Securing Agentic AI through Logic Tensor Networks

Kurzfassung

Logic Tensor Networks (LTNs) provide a neurosymbolic framework integrating first-order fuzzy logic with differentiable tensor neural networks to unify learning and reasoning. Logical constants become feature vectors, formulas yield truth values in [0,1], and computations are fully differentiable. LTNs solve diverse AI tasks—multi-label classification, relational learning, clustering, regression, embedding learning, query answering—enabling end-to-end training with symbolic constraints. Recent extensions like FOLAR leverage LLMs for axiom generation in interpretable stance detection, training LTNs and fusing outputs via multi-decision mechanisms for robustness. Similarly, LTP uses shallow LTN probes on LLM embeddings to localize and quantify logical reasoning across layers. We aim to test similar approaches on Agentic AI for train driving applications.