打破马丁戈尔诅咒：通过不对称认知潜能能量实现多智能体辩论

Multi-Agent Debate (MAD) has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models. However, a significant limitation has been identified: standard MAD cannot improve belief correctness beyond majority voting, a phenomenon referred to as the Martingale Curse. This curse occurs due to correlated errors that lead agents to converge on erroneous consensus, where debate merely reinforces collective mistakes instead of filtering out noise. To address this issue, the AceMAD framework is proposed, which utilizes asymmetric cognitive potential energy to transform MAD from a random walk into a directed convergence process with positive drift. Through a peer-prediction mechanism, agents predict their peers' belief distributions, revealing asymmetries that can enhance the overall reasoning performance of the multi-agent system.