Muhammad Kamran Janjua

LLMs struggle with raw visual tool outputs, but a simple trick—translating pixels into language-native summaries—unlocks massive gains in visual reasoning, outperforming even fine-tuned agents.