A new AI benchmark called SOOHAK reveals that while models like Google's Gemini 3 Pro excel at solving certain research-level math problems, they struggle to recognize unsolvable tasks. This highlights a significant gap in AI's problem-solving capabilities despite advancements in computational power.