💡 This paper proposes ARM-Thinker, an agentic multimodal reward model that leverages external tool use and multi-stage reinforcement learning to jointly optimize tool-calling decisions and scoring accur