Enhancing Factual Question Answering with Web Knowledge and Chain-of-Reevaluation

The miniaturization of Large Language Models (LLMs) has facilitated their deployment on mobile devices; however, limited computational resources significantly restrict inference capabilities, particularly diminishing the effectiveness of Chain-of- Thought (CoT) prompting. Existing CoT methods, designed primarily for large LLMs, struggle to perform effec-tively on smaller models. This paper introduces Web Knowl-edge and Chain-of-Reevaluation (WKCoR), a novel prompting approach tailored for small LLMs. WKCoR supplements the model's knowledge base through web searches and employs a modular Chain-of-Reevaluation strategy to minimize computational overhead. Additionally, WKCoR integrates a mechanism to detect and correct hallucinations in LLM responses. Experi-mental results demonstrate that, with WKCoR, the Phi-3:14B model outperforms leading models, including GPT-3.5 Turbo and GEMINI 1.5 PRO, on over half of six question-answering benchmarks and achieves performance close to GPT-4, validating its efficacy for resource-limited mobile environments.