PaperlayerView exact revision on arXiv

Source-linked AI summary

DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models

Zefeng He, Xiaoye Qu, Yafu Li, Tong Zhu, Siyuan Huang, Yu Cheng

arXiv:2512.24165v1cs.CV

TL;DR

MLLM reasoning remains predominantly text-centric, limiting performance on complex long-horizon, vision-centric tasks. DiffThinker reformulates multimodal reasoning from symbolic mapping into a native generative image-to-image process using diffusion models. Across seven tasks, DiffThinker significantly outperforms SOTA MLLMs, including GPT-5 (+314.2%) and Gemini-3-Flash (+111.6%).

  • Problem

    MLLM reasoning remains predominantly text-centric, limiting performance on complex long-horizon, vision-centric tasks.

  • Method

    DiffThinker reformulates multimodal reasoning from symbolic mapping into a native generative image-to-image process using diffusion models.

  • Results

    Across seven tasks, DiffThinker significantly outperforms SOTA MLLMs, including GPT-5 (+314.2%) and Gemini-3-Flash (+111.6%).

  • Takeaways & Limitations

    Generative multimodal reasoning is presented as a promising approach for vision-centric reasoning, with efficiency, controllability, native parallelism, and collaboration as core properties.

  • Takeaways & Limitations

    Out-of-distribution generalization is constrained by the limited zero-shot reasoning proficiency of current generative foundation models.

Abstract

from arXiv · show

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.

1. Introduction

Current MLLMs advance multimodal reasoning but remain text-centric, creating inefficiency and difficulty tracking changing visual states in complex long-horizon tasks. DiffThinker addresses this by shifting reasoning into visual space through generative image-to-image modeling and shows strong results across diverse benchmarks.

  • 1. Introduction: MLLMs’ lengthy Chain-of-Thought and iterative visual interactions produce uncontrollable generation, high latency, and difficulty tracking visual state over long sequences.These limitations particularly affect complex, long-horizon, vision-centric tasks.
  • 1. Introduction: The paper identifies four properties of generative multimodal reasoning: efficiency, controllability, native parallelism, and collaboration.Collaboration allows DiffThinker to partner with MLLMs, while parallel generation explores and prunes multiple candidate solutions.
  • 1. Introduction: DiffThinker reformulates multimodal reasoning from text-centric symbolic mapping into a native image-to-image generative process in visual space.The paradigm is designed to improve reasoning for vision-centric tasks.
  • 1. Introduction: DiffThinker’s evaluation systematically compares the new paradigm with MLLMs across multiple domains and investigates its intrinsic characteristics.The experiments are presented as a broad assessment rather than a single-task demonstration.
  • 1. Introduction: +314.2% versus GPT-5, +111.6% versus Gemini-3-Flash, and +39.0% versus fine-tuned Qwen3-VL-32B demonstrate DiffThinker’s benchmark advantage.Experiments span seven tasks across sequential planning, combinatorial optimization, constraint satisfaction, and spatial configuration.

2. Related Works

Prior multimodal reasoning remains largely text-centric, while image and video interaction paradigms add iterative tool use or high computational cost. DiffThinker instead uses image generation as a unified and more efficient visual reasoning process.

  • 2. Related Works: Existing multimodal reasoning paradigms are predominantly text-centric, while Thinking with Image extends them through multi-turn visual interactions.The related-work discussion frames DiffThinker as advancing beyond symbolic reasoning and iterative tool invocation.
  • 2. Related Works: Thinking with Video enables multi-turn interaction with video, but recent video-generation reasoning studies often benchmark closed-source models with undisclosed internal reasoning mechanisms.The passage also notes that video generation carries prohibitive computational costs.
  • 2. Related Works: DiffThinker diverges from video-generation approaches by establishing image generation as a more efficient paradigm for multimodal reasoning.The comparison concerns computational efficiency rather than a claim that image generation dominates every reasoning setting.
  • 2. Related Works: Diffusion research provides the generative-modeling foundation underlying DiffThinker’s visual reasoning formulation.The related work situates the framework within diffusion, flow-based, latent-diffusion, and diffusion-transformer developments.
  • 2. Related Works: DiffThinker formulates diverse multimodal reasoning tasks as a unified generative process in visual space, enabling rapid adaptation without task-specific architectures or training from scratch.This positions visual generation as the central alternative to prior specialized approaches.

3. Generative Multimodal Reasoning

DiffThinker reformulates multimodal reasoning as a direct multimodal-to-image generation process, replacing text-centric symbolic reasoning with visual reasoning. Its flow-matching pipeline generates and parses visual solutions across diverse long-horizon, vision-centric tasks.

