Thinking with Drafting: Optical Decompression via Logical Reconstruction

Jingxuan Wei, Honghao He, Caijun Jia +9

Existing multimodal large language models have achieved high-fidelity visual perception and exploratory visual generation. However, a precision paradox persists in complex reasonin…