Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning and fact verification tasks, yet their performance remains constrained by inherent uncertainties in multi-step reasoning processes. Although the Multi-Agent Debate (MAD) framework significantly enhances reasoning accuracy through collaborative interactions among agents, these agents often generate divergent answers to the same question, which may inadvertently propagate incorrect responses. In such scenarios, greater priority should be granted to agents that consistently provide accurate answers. To address this challenge, this paper proposes a novel Dynamic Weighted Consensus Framework for LLM Multi-Agent Debate (DWC-MAD). This framework integrates real-time confidence quantification (derived from agents’ iterative responses) and longitudinal accuracy metrics (based on agents’ historical performance). By employing a dynamic weight allocation algorithm, DWC-MAD assigns differentiated weights to agent inputs, thereby constructing an optimized multi-agent debate system.