Symbolic-JEPA: Where Symbolic AI Meets Joint-Embedding Predictive Architecture for NL–FOL Conversion

Bridging natural language (NL) and formal logic, such as first-order logic (FOL), remains a fundamental challenge for developing trustworthy and explainable artificial intelligence (AI) systems. Despite recent progress, current NL–FOL systems still lack unified semantic–logical representation learning; most adopt late-fusion architectures that loosely couple large language models (LLMs) with symbolic solvers. Furthermore, extended reasoning chains often introduce inefficiencies and lead to violations in logical consistency, structural validity, and formal constraints. In this paper, inspired by Neuro-Symbolic Type-4&5 architectures and the Joint Embedding Predictive Architecture (JEPA), we propose Symbolic-JEPA , a neuro-symbolic predictive representation model that unifies semantic and logical learning within a single framework. The model incorporates symbolic priors into JEPA, ensuring that the generated FOL expressions remain both semantically faithful and logically consistent. Symbolic-JEPA comprises four key components: a Symbolic Graphizer that unifies NL and FOL inputs into a shared graph space; a Structure-Aware Node & Path Encoder (SANE) that fuses linguistic and structural cues; a Structure-Guided Attention Transformer (SGAT) that injects structure-biased attention; and a Predictor with Micro-Attention that performs clause- and alignment-aware inference efficiently. To support paragraph-level conversion and evaluation, we further introduce ParaLogic , a large-scale NL–FOL dataset with inter-sentence logical annotations (9,344 samples, 28K expressions), along with a three-dimensional evaluation framework spanning parse-, conversion-, and reasoning-level metrics. Experimental results show that Symbolic-JEPA achieves consistently higher logical accuracy, structural well-formedness, and generalization capability compared to neural and LLM-based baselines.