Since the introduction of GPT-3, large language models (LLMs) have become a major influence in Artificial Intelligence research, delivering impressive results across various domains. However, despite their successes, LLMs face significant challenges that raise doubts about their potential to achieve Artificial General Intelligence (AGI). There are many concerns surrounding their explainability, safety, and hallucination—where models generate plausible but incorrect or nonsensical outputs—which highlight the limitations of these models. To tackle these challenges, NeuroSymbolic AI has emerged as a hybrid approach that combines the strengths of Symbolic AI with deep learning techniques to address the weaknesses of LLMs. Within this paradigm, Logical Neural Networks represent a key direction, where propositional logic or first-order logic is integrated into a fully differentiable neural network, enabling end-to-end training while preserving the advantages of logical reasoning. This paper presents a systematic review of Logical Neural Networks, examining common patterns across various architectures and assessing their ability to enhance interpretability, enforce logical constraints, and improve model reliability. Additionally, we identify the remaining challenges and drawbacks associated with these architectures, which present critical areas for future research. Addressing these issues is essential for advancing the field of NeuroSymbolic AI and unlocking its full potential in achieving more robust and reliable AI systems.