Research map / Vision-language and multimodal / Multimodal LLMs

DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiffThinker, which reformulates multimodal reasoning as native image-to-image generation with a diffusion model (fine-tuned Qwen-Image-Edit) and shows it beats GPT-5, Gemini-3-Flash and fine-tuned Qwen3-VL-32B on vision-centric puzzles.

Problem

Text-centric MLLM reasoning struggles to track evolving visual state over long-horizon, vision-centric tasks and is slow and variable in latency.

Approach

Model multimodal reasoning directly as a generative image-to-image task with a diffusion/flow-matching model, so solutions are produced as images in visual space and parsed to symbolic answers, giving native parallel exploration, fixed inference budget and collaboration with MLLMs.

Key results

• DiffThinker outperforms GPT-5 by +314.2% and Gemini-3-Flash by +111.6% on average across seven tasks.
• DiffThinker outperforms the fine-tuned Qwen3-VL-32B baseline by +39.0% with identical training data.
• DiffThinker inference latency is 1.1s versus 1.4s for Qwen3-VL-32B (SFT) and 1.0s for Qwen3-VL-8B (SFT) on VSP-Super level-16.

Surprising findings

• DiffThinker-Video (Wan2.2-TI2V-5B) yields lower accuracy and higher training cost than the image-based variant on Maze level-8 despite fewer parameters.
• Video-based reasoning is about 2.0s per inference versus 1.1s for the image model.

Limitations

• Out-of-distribution generalization is limited by the zero-shot reasoning ability of the generative base model.
• MLLMs retain an advantage in text-centric domains such as complex math, which the method does not address.
• Video-based variant is costly and currently weaker at reasoning.

Experimental setup

Datasets: VSP VSP-Super Maze TSP Sudoku Jigsaw VisPuzzle COCO

Benchmarks: VSP VSP-Super Maze TSP Sudoku Jigsaw VisPuzzle

Metrics: accuracy training time (h) inference latency (s)

Baselines: GPT-5 Gemini-3-Flash Qwen3-VL-32B (SFT) Qwen3-VL-8B (SFT) Gemini-3-Pro GRPO-trained Qwen3-VL

Builds on: Qwen-Image-Edit flow matching GRPO Thinking with Images classifier-free guidance

Closest papers by meaning

• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.92)
• Thinking Diffusion: Penalize and Guide Visual-Grounded Reasoning in Diffusion Multimodal Language Models (CVPR 2026 · similarity 0.92)
• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.92)
• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.91)
• I Think, Therefore I Diffuse: Enabling Multimodal In-Context Reasoning in Diffusion Models (ICML 2025 · similarity 0.91)
• GGBench: A Geometric Generative Reasoning Benchmark for Unified Multimodal Models (CVPR 2026 · similarity 0.91)