PAVE: AI Cognitive Architecture for Rule-Breaking in Emergencies
Researchers propose PAVE (Perception, Assessment, Verdict, Emulation), a four-module cognitive architecture enabling generative agents to reason about legitimate rule-breaking in emergencies like fire evacuations. The system extracts context with authority distance, peer behaviors, and severity cues; scores necessity, proportionality, and alternatives; decides compliance or violation via a hard legitimacy gate based on agent persona; and enacts the verdict. This addresses a gap in AI behavior for cooperative settings where rule-breaking may be required.
Key facts
- PAVE stands for Perception, Assessment, Verdict, Emulation
- Addresses rule-breaking reasoning in emergencies like fire evacuation
- Four modules: Perception, Assessment, Verdict, Emulation
- Perception extracts authority distance, peer behaviors, severity-tagged cues
- Assessment scores context on five scalars including legitimacy judgment
- Verdict uses hard legitimacy gate with per-agent threshold from persona
- Emulation enacts verdict and scopes violation to justified rule
- Based on large language models for generative agent societies
Entities
Institutions
- arXiv