Generative artificial intelligence (AI) solutions, such as transformer models, represent a significant step toward building next-generation wireless networks. However, their adoption is hindered by limitations such as their black-box nature (lack of interpretability), poor generalizability, and large model sizes, which result in substantial power consumption. These challenges prevent networks from achieving critical qualities like near-zero latency (due to retraining overhead), trustworthiness, and resiliency—essential for applications such as digital twin-driven smart industries, intelligent transportation systems, the metaverse, and autonomous network experiences. Addressing these challenges requires a paradigm shift in how generative AI is integrated into future networks. The focus must shift toward reasoning-driven AI models that offer out-of-domain and out-of-data generalizability, decision explainability, and sustainability. Promising approaches in this direction include causal reasoning and neuro-symbolic AI. This chapter explores how incorporating causality into future generative AI models can ground them in wireless concepts and the underlying physics of wireless transmission and reception. Additionally, combining the strengths of symbolic AI and neural networks through neuro-symbolic systems can imbue future wireless systems with logical problem-solving capabilities. These problem-solving capabilities enable the identification of network failures and rapid remediation with minimal human intervention, enhancing the overall resilience and efficiency of wireless networks. Finally, we present initial experimental results demonstrating how retrieval-augmented generation (RAG) enhances accuracy, mathematical reasoning, explainability, and the quality of assertions in wireless question-answering tasks, when compared to baseline (vanilla) large language models.