SOTA models for multimodal reasoning, fine-tuned with VPPO. Achieves superior performance by focusing on critical visual tokens.