Hybrid AI Systems (Symbolic + Neural)

Abstract

Hybrid AI systems that combine neural learning with symbolic reasoning offer a promising middle path between the complementary limitations of each paradigm: neural systems excel at perceptual pattern recognition and learning from data, while symbolic systems provide structured reasoning, interpretability, and systematic generalisation. Hybrid architectures range from loosely coupled systems (neural perception feeding symbolic reasoning modules) to tightly integrated approaches where neural components participate in differentiable logical inference. This study provides a comprehensive evaluation of ten hybrid AI architectures across six task domains: visual relational reasoning, natural language inference with logical constraints, scientific hypothesis generation and validation, robotic task planning, medical diagnosis with clinical guidelines, and autonomous code generation with formal verification. Architectures evaluated include NS-CL, LRNN, NTP-NN, DPL (DeepProbLog), LLM+SAT (GPT-4 with SAT solver), AlphaGeometry-inspired hybrid, Neurosymbolic Transformers (NST), Logic-LM, ARGO (Argumentation + Neural), and a proposed Unified Hybrid Reasoning Architecture (UHRA). UHRA achieves the best mean performance across all six domains (index score 7.84/10), combining differentiable logic programming with pre-trained LLM reasoning. LLM+SAT reduces logical inconsistency rates by 74.2% vs. GPT-4 alone. AlphaGeometry-inspired hybrid achieves 87.4% on mathematical proof generation. All hybrid architectures outperform pure neural baselines on systematic generalisation tasks (mean +22.4pp). A hybrid architecture selection taxonomy and design principles are proposed.