DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models
Zefeng He, Xiaoye Qu, Yafu Li, Tong Zhu, Siyuan Huang, Yu Cheng
2026-01-02
Summary
This paper introduces a new way for AI to solve problems that involve both images and text, focusing on tasks where understanding the visual information is key.
What's the problem?
Current AI models that handle both images and text often rely too heavily on the text part when making decisions, even when the images contain important clues. This makes them struggle with complex tasks that require careful visual reasoning, like planning a sequence of actions based on what's seen in a picture or figuring out how objects fit together in space.
What's the solution?
The researchers developed a system called DiffThinker that treats visual reasoning as a process of *generating* new images. Instead of just describing what it sees, the AI essentially 'thinks' by creating images that represent its reasoning steps. This is done using a technique called diffusion, which is good at creating detailed and consistent images. DiffThinker directly manipulates images to solve problems, leading to more accurate and logical results.
Why it matters?
This research shows that a new approach to multimodal reasoning – one that focuses on generating images rather than just processing text – can significantly outperform existing AI models on vision-centric tasks. It suggests that letting AI 'show its work' visually can lead to more reliable and capable AI systems, especially for problems where visual understanding is crucial, and it beats out even the most advanced closed-source models like GPT-5 and Gemini.
Abstract
While recent Multimodal Large Language Models (MLLMs) have attained significant strides in multimodal reasoning, their reasoning processes remain predominantly text-centric, leading to suboptimal performance in complex long-horizon, vision-centric tasks. In this paper, we establish a novel Generative Multimodal Reasoning paradigm and introduce DiffThinker, a diffusion-based reasoning framework. Conceptually, DiffThinker reformulates multimodal reasoning as a native generative image-to-image task, achieving superior logical consistency and spatial precision in vision-centric tasks. We perform a systematic comparison between DiffThinker and MLLMs, providing the first in-depth investigation into the intrinsic characteristics of this paradigm, revealing four core properties: efficiency, controllability, native parallelism, and collaboration. Extensive experiments across four domains (sequential planning, combinatorial optimization, constraint satisfaction, and spatial configuration) demonstrate that DiffThinker significantly outperforms leading closed source models including GPT-5 (+314.2\%) and Gemini-3-Flash (+111.6\%), as well as the fine-tuned Qwen3-VL-32B baseline (+39.0\%), highlighting generative multimodal reasoning as a promising approach for vision-centric reasoning.