Thinking with Images as Continuous Actions: Numerical Visual Chain-of-Thought
Numerical Visual Chain-of-Thought (NV-CoT) allows Multimodal Large Language Models (MLLMs) to reason over images using continuous numerical coordinates instead of discrete tokens. This framework expands the model's action space to a Euclidean space, enabling direct generation of bounding box coordinates as actions. This significantly improves localization precision and final answer accuracy across both supervised fine-tuning and reinforcement learning settings.