Normative Conflict Resolution for Autonomous Agents

59 concepts35 papers1 proposal

Imagine a social robot programmed with two simple rules: "Keep the floor clean" and "Never disturb a human." What should it do if a person is standing directly over a spill?

In classical logic, a system facing flatly contradictory obligations suffers from logical explosion or computational paralysis. Yet, as autonomous systems move from isolated environments into complex human spaces, they must navigate a dense, overlapping web of social norms, legal obligations, and organizational codes. As shown by Özgür Kafali et al. in Kont: Computing Tradeoffs in Normative Multiagent Systems, real-world rule sets are rarely perfectly consistent. An agent's internal objectives frequently collide with external social constraints David R. Traum et al., [Zohreh Shams](/author/zohreh-shams) et al., such as when a safety protocol conflicts with an ownership norm Zhi-Xuan Tan et al..

To prevent logical deadlocks and ensure safe operation Martin J. Kollingbaum et al., the field of normative conflict resolution provides formal mechanisms to handle these inevitable collisions. Instead of treating rules as rigid, unbreakable laws, these frameworks give agents the mathematical tools to weigh trade-offs, break logical loops, and choose the "lesser of two evils."

Soft Constraints and the Economy of Violations

The first core intuition for resolving rule conflicts is to transform a binary logical contradiction into an optimization problem. If an agent cannot satisfy every rule simultaneously, it should find the path that minimizes total normative harm.

In Norm Conflict Resolution in Stochastic Domains, Daniel Kasenberg et al. propose treating norms as suspendable constraints with associated violation costs. When an agent enters a "no-win" situation, it calculates a compliance score for each potential sequence of actions.

This mechanism is formalized as a "lesser evil" logic by Ezio Bartocci et al. in A Normative Supervisor for Reinforcement Learning Agents and Enforcing ethical goals over reinforcement-learning policies. When all available actions violate at least one constraint, a selection algorithm minimizes the total weight or count of the violated norms.

This operates on an economy principle of normative resistance: the agent chooses the path of least resistance. For example, an autonomous vehicle might temporarily violate a minor traffic rule (like crossing a solid white line) to satisfy a high-stakes safety obligation (avoiding an obstacle). This optimization framework allows systems to distinguish between avoidable harm and necessary, justifiable trade-offs [Sander Beckers](/author/sander-beckers) et al..

Argumentation and Non-Dominated Commitments

A second approach models rule conflicts as a dialectical debate rather than a numerical optimization problem. Here, rules, goals, and plans are treated as discrete arguments within a formal graph, where contradictions are represented as directed "attacks."

Zohreh Shams in Normative Practical Reasoning: An Argumentation-Based Approach demonstrates that by analyzing these attack relationships, an agent can compute a "justified extension"—a set of actions that balances internal utility with external regulations. This framework relies on defeasible reasoning, meaning a general norm remains valid in the agent's database but can be defeated or overridden by a more specific situational context Alex Raymond et al..

Under this paradigm, system compliance is fundamentally redefined. Instead of requiring the fulfillment of all active duties, Nirav Ajmeri et al. argue in Coco: Runtime Reasoning about Conflicting Commitments that compliance means fulfilling a maximal set of nondominated commitments.

[Lane Keeping Norm] <------- (Attacks) ------- [Obstacle Avoidance Norm]
                                                    |
                                             (Dominates Contextually)
                                                    v
                                        [Justified Action: Swerve]

If a situation-dependent dominance rule dictates that Obligation A supersedes Obligation B, ignoring B is formally recognized as a justifiable non-performance rather than a true norm violation. Because this process relies on explicit, human-readable logical relationships, the agent can engage in structured dialogues to explain and justify exactly why it prioritized a specific goal over a rule [Zohreh Shams](/author/zohreh-shams) et al., Alex Raymond et al..

Deontic Cycle Elimination and Permission Consistency

Sometimes, conflicting rules create structural loops, or "deontic cycles," where Rule A mandates an action that Rule B implicitly forbids, paralyzing the agent's reasoning cycle.

To resolve these structural deadlocks, Leendert W. N. van der Torre et al. developed a framework in Rights, Duties and Commitments between Agents that ranks normative worlds and applies a cycle elimination procedure. By identifying the loop and systematically removing its weakest logical links, the system derives a clean, consistent set of all-things-considered obligations.

A parallel challenge occurs when multiple agents interact in a shared environment and are granted permissions that lead to mutually exclusive outcomes. For instance, if Agent 1 is permitted to close a valve and Agent 2 is permitted to keep it open, the system contains a normative inconsistency.

To prevent this, Qi Shi in Agentive Permissions in Multiagent Systems introduced specific inference rules directly into the multiagent logical system. These rules mathematically restrict the system from issuing permissions that allow different participants to simultaneously ensure contradictory states, ensuring that the collective action space remains executable.

The Challenge of Value Hierarchies

While these mathematical and logical engines successfully resolve conflicts given a set of priorities, determining those priorities dynamically remains a major hurdle. Autonomous systems require a robust, context-aware hierarchy of values to guide their choices Vahid Yazdanpanah et al., Louise A. Dennis et al..

This limitation is starkly visible in modern AI models. As demonstrated by MohammadHossein Rezaei et al. with the EgoNormia: Benchmarking Physical-Social Norm Understanding benchmark, state-of-the-art vision-language models frequently fail when forced to navigate trade-offs between overlapping social values like politeness, helpfulness, and personal safety. They lack the structured hierarchical value systems necessary to adjudicate between conflicting rules of thumb in nuanced scenarios [Tao Shen](/author/tao-shen-dc77) et al..

To bridge this gap, researchers are exploring grounding agent behaviors in universal safety baselines, such as negative human rights Ondrej Bajgar et al., and integrating formal verification with norm-aware mechanism design Nicola Basilico et al.. This ensures that when rules inevitably collide, the agent's resolution strategy is both mathematically sound and socially acceptable.

Key Foundational Work

A Logic of Deliberation (1986) — contributed 1 concept elements to this direction.
Discourse Obligations in Dialogue Processing (1994) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 33 papers
  2. Version 1 · 2026-10-04 · 70 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_normative_conflict_resolution_ebe445bc,
  title = {normative conflict resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-conflict-resolution/rdr_ebe445bc7ee847bd811e293839faf62c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_normative_conflict_resolution_ebe445bc_v2,
  title = {normative conflict resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/normative-conflict-resolution/rdr_ebe445bc7ee847bd811e293839faf62c/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake