Mirage: Multimodal Reasoning in VLMs Without Rendering Images

The article highlights a significant advancement in multimodal reasoning with the introduction of Mirage, a model that enables visual reasoning in vision-language models (VLMs) without relying on image rendering or generation. Traditional VLMs excel at understanding text and images but struggle with tasks requiring visual thinking, such as spatial puzzles, because they primarily depend on textual descriptions during reasoning processes. Existing approaches that generate images or incorporate visual annotations often face limitations like increased computational costs and weakened reasoning capabilities, underscoring the need for more efficient solutions. Mirage addresses this challenge by embedding visual reasoning directly within the model's internal representations