We introduce an innovative method that combines large language models (LLMs) with logic programming (LP) to address complex reasoning tasks. This approach leverages the formal structure of LP to enhance the consistency of problem-solving by LLMs. In our framework, the LLM operates independently to generate reasoning steps and constructs a corresponding LP representation. The LP module then processes these reasoning steps, providing formalized results. The LLM subsequently interprets these LP outputs and formulates adversarial challenges against its initial conclusions to reconcile inconsistencies. This adversarial interaction between the LLM and LP—where each agent aims to refine or challenge the other’s conclusions—improves the reliability and accuracy of the LLM’s predictions and recommendations. We validate our LP-based adversarial neuro-symbolic framework using various reasoning datasets, comparing its performance to state-of-the-art neuro-symbolic systems. While our approach demonstrates comparable performance across the full dataset, it significantly outperforms competing systems on subsets containing contentious or highly complex tasks, underscoring its robustness in handling intricate reasoning challenges.