A collaborative effort from Southeast University, The Chinese University of Hong Kong, and Ant Group has unveiled the innovative lmm-r1 framework. This groundbreaking approach employs a two-stage training strategy to dramatically bolster the mathematical reasoning capabilities of smaller multimodal large models. Remarkably, this enhancement does not necessitate expensive multimodal labeled data, requiring only a modest GPU cost equivalent to 240 yuan. Experimental results have demonstrated that the lmm-r1-enhanced QwenVL-2.5-3B model surpasses larger models like GPT-4 in tackling complex path planning tasks. Since its open-source release, the project has garnered significant attention from the academic community and effectively established a self-sustaining technical ecosystem.
