Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models | ScienceToStartup | ScienceToStartup