Multi-modal multi-agent debate-based dialectical knowledge transfer learning

Transferring knowledge from large to smaller models can enhance their reasoning abilities and expedite deployment. However, for multi-modal complex reasoning tasks, existing large models often produce knowledge with factual errors. Transferring this flawed multi-modal knowledge to smaller models can lead to error accumulation, thereby degrading their performance. To address this challenge, we propose a novel Multi-modal Multi-agent Debate-based Dialectical Knowledge Transfer Learning framework (M2D-DKTL). Initially, we introduce a multi-modal multi-agent debate framework to uncover deep logical relationships and enhance knowledge quality. Then, we propose using chain-of-debate to condense dialectical knowledge and transfer it to smaller models, thereby improving their complex reasoning ability. Experimental results show that debates can enhance the model's understanding and reasoning in complex scenarios. Moreover, our method achieved average improvements of +4.3%, +3.8%, and +2.6% on the MMMU, MathVista, and CMMMU datasets, respectively, validating its effectiveness and generalizability. Further analysis reveals that our method enhances the logical reasoning and deep analysis capabilities of smaller models.