Enhancing EM Score on HotpotQA with Small Models Via Prompt Engineering, Answer Generation Strategies and Knowledge Graphs

Improving multi-hop question answering (QA) with lightweight models remains challenging. This paper presents a multi-stage framework to enhance exact match (EM) performance on the HotpotQA dataset using FLAN-T5-Large. Our method combines diverse answer generation, policy-based answer selection, answer post-processing, and knowledge graph-based multi-round detection. Exper-iments show that it improves EM scores, demonstrating that compact models can deliver competitive QA results.