ANDRE: An Attention-based Neuro-symbolic Differentiable Rule Extractor

ANDRE is a novel neuro-symbolic AI framework that combines deep learning with interpretable logic programming to extract first-order rules from data. The method addresses long-standing scalability and robustness issues in Inductive Logic Programming by using attention-based differentiable operators instead of rigid rule templates or fuzzy approximations.

ANDRE represents a meaningful advance in making symbolic AI systems more practical for real-world applications. Traditional Inductive Logic Programming excels at producing interpretable rules but struggles with noisy data and computational scaling, while existing neuro-symbolic approaches either require predefined rule structures or rely on fuzzy logic operators that introduce gradient instability and misrepresent logical semantics. This creates a fundamental tension between interpretability and scalability that has limited ILP deployment in industry settings. The framework addresses this by introducing fully differentiable attention mechanisms to replace hard logical operators, allowing the system to learn soft selections of predicates during rule induction. This architectural choice enables training through standard gradient descent while maintaining symbolic interpretability—the extracted rules remain human-readable first-order logic programs rather than opaque neural network weights. The experimental validation across classical benchmarks, large knowledge bases, and probabilistic datasets demonstrates ANDRE's ability to recover correct symbolic rules while outperforming existing differentiable ILP methods. Notably, robustness to label noise suggests practical viability in messy real-world datasets where ground truth labels are unreliable. For the AI research community, ANDRE advances the neuro-symbolic frontier by showing how attention mechanisms can bridge discrete symbolic reasoning and continuous optimization. The implications extend to automated knowledge discovery, logical inference in noisy settings, and potential applications in scientific discovery where both accuracy and interpretability are essential. Following developments in this space will reveal whether ANDRE's stability advantages translate into adoption for knowledge base completion, rule mining, and other industrial ILP applications.

• →ANDRE uses differentiable attention operators to learn interpretable first-order logic rules while maintaining gradient flow during training
• →The framework outperforms existing neuro-symbolic ILP methods in both rule extraction quality and robustness to label noise
• →Attention-based conjunction and disjunction operators approximate logical semantics more accurately than traditional fuzzy operators
• →The method successfully scales to large knowledge bases and probabilistic settings where classical ILP fails
• →Extracted rules remain human-interpretable symbolic structures rather than black-box neural representations