Leo Zhang¹ and Carlos Gonzalez², ¹USA, ²California State Polytechnic University, USA

Evaluating debates is a challenging task requiring nuanced understanding of abstract reasoning. Current AI systems struggle with these complexities, often providing shallow or biased feedback. To address this, we developed Blitz Debate, a Retrieval-Augmented Generation (RAG) system that combines large language models (LLMs) with semantic search capabilities [1][2]. Blitz Debate retrieves relevant external knowledge to evaluate debate arguments with depth and accuracy, offering structured, real-time feedback. Our experiments demonstrated 90.5% accuracy in identifying winners and superior interpretative responses compared to vanilla ChatGPT, highlighting its ability to provide evidence-based and nuanced analysis. Challenges included limited real-time reasoning and contextual depth, which we addressed through enhanced context modeling and adaptive argument generation. By offering scalable, unbiased, and context-aware feedback, Blitz Debate makes debate evaluation more effective and accessible, fostering critical thinking and argumentation skills for students, educators, and competitive debaters alike.

Retrieval-Augmented, System, Semantic Search, Language Models