Why do AI models hit a wall in visual reasoning? Researchers from Zhejiang University, Hunan University, and Xiaomi introduce Visual Para-Thinker. It uses parallel divide-and-conquer reasoning to avoid sequential thinking plateaus. Outperforms on V*, CountBench, RefCOCO, and

→ View original post on X — @jiqizhixin