Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It

Look Light, Think Heavy: What Multimodal Chain-of-Thought Reasoning Can and Cannot Do