Agentic Debate: Mitigating Hallucinations through Multi-Agent Adversarial Verification

ABSTRACT

The propensity of large language models to hallucinate facts severely limits their deployment in autonomous agentic workflows where verification is costly. This paper introduces 'Agentic Debate,' a multi-agent adversarial framework where a 'Proposer Agent' generates a solution, while a secondary 'Devil's Advocate Agent' actively attempts to falsify the reasoning using external search tools. The agents engage in a structured, multi-turn debate overseen by a 'Judge Agent' that evaluates the logical soundness of the arguments. Empirical results on the TruthfulQA and complex legal datasets demonstrate that adversarial verification reduces ungrounded hallucinations by 81% compared to self-reflection prompting in single agents.