Confusing charges are challenging to predict due to the substantial overlap in their legal elements, which complicates the identification of subtle differentiating factors. Accurately distinguishing these charges necessitates a deep understanding of domain-specific legal nuances and the application of intricate reasoning processes. To enable a more distinctive and nuanced analysis of multiple confusing charges, this paper introduces the Debate-Driven Legal Reasoning (DDLR) framework, a novel multi-agent debate approach that simulates legal argumentation to disambiguate charges. In DDLR, agents take specific stances on candidate charges, refining their reasoning via contradiction detection and fact-statute alignment in turn-based debates. The debate is managed by adjudication mechanisms to ensure both efficiency and accuracy. Evaluated on three Chinese legal datasets, DDLR achieves significant improvements, outperforming baselines by 4.3%-17.0% in charge prediction accuracy and generating court views with higher semantic fidelity.