neuro-symbolic integration

Neural-symbolic AI research has long prioritized learning and reasoning while critically neglecting interpretability, trustworthiness, and metacognition—key dimensions for human understanding and reliable AI deployment. This study conducts a systematic literature review (2020–2024) of 167 high-quality papers from IEEE Xplore, arXiv, ACM, and other venues, adhering to the PRISMA framework. We quantitatively identify three critical gaps: limited interpretability (28% of papers), absence of explicit trustworthiness modeling, and scarce metacognitive integration (only 5%). While learning & reasoning (63%), knowledge representation (44%), and logical reasoning (35%) dominate, interdisciplinary convergence remains weak. Our contribution is a novel integrative roadmap grounded in cognitive science, philosophy, and formal verification to advance robust, trustworthy, and context-aware neural-symbolic AI. All analytical code is publicly released to ensure full reproducibility.

Formal verification of software architecture remains impractical in industry due to prohibitively high modeling costs. Method: This paper proposes Neural Architecture Inference—a novel approach that automatically learns structured, verifiable architectural models from source code or runtime traces. It establishes the first neuro-symbolic paradigm for architecture inference, integrating graph neural networks and sequence modeling with formal specification languages (e.g., TLA+, Alloy) to close the learning–verification loop. Contribution/Results: We define a six-dimensional research roadmap and present a framework for automated generation of interpretable, formally verifiable architecture models. Experiments demonstrate substantial reduction in modeling effort and enable symbolic verification of architectural constraints—including layer isolation and communication protocols—thereby providing both theoretical foundations and practical pathways for industrial-scale architecture governance.

This work addresses the fundamental challenge of lacking formal verifiability in neural networks by introducing the novel paradigm of *Proof-Carrying Neural-Symbolic Code*, which unifies deep learning models and machine-checkable correctness proofs within a single executable code artifact—marking the first such integration. Methodologically, it synergistically combines formal verification (using Coq/Lean), neural program synthesis, differentiable symbolic execution, and automated theorem proving. We implement the first end-to-end prototype system capable of automatically generating and verifying machine-checkable proofs for small-scale neural-symbolic functions. Empirical evaluation demonstrates successful verification on multiple safety-critical micro-benchmarks. By bridging neural flexibility with symbolic rigor, this work establishes a principled foundation for building trustworthy AI systems that simultaneously possess strong learning capabilities and mathematically grounded safety guarantees.

Current artificial intelligence systems predominantly rely on monolithic models that tightly couple perception, reasoning, and decision-making, resulting in low transparency, limited scalability, and difficulty in continuous evolution. This work proposes a neuro-symbolic architecture centered on composability, introducing an innovative “symbolic seam” mechanism that explicitly defines typed objects, versioned constraint bundles, and decision traces at module boundaries. This approach enables the organic integration of data-driven components with formal symbolic constraints. The architecture supports modular composition and dynamic evolution, significantly enhancing system verifiability, transparency, and scalability, thereby offering a new paradigm for building evolvable intelligent systems.

Neural-symbolic program verification suffers from a semantic gap—termed the “embedding gap”—between neural components and symbolic logic. Method: This paper formally defines the embedding gap and proposes an end-to-end formal verification framework comprising: (1) a domain-specific language (DSL) to declaratively specify problem-space properties; (2) a multi-backend compiler that enables declarative, compilable mapping from the problem space to the embedding space, seamlessly interfacing PyTorch, Marabou, and Lean; and (3) modular, co-verification across training environments, neural verifiers, and theorem provers. Contribution/Results: We demonstrate fully automated, reproducible, and mathematically rigorous safety verification on a simplified autonomous driving system equipped with a neural controller. Our approach systematically bridges the semantic divide between neural and symbolic verification, enabling principled integration of learning-based and logic-based reasoning within a unified formal framework.

This work investigates whether neural networks can learn to execute full programs end-to-end—not merely subtasks such as arithmetic or logical reasoning. Addressing prior models’ reliance on structural biases or restricted program spaces, we adopt Turing-complete λ-calculus as a formal benchmark and introduce the first purely data-driven neural-symbolic system: no syntax or reduction strategy is hard-coded; instead, a Transformer learns functional computation intrinsically via supervised learning on β-reduction traces. Our method integrates symbolic token embedding, differentiable reduction modeling, and sequence-to-sequence learning. On synthetic benchmarks, the model achieves 98.2% step-level reduction accuracy and 92.1% full-program execution correctness, demonstrating that neural networks can generalize to semantically execute arbitrary λ-terms. This establishes a novel paradigm for neural-symbolic AI capable of true program-level reasoning.

本文提出了一种基于神经元激活的符号框架，使用SMT求解器等逻辑引擎有效计算深度神经网络行为的解释，解决了现有技术无法处理深层架构的问题。

This work addresses the insufficient reliability and trustworthiness of AI systems in high-stakes or data-scarce scenarios by proposing RAIL—a unified design framework for neuro-symbolic AI grounded in four principles: Reasoning, Assurance, Interface, and Learning. Integrating cutting-edge techniques such as physics-informed learning, causal inference, and tool-augmented large language models, RAIL offers engineers actionable design guidelines through neuro-symbolic integration, formal reasoning, and neural-guided search. The framework not only fosters deep synergy between neural and symbolic approaches but also substantially enhances AI system performance in reliability, efficiency, and trustworthiness, demonstrating broad applicability across real-world domains.

This study addresses the lack of systematic guidance for selecting logic programming languages in neuro-symbolic AI. It surveys rule-based languages—including Datalog, Answer Set Programming, and probabilistic logic programs—across four dimensions such as semantics and expressiveness, while analyzing over fifty systems. By constructing application-feature mapping and decision matrices, this work compares formal methodological differences across domains and examines integration techniques with neural networks. Ultimately, it establishes a selection guideline for rule-based languages alongside a future research roadmap, providing theoretical support for system design and the resolution of open problems in this field.

Reactive synthesis faces dual challenges of high algorithmic complexity and the difficulty of writing formal specifications. This work proposes a neurosymbolic approach that, for the first time, incorporates natural language specifications into reactive synthesis by leveraging a large reasoning model to generate Verilog circuits and integrating a model checker to provide symbolic feedback for iterative refinement. The method establishes an end-to-end natural synthesis pipeline that outperforms existing specialized tools on benchmarks from the annual synthesis competition. Notably, it achieves performance comparable to hand-crafted formal specifications when using natural language inputs and scales to the synthesis of undecidable parameterized systems.

This work addresses the challenge of neural-symbolic learning in achieving differentiable semantics while preserving the formal properties of logical connectives. The paper proposes Quantitative Linear Logic (QLL), the first differentiable logic framework grounded in the principle of “naturality,” which directly maps logical constraints to additive and log-sum-exp operations commonly used in machine learning. By defining semantics in logit space, QLL seamlessly integrates neural network training with formal verification, maintaining both logical rigor and compatibility with gradient-based optimization. Experimental results demonstrate that QLL exhibits robustness under adversarial attacks that closely aligns with its formal verification guarantees, significantly outperforming existing neural-symbolic approaches.