#nlp

MAD (Multi-Agents-Debate) is an open-source framework exploring the collective debating capabilities of LLMs. It addresses the 'Degeneration of Thoughts' (DoT) issue—where self-reflection leads to biased or rigid outputs—by implementing a tit-for-tat interaction between multiple agents. By assigning roles like 'devil' and 'angel' to provide mutual external feedback, MAD corrects distorted thinking and breaks cognitive rigidity. It achieves significant performance gains in Counterintuitive QA and Commonsense Machine Translation tasks.