Recently, there are only a few studies exploring cross-domain constituency parsing (CDCP), which usually rely on single LLM and use unsupervised or semi-supervised learning methods. This not only requires a lot of training resources and time, but also completely ignores the respective advantages and complementarity brought by multiple homologous models. Therefore, this paper proposes a multi-LLM debate framework (MLD), which needs only a small amount of source domain data for fine-tuning, and adopts a multi-round dialogue strategy to encourage multiple LLMs to debate, thereby helping them identify and fill in gaps to produce accurate and consistent CDCP results. Systematic experiments and analyses have verified the effectiveness of our method.