The emergence of large language model (LLM) has promoted the research progress in many fields, but it still faces challenges in imitating human logical reasoning, especially in the step-by-step reasoning of complex tasks and zero-shot logical cognition. To address these challenges, we propose a multi-step symbolic reasoning strategy that decomposes complex tasks into subtasks and optimizes the decomposition using a subtask verification module. Moreover, we also introduce a new zero-shot symbolic module which can help improve the model’s reasoning ability on unseen samples with symbolic representation and logical schemes. We evaluated our method on four reasoning datasets: the industrial private dataset Ship Assembly Technology and the public datasets ProntoQA, ProofWriter, and OpenBookQA. Our framework demonstrates substantial improvements in reasoning interpretability and generalization capacity compared to existing prompting paradigms. The proposed method establishes a new pathway for enhancing LLMs' cognitive architectures through symbolic system integration, showing strong potential for efficient knowledge transfer to downstream applications while preserving human-understandable reasoning traces.