Can Atomic Step Decomposition Enhance the Self-structured Reasoning of Multimodal Large Models? (arXiv)
AtomThink introduces four key modules — a data engine, atomic step fine-tuning, policy-guided multi-turn inference, and an atomic capability metric — to bring structured slow-thinking capabilities into visual understanding models, while improving data utilization by 5× and inference efficiency by 85.3% over state-of-the-art structured CoT methods.