Debating Truth: Debate-driven Claim Verification with Multiple Large Language Model Agents

DebateCV is a multi-agent LLM framework where debaters argue opposing stances on claims while a moderator weighs evidence to render verdicts. The authors introduce Debate-SFT, a post-training method using synthetic data to fix zero-shot moderator bias toward neutrality. The framework outperforms single-agent baselines on accuracy and justification quality.