds evidence on. Can AI systems learn social norms without embodied experience? shows a related but distinct capability already working at scale (predicting the social appropriateness of scenarios), suggesting the "norm discovery" half of this paper's vision is closer to existing capability than the paper could have known. Can language modeling close the knowing-doing gap in AI? tackles the same traditional-symbolic-agent-versus-LLM gap this paper names, but through online RL rather than through modular prompting — a different route to the same know-how-versus-know-why problem the normative-agent authors describe.

The paper does not test whether LLMs reliably discover or apply norms in the way it proposes — no agent architecture is built or evaluated here, only sketched. It also does not resolve how a dynamically discovered norm, once stored in a vector database, gets reconciled with conflicting norms already held — the paper names norm-conflict detection as future work, not a solved problem. The implication the excerpt supports is narrow: a plausible mapping exists between normative capabilities and an existing modular LLM-agent architecture, not that the mapping has been shown to produce norm-compliant agent behavior.