Relational Norms: Definitions & Applications

• Relational norms are relationship-specific, context-dependent rules that prescribe appropriate cooperative behaviors across varied social roles.
• They are operationalized using empirical annotations, Bayesian inference, network-dynamical models, and temporal logic frameworks to capture social context.
• In AI and multi-agent systems, relational norms guide role-appropriate behavior and sanction design, enabling robust norm inference and sustainable cooperation.

Relational norms are relationship-specific, context-dependent rules that prescribe or proscribe behaviors in social interactions, often structuring cooperation, communication, and expectations between agents. Unlike universal norms, relational norms are explicitly tied to particular social roles, relationships, or network structures, determining what is judged appropriate within a given dyadic or multi-agent context. Their formalization spans empirical annotation, Bayesian inference, network-dynamical modeling, multi-objective regulatory design, and temporal logic frameworks.

1. Defining Relational Norms: Taxonomies, Formalisms, and Context

Relational norms are defined as “relationship-specific patterns of prescription and proscription” that determine which cooperative behaviors are appropriate in dyadic or group relations (Earp et al., 17 Feb 2025, Jurgens et al., 2023). In the context of human interactions, these norms emerge because distinct social roles (e.g., teacher–student, parent–child, employer–employee, friends, romantic partners) solve different coordination problems and generate unique relationship “goods” such as care, mutual support, hierarchical ordering, fair exchange, and mating. Formally, the normative profile of a relationship type $r$ can be encoded as a four-tuple:

$$N(r) = \langle h_{(r)}, c_{(r)}, t_{(r)}, m_{(r)} \rangle$$

where $h$ (hierarchy), $c$ (care), $t$ (transaction), and $m$ (mating) specify the prescriptive, neutral, or proscriptive strength of each cooperative function in role $r$ (Earp et al., 17 Feb 2025).

Relational norms operate not only at the dyad level but also within higher-order social structures, such as triads or cliques, where an agent’s behavior is adapted based on the collective local action profile, not simply pairwise interactions (Ma et al., 2024). In this sense, norms can be distributed as shared constraints across agents, specifying injunctive regularities that are contextually shared rather than idiosyncratic preferences (Tan et al., 2019).

2. Empirical Operationalization and Inference of Relational Norms

Relational norms can be directly annotated or inferred from observed behaviors and relationship labels. For interpersonal communication, appropriateness is modeled as a binary variable $y(m, R) \in \{0,1\}$, representing whether utterance $m$ is judged as appropriate in relationship $R$ (Jurgens et al., 2023). Data collected from large-scale human annotation tasks demonstrates that contextually-appropriate communication depends critically on relationship categories—what is acceptable from a friend is not necessarily appropriate from a boss or stranger. The functional distinctions between categories (e.g., family vs. peer vs. organizational) are reflected in block-structure of appropriateness matrices, with substantial context-sensitivity: for ∼19% of utterances that are appropriate in one relationship, annotated judgments flip to “inappropriate” when presented in another (Jurgens et al., 2023).

Bayesian Theory of Mind models operationalize relational norms as latent shared constraints. Norms $N\in\{0,1\}$, desires $D_i$, and observed actions $A_i$ for each agent are incorporated in a graphical model. Posterior inference then decomposes observed behavior into private desires and contextually-shared norms, enabling the estimation of norm existence and generalization across agents (Tan et al., 2019). Experimental validation of this model shows that participants’ inferences on norm existence closely match Bayesian predictions, with multi-agent evidence used to attribute causality to either individual preference or social norms.

3. Formal Representation in Multi-Agent Systems and Logic

In regulatory and multi-agent settings, relational norms are made explicit through algebraic, combinatorial, and logical frameworks that capture dependencies and mutual constraints among norms. In norm nets $NN = \langle N, R \rangle$, relational structure is encoded via three binary relations:

• Generalisation ($R_n$): one norm subsumes another.
• Exclusivity ($R_x$): mutually incompatible norms.
• Substitutability ($R_s$): norms that are functionally interchangeable.

