Paper 2512.24165
DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models
- Published
- Dec 2025
- Research lab
- Independent
- Citations
- 12
- GitHub
- 185 stars
01 In brief
Summary
DiffThinker is a diffusion-based framework that introduces Generative Multimodal Reasoning, reformulating multimodal reasoning as a native image-to-image generative task rather than text-centric symbolic mapping.
Built on Qwen-Image-Edit with Flow Matching and a Multimodal Diffusion Transformer, it directly produces solution images that are parsed into symbolic outputs for evaluation.
Across seven tasks in four domains (sequential planning, combinatorial optimization, constraint satisfaction, spatial configuration), DiffThinker outperforms GPT-5 by 314.2%, Gemini-3-Flash by 111.6%, and a fine-tuned Qwen3-VL-32B baseline by 39.0%.
The paper identifies four core properties: efficiency (training and inference comparable to SFT baselines), controllability (fixed-step generation), native parallelism (explores multiple candidate paths simultaneously), and collaboration (partners with MLLMs to surpass either alone).
Ablations show optimal performance at 20 inference steps, 30,000 training samples, and a classifier-free guidance scale of 4.
A video-based variant, DiffThinker-Video, shows lower accuracy and higher latency than the image-based approach.
02 From the paper
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.