DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

TL;DR

DiffThinker introduces Generative Multimodal Reasoning by recasting multimodal tasks as image-to-image generation with diffusion models, delivering superior logical consistency and spatial precision for vision-centric, long-horizon reasoning. It identifies four core advantages—efficiency, controllability, native parallelism, and collaboration—and benchmarks seven tasks across four domains, consistently surpassing state-of-the-art closed-source models (+314.2% vs GPT-5, +111.6% vs Gemini-3-Flash, +39.0% vs a fine-tuned Qwen3-VL). An explicit comparison with MLLMs highlights the benefits of visual generative reasoning, while ablation studies reveal optimal inference budgets (≈20 steps) and the impact of data scale and CFG guidance. The work also explores a DiffThinker-Video variant, showing both the promise and current limitations of video-based reasoning, and argues for future integration with text-centric models to broaden multimodal reasoning capabilities.

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.

Figures (36)

• Figure 1: (a) Quantitative results across seven tasks. (b) DiffThinker produces solution images directly, whereas baseline results are post-processed visualizations of textual outputs with errors highlighted. By reformulating reasoning as a native image-to-image generative task, DiffThinker achieves superior logical consistency and spatial precision in complex long-horizon, vision-centric reasoning tasks.
• Figure 2: Overview of different multimodal reasoning paradigms. (a) Standard MLLMs map inputs directly to symbolic solutions. (e.g., 'R' and 'D' representing 'Right' and 'Down' actions) (b) "Thinking with Images" MLLMs interact with multimodal inputs through iterative tool calls. (c) DiffThinker reformulates multimodal reasoning as a direct generative image-to-image task, where solutions are produced in visual space and then parsed to symbolic solutions to ensure a fair comparison.
• Figure 3: Computational efficiency analysis. (a) compares training duration in hours, and (b) shows inference latency in seconds.
• Figure 4: DiffThinker as a collaborative partner. (a) The partnership framework where DiffThinker generates $N$ candidates for MLLM verification. (b) Performance on Jigsaw level-4, demonstrating that collaboration surpasses individual models and accuracy further scales with the number of candidates $N$.
• Figure 5: Trade-off between accuracy and inference time across varying inference steps. The horizontal axis denotes the number of inference steps, while the vertical axis denotes accuracy or inference time. An optimal balance between reasoning performance and computational cost is achieved at approximately 20 steps.