These relations are translated to formal constraints in integer linear programming formulations for optimal norm selection, with each relation type restricting feasible sets of enacted rules (Lopez-Sanchez et al., 2017). Selection objectives may then maximize regulatory coverage, minimize cost, and optimize for supported moral values, subject to relational constraints encoded in the LP.

In temporal logic (CTL + STIT + TPTL), relational norms are constructed using first-class violation modalities. For example, violation of norm $i$ by action (acting-violation) or omission (omission-violation) can activate new obligations (“pay fine” on violation, “call out” on bystander detection), chaining norms parametrically upon violation events. This enables mapping complex sanction and reparation structures as directed graphs over norms and agents, where obligations and permissions are recursively defined through violation patterns (Mellema et al., 2022).

4. Dynamics: Evolution, Network Effects, and Emergence

Norms are not static; their emergence and stability depend on evolutionary pressures, network topology, and information flow. In higher-order networks, relational norms can leverage triadic (or higher) interactions: agents revise their behavior toward a neighbor $j$ not just on individual payoff signals, but based on the collective cooperation state $s'_{ij}$ aggregated from the local group (Ma et al., 2024). Evolutionary dynamics show prosocial norms (SIC, AIC) spontaneously dominate under moderate triangle density and payoff structure, generating stable cooperation even where pairwise-imitation yields only defection.

The effect of information privacy is nonmonotonic: moderate privacy levels (e.g., partial observation of local actions) maximally sustain both the prevalence of prosocial relational norms and cooperation. With full transparency or extreme isolation, both antisocial behavior and norm breakdown increase. Thus, relational norms help maintain robust cooperation via higher-order contextual cues and limited transparency.

5. Relational Norms in AI, Human-AI Interaction, and Norm Reasoning

For artificial agents, especially those deployed in human-facing social or embodied contexts, relational norms are increasingly treated as essential design primitives. As AI systems occupy roles analogous to human social positions (assistant, therapist, tutor, romantic partner), relational norms guide the specification of which actions are expected, permissible, or prohibited, given the intended role (Earp et al., 17 Feb 2025). Lack of appropriate relational norm modeling in LLMs and robots results in systematic reasoning failures—including the inability to utilize implicit norms for reference resolution and to adapt language and behavior to socio-relational context (Abrams et al., 3 Feb 2026).

Experimentally, even SOTA LLMs perform poorly on norm-based reference resolution unless norms are explicitly provided in the prompt (raising accuracy from ∼44% to ∼70–99%) (Abrams et al., 3 Feb 2026). This demonstrates that social grounding and explicit norm modeling are prerequisites for robust, context-aware AI in human environments.

6. Implementation, Reasoning, and Practical Constraints

In normative MAS and social simulation, implementation of relational norms involves:

• Maintaining per-agent, per-norm states including deadlines, flagged violations, repairs, and sanctions.
• Real-time violation detection via deontic reduction (e.g., acting vs. omission).
• Automated obligation spawning, sanction assignment, and closure on repair.
• Chaining of norms parametrically—supporting meta-norms, reparations, and sanctions as explicit relational dependencies.
• Explicit conflict detection and resolution using the analytic properties of violation modalities (persistence, exclusivity, splitting, normality) (Mellema et al., 2022).

In regulatory optimization, relational norm constraints are directly mapped to LP exclusion constraints, with selection criteria determined by application prioritization (coverage, cost, value alignment) (Lopez-Sanchez et al., 2017).

7. Empirical Impact and Theoretical Significance

Relational norms operationalize the social grammar that underpins human cooperation, enforcement, and context-appropriate behavior across a spectrum of domains. Their explicit representation allows empirical quantification of context sensitivity, model-based inference of latent socio-normative constraints, and the design of normative frameworks for artificial agents. The mathematical and computational formalizations reviewed here enable precise, modular, and scalable deployment of relational norms in human, multi-agent, and human–AI systems, supporting both foundational research and practical engineering of norm-guided behavior (Earp et al., 17 Feb 2025, Jurgens et al., 2023, Mellema et al., 2022, Ma et al., 2024, Abrams et al., 3 Feb 2026, Lopez-Sanchez et al., 2017, Tan et al., 2019).