A new framework called PTD-PO enhances the reasoning capabilities of Large Vision-Language Models by providing dense guidance without revealing answers, addressing inefficiencies in multimodal reasoning tasks. Experiments show that PTD-PO outperforms existing methods, suggesting a significant advancement in AI training techniques.