neuro-symbolic integration

To address the inefficiency and limited semantic modeling capability of neuro-symbolic AI in discriminative learning tasks involving numerous constants, this paper proposes a lightweight neural-symbolic integration method. Specifically, it injects differentiable sub-symbolic constant embeddings directly into the similarity predicates of the symbolic learning framework TILDE, enabling joint optimization of embeddings and logical rules under formal logical constraints. This approach achieves, for the first time, end-to-end co-tuning of embeddings and symbolic learners without requiring complex end-to-end architectures, thereby preserving both interpretability and computational efficiency. Evaluated on three real-world tasks, the method significantly outperforms mainstream baselines—including Logic Tensor Networks (LTN) and DeepProbLog—in F1 score. Results demonstrate its simplicity, effectiveness, and strong generalization capability across diverse domains.

Current AI frameworks suffer from a fundamental dichotomy: Python-based ecosystems (e.g., PyTorch, TensorFlow) enable efficient learning but lack formal reasoning and declarative knowledge representation; conversely, symbolic languages (e.g., Prolog, LISP) support rigorous logical inference yet are inherently non-differentiable, non-scalable, and incompatible with gradient-based optimization. To bridge this gap, we propose **Tensor Logic**, a novel programming language foundation unifying neural and symbolic AI. Its core innovation is a syntax grounded in tensor equations—integrating first-order logic rules with tensor operations (via Einstein summation) for the first time. Tensor Logic natively supports automatic differentiation, GPU acceleration, formal inference, Transformer architectures, and graph-based modeling. Experiments demonstrate seamless expression of diverse neuro-symbolic paradigms, achieving superior reasoning reliability, learning efficiency, and knowledge integration compared to state-of-the-art approaches. This work establishes a linguistic foundation for interpretable, scalable, and end-to-end trainable AI systems.

This work addresses the long-standing challenge in neuro-symbolic artificial intelligence of integrating logical reasoning and probabilistic learning due to the absence of a unified mathematical framework. The authors propose a novel formalism based on tensor networks, wherein logical formulas and probability distributions are jointly represented as structured tensor decompositions. For the first time, tensor contraction is employed as a universal inference mechanism, enabling the construction of Hybrid Logic Network—a trainable hybrid logic-probabilistic model. Through tensor decomposition, basis encoding, contraction-based message passing algorithms, and a custom Python library (tnreason), the framework achieves a unified and efficient computational treatment of both logical and probabilistic reasoning. This approach establishes a scalable, differentiable paradigm for neuro-symbolic systems, bridging symbolic expressiveness with data-driven learning.

This work addresses sequential classification and labeling tasks by introducing the first end-to-end differentiable neurosymbolic model. Methodologically, it formalizes symbolic finite-state automata as differentiable neurosymbolic primitives and integrates them with neural perception modules (e.g., LSTM or Transformer) via probabilistic semantics, enabling gradient-based joint optimization while preserving logical interpretability. The approach explicitly encodes temporal constraints and propositional logic knowledge, thereby balancing generalization capability and model transparency. Empirically, the method achieves substantial improvements over existing neurosymbolic baselines on synthetic benchmarks. Moreover, on real-world event recognition tasks, it demonstrates superior out-of-distribution generalization, higher accuracy, and enhanced robustness compared to purely neural models.

Existing neural-symbolic systems face fundamental bottlenecks in integration, generalization, and interpretability. To address these challenges, we propose NSF-SRL, the first framework leveraging Statistical Relational Learning (SRL) to establish a bidirectional closed-loop synergy between neural and symbolic components: (i) first-order logic rules guide neural predictions via differentiable symbolic execution and rule-driven prediction correction; and (ii) deep neural networks accelerate symbolic inference. The framework supports supervised, weakly supervised, and zero-shot learning, achieving significant improvements in cross-task generalization. Technically, we introduce a novel, quantifiable interpretability evaluation strategy that enables logic-rule visualization and decision attribution analysis—thereby unifying enhanced cognitive reasoning capability with model transparency.

Neuro-symbolic (NeSy) research suffers from fragmented frameworks, lack of standardized tooling, and high development barriers—most existing efforts prioritize algorithmic innovations over declarative modeling support. This paper introduces the first multidimensional analytical framework for NeSy systems, systematically comparing DeepProbLog, Scallop, and DomiKnowS along three axes: symbolic representation formalisms, neural integration mechanisms, and differentiable inference capabilities. We unify probabilistic programming, differentiable logical reasoning, and declarative modeling within a coherent computational architecture, exposing fundamental limitations in expressive power and problem-solving scope across these frameworks. Our analysis provides principled guidance for NeSy framework selection and advances the design of unified, interpretable, data-efficient, and user-friendly NeSy platforms.

Existing neuro-symbolic approaches struggle to effectively model first-order linear temporal logical reasoning involving the evolution of object attributes and relations over time. This work proposes the first end-to-end differentiable neuro-symbolic framework capable of handling quantified first-order linear temporal logic. By integrating Logic Tensor Networks with fuzzy real-valued semantics, the method enables a unified differentiable treatment of both temporal operators and quantificational structures. It thereby addresses a critical gap in differentiable neuro-symbolic modeling for dynamic knowledge reasoning. Empirical evaluation demonstrates that the proposed approach significantly outperforms specialized purely neural models on two synthetic temporal knowledge graph completion tasks.

This work addresses the logical inconsistencies that large language models often exhibit in complex reasoning due to the absence of formal axiomatic systems. To this end, the authors propose Mathesis, a novel neuro-symbolic architecture that uniquely integrates a differentiable logic engine with higher-order hypergraph representations. The framework employs a Symbolic Reasoning Kernel (SRK) to map logical constraints onto a continuous energy landscape, thereby recasting proof search as an energy minimization problem. It further combines Monte Carlo tree search with evolutionary proof strategies to enable multi-step deductive reasoning. By leveraging a Hypergraph Transformer and a semantic unification mechanism within an end-to-end trainable framework, Mathesis effectively unifies symbolic reasoning with neural learning, significantly enhancing both logical consistency and multi-step reasoning capabilities in theorem proving.

Existing approaches to learning neural representations of formal specifications either rely on computationally expensive and irreversible symbolic kernels or employ syntactic embeddings that disregard semantic structure. This work proposes a teacher–student distillation framework that transfers the semantic geometric information encoded in symbolic robustness kernels into a Transformer encoder. The student model is trained under a novel geometric alignment loss based on continuous kernel weighting, enabling supervised learning that preserves semantic fidelity. The resulting method achieves, for the first time, a semantic-preserving, computationally efficient, and invertible neural representation of formal logic. The learned embeddings accurately predict the robustness and constraint satisfaction of Signal Temporal Logic (STL) formulas, faithfully maintain semantic similarity, and support efficient formula reconstruction.

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.