Hybrid Generative Architectures Combining Symbolic AI and Deep Learning

Abstract

The dichotomy between symbolic AI -- with its strengths in logical reasoning, constraint satisfaction, and interpretability -- and deep learning -- with its strengths in perceptual pattern recognition, flexible representation, and scalable learning -- has historically produced two largely separate research communities and system architectures. Hybrid approaches that integrate both paradigms have long been advocated as a path to systems that combine the statistical generalisation of deep learning with the structured reasoning of symbolic AI, but practical integration architectures that achieve this combination at scale for generative tasks remain scarce. This paper proposes the Neurosymbolic Generative Architecture (NGA), a hybrid framework that integrates a large language model backbone with three symbolic reasoning modules: a formal logic verifier (FLV) that checks LLM-generated outputs against first-order logic constraints; a constraint-guided generation controller (CGC) that steers LLM decoding toward constraint-satisfying outputs; and a knowledge graph reasoning engine (KGRE) that augments LLM generation with structured symbolic reasoning over factual knowledge. The NGA is evaluated on four hybrid reasoning-generation benchmarks: logical constraint satisfaction in text generation, structured knowledge synthesis, formal specification-to-code generation, and mathematical proof sketch generation. NGA achieves a 42.6% improvement in constraint satisfaction rate (SD = 4.8%) and a 28.4% improvement in logical consistency (SD = 3.9%) relative to pure LLM baselines, while maintaining 94.8% of LLM fluency. The study contributes the NGA specification, a hybrid reasoning-generation evaluation suite, and empirical evidence that neurosymbolic integration substantially improves constrained and logically consistent generative AI outputs.