Abstract: As an alternative to LLM deployment in low-resource scenarios, Small-scale Large Language Models (SLLM) still have a significant gap in task processing capabilities compared to LLM. This paper proposes a question reasoning framework for SLLM. By introducing a reasoning path ranking selection mechanism based on Step-Semantic Verifier on Reasoning Paths (SSVRP) between the two stages of reasoning path generation and answering generation, the reasoning capability of SLLM is improved without human intervention. Experimental results show that SSVRP effectively enhances the reasoning performance of SLLM, with average accuracy of 54.3%, 90.6%, 64.3% and 63.7% in four reasoning tasks, respectively, which are the latest SOTA results in three of the reasoning tasks.