This paper introduces Deductive Beam Search (DBS), a novel approach that integrates chain-of-thought (CoT) reasoning with deductive reasoning using step-wise beam search to enhance the reasoning capabilities of Large Language Models (LLMs). DBS addresses the issue of accumulative errors in intermediate steps by deploying a verifier to ensure the deducibility of each reasoning step and its premises. The verifier, trained on synthetic and diverse deductive reasoning errors, evaluates the logical coherence between steps. The paper also presents a scalable and labor-free data construction method to train the verifier effectively. Extensive experiments across various reasoning tasks and model scales (7B, 13B, 70B, and ChatGPT) demonstrate that DBS significantly improves the performance of LLMs in arithmetic, commonsense, and symbolic reasoning tasks. The analysis shows that DBS can detect subtle and diverse reasoning errors and is robust across different model scales.This paper introduces Deductive Beam Search (DBS), a novel approach that integrates chain-of-thought (CoT) reasoning with deductive reasoning using step-wise beam search to enhance the reasoning capabilities of Large Language Models (LLMs). DBS addresses the issue of accumulative errors in intermediate steps by deploying a verifier to ensure the deducibility of each reasoning step and its premises. The verifier, trained on synthetic and diverse deductive reasoning errors, evaluates the logical coherence between steps. The paper also presents a scalable and labor-free data construction method to train the verifier effectively. Extensive experiments across various reasoning tasks and model scales (7B, 13B, 70B, and ChatGPT) demonstrate that DBS significantly improves the performance of LLMs in arithmetic, commonsense, and symbolic reasoning tasks. The analysis shows that DBS can detect subtle and diverse reasoning errors and is robust across different model scales.