Towards Neuro-symbolic Causal Rule Synthesis, Verification, and Evaluation Grounded in Legal and Safety Principles

Problem
This paper addresses the limitations of rule-based systems in safety-critical domains, particularly their scalability, brittleness, and susceptibility to goal misspecification. These issues can lead to reward hacking and failures in formal verification, as AI systems often optimize for overly narrow objectives. The authors build upon their previous work on a neuro-symbolic causal framework, which integrates first-order logic abduction trees, structural causal models, and deep reinforcement learning, by introducing a meta-level layer to enhance rule synthesis and verification. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the introduction of a meta-level layer comprising a Goal/Rule Synthesizer and a Rule Verification Engine. The synthesis pipeline utilizes large language models (LLMs) to process high-level natural language goals and principles from human experts. The steps include: (1) decomposing goals into candidate causes, (2) consolidating semantics to eliminate redundancies, (3) translating these into candidate first-order rules, and (4) composing necessary and sufficient causal sets. The verification pipeline performs (1) syntax and schema validation, (2) logical consistency analysis, and (3) safety and invariant checks. The authors implemented a proof-of-concept in two autonomous driving scenarios, demonstrating the framework’s ability to derive minimal necessary and sufficient rule sets and formalize them as logical constraints.

Results
The proposed framework successfully synthesized and verified rule sets in the context of autonomous driving, although specific quantitative results (e.g., accuracy, efficiency metrics) are not detailed in the abstract. The authors claim that the pipeline supports incremental, modular, and traceable rule synthesis, grounded in established legal and safety principles. The effectiveness of the approach is implied through the successful derivation of rule sets that meet human-specified goals, but exact performance metrics against named baselines are not provided.

Limitations
The authors acknowledge that their approach is still in the proof-of-concept stage, which may limit its applicability in real-world scenarios. They do not discuss potential challenges related to the scalability of LLMs in diverse contexts or the generalizability of the synthesized rules across different domains. Additionally, the reliance on human-specified goals may introduce biases or inconsistencies that could affect the rule synthesis process.

Why it matters
This work has significant implications for the development of AI systems in safety-critical applications, as it proposes a structured approach to rule synthesis and verification that is both explainable and grounded in legal and safety principles. By addressing goal misspecification and enhancing rule maintenance, the framework could lead to more robust and reliable AI systems capable of adapting to distribution shifts. The integration of neuro-symbolic methods with LLMs may also pave the way for future research in explainable AI and formal verification, potentially influencing how AI systems are designed and deployed in regulated environments.

Authors: Zainab Rehan, Christian Medeiros Adriano, Sona Ghahremani, Holger Giese
Source: arXiv:2604.28087
URL: https://arxiv.org/abs/2604.28087v1