Beyond Dominant Patches: Spatial Credit Redistribution For Grounded Vision-Language Models | ScienceToStartup | ScienceToStartup