Semi-off-Policy Reinforcement Learning for Vision-Language Slow-Thinking Reasoning | Read Paper on Bytez