Advancing Neuro-symbolic Causal Rule Synthesis and Verification for Safety-critical AI Systems

This paper addresses the limitations of rule-based systems in safety-critical domains by extending a neuro-symbolic causal framework that integrates first-order logic and deep reinforcement learning. The authors introduce a meta-level layer to enhance scalability and mitigate goal misspecification through a Goal/Rule Synthesizer and a Rule Verification Engine. The synthesis process utilizes large language models to decompose goals, consolidate semantics, and translate them into formal rules. The verification process checks for syntax, logical consistency, and safety before integrating the rules into a knowledge base. Evaluation in autonomous driving scenarios demonstrates the framework's ability to derive minimal rule sets based on human-specified goals, supporting a traceable synthesis grounded in legal and safety principles.