A Semantic Verifier for Optimizing Small-Scale Large Language Models on Reasoning Tasks

Large language models (LLMs) with more than 100 billion parameters have revolutionized various tasks related to natural language processing and have had a profound impact in the field of artificial intelligence. However, deploying LLMs in the real world could also result in increased production costs. Small-scale Large Language Models (SLLMs), which are smaller, compact LLMs with fewer than 10 billion parameters, could significantly reduce production costs compared to LLMs. However, they typically perform less effectively than LLMs in general. Although In-context learning prompting has successfully enhanced the capabilities of SLLMs, the construction of prompts requires a certain level of human expertise. In this study, we explore enhancing SLLMs in emulating the performance of LLMs in reasoning tasks at a minimal cost, without any prompts provided by humans. We employ two SLLMs and incorporate a ranking model based on a Semantic Verifier between them to facilitate reasoning tasks. Experiments conducted on four publicly available datasets for reasoning tasks demonstrate that our approach effectively enhances the inference performance of SLLMs, and it achieves new state-of-the-art results.