Search for a command to run...
GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training