  • 3.1. Problem Reformulation: Standard MLLMs represent reasoning as multimodal-to-text mapping, while iterative image-based variants interleave reasoning, tool calls, and visual observations.These multi-turn loops create computational overhead when scaling to complex long-horizon, vision-centric tasks.
  • 3.1. Problem Reformulation: DiffThinker directly transforms multimodal inputs into visual solution images instead of sequential textual reasoning traces.The solution image visually encapsulates the reasoning trajectory and solution, then a parsing function maps it back to symbolic space for evaluation.
  • 3.2. Flow Matching: DiffThinker trains a multimodal diffusion model with flow matching to transform noise into task-conditioned solution latents.The conditioning latent comes from an MLLM encoding the user’s text and visual inputs; the model predicts the target velocity field with mean squared error.
  • 3.2. Flow Matching: DiffThinker solves the learned ordinary differential equation from initial noise and decodes the resulting latent into a visual solution after T Euler steps.The final latent approximates the data latent and is decoded through the VAE decoder into pixel space.
  • 3.3. Benchmark Tasks: The evaluation covers seven tasks across sequential planning, combinatorial optimization, constraint satisfaction, and spatial configuration with scalable difficulty and structurally parseable outputs.Examples include grid navigation, maze routing, TSP, Sudoku, jigsaw reconstruction, and VisPuzzle; DiffThinker renders trajectories, paths, completed grids, or reconstructed images while MLLMs produce text plans or indices.
  • 3.4. Native Parallel Reasoning: Figure 4 depicts native parallel reasoning, where multiple candidate paths are explored early and iteratively refined into one valid trajectory.This contrasts with sequential MLLM reasoning that may require explicit reflection or backtracking to correct early errors.

4. Experiments

DiffThinker achieves strong performance across diverse vision-centric reasoning tasks while offering efficient, controllable, parallel, and collaborative inference. Ablations show useful trade-offs involving inference steps, training data, guidance scale, and video generation.

  • 4.1. Main Results: DiffThinker achieves state-of-the-art performance across seven challenging tasks in four domains, surpassing GPT-5 by 314.2%, Gemini-3-Flash by 111.6%, and fine-tuned Qwen3-VL-32B by 39.0%.The main results use DiffThinker++ based on Qwen-Image-Edit-2511, while subsequent analyses use DiffThinker.
  • 4.1. Main Results: As task complexity increases, MLLM performance decays rapidly, whereas DiffThinker maintains high accuracy across sequential planning, spatial configuration, optimization, and constraint-satisfaction tasks.DiffThinker reaches near-perfect performance on Jigsaw and VisPuzzle and performs exceptionally on TSP and Sudoku.
  • 4.2. Discussion and Observation: DiffThinker delivers 1.1s inference latency, comparable to Qwen3-VL-8B’s 1.0s and with training duration nearly identical to Qwen3-VL-32B SFT.Its training duration is substantially lower than GRPO overhead.
  • 4.2. Discussion and Observation: Fixed-step Euler integration makes DiffThinker’s computational budget deterministic, avoiding the fluctuating latency associated with verbose or collapsed autoregressive MLLM outputs.The method is described as preserving execution stability and reliable solutions under controllable generation.
  • 4.2. Discussion and Observation: DiffThinker generates multiple candidate images for MLLM verification, and their collaboration surpasses either model alone on Jigsaw level-4.DiffThinker supplies visual imagination while the MLLM filters candidate errors through reflection; accuracy scales with candidate count N.
  • 4.3. Ablation Studies: Accuracy remains high with 10 inference steps, improves noticeably at 20 steps, and then plateaus; training performance scales consistently with expanded data.With limited data, the model initially prioritizes task-specific rendering syntax such as grid alignment and trajectory continuity.
  • 4.3. Ablation Studies: Classifier-free guidance peaks at w = 4 across most levels, while w = 1 yields weak trajectories and w = 7 causes oversaturation and visual artifacts.The authors adopt w = 4 as the default configuration.
  • 4.4. Image Generation vs. Video Generation: DiffThinker-Video shows visual maze reasoning but achieves lower accuracy, higher training overhead, and 2.0s inference latency versus DiffThinker’s 1.1s.The comparison highlights the computational cost of video generation despite the video variant’s smaller parameter count.

5. Conclusion

The paper presents DiffThinker as a diffusion-based generative multimodal reasoning paradigm for complex vision-centric tasks. Its image-to-image formulation achieves strong results while exhibiting efficiency, controllability, native parallelism, and collaboration with MLLMs.

  • 5. Conclusion: DiffThinker reformulates multimodal reasoning from text-centric symbolic mapping into a native generative image-to-image process in visual space.The paper associates this formulation with logical consistency and spatial precision.
  • 5. Conclusion: Across four domains, DiffThinker significantly outperforms state-of-the-art MLLMs and demonstrates efficiency, controllability, and native parallelism.The paper also presents DiffThinker as a collaborative backend for augmenting MLLMs.

A.1.1. DATA PREPARATION.

The study prepares task-specific datasets and training configurations for independent models across several multimodal reasoning tasks. It uses deduplicated data and LoRA-based fine-tuning procedures.

  • A.1.1. DATA PREPARATION: Five independent models are specialized for VSP/VSP-Super, Maze, TSP, Sudoku, and Jigsaw, then evaluated on their respective test benchmarks.COCO is used to synthesize training and testing samples for jigsaw puzzles, and all training datasets undergo deduplication.
  • A.1.1. DATA PREPARATION: The study summarizes Flow Matching, SFT, and GRPO configurations and applies LoRA to Qwen-Image-Edit fine-tuning and Qwen3-VL SFT.GRPO is limited to one epoch and fewer rollouts to control computational overhead.

A.1.3. REWARD FUNCTIONS FOR GRPO

The GRPO setup uses task-specific partial rewards rather than sparse exact-match rewards, tailoring feedback to navigation, optimization, Sudoku, and spatial configuration tasks.

  • A.1.3. REWARD FUNCTIONS FOR GRPO: Task-specific partial rewards address sparse binary signals caused by low zero-shot accuracy on complex reasoning tasks.The reward functions are designed to provide more informative optimization feedback.
  • A.1.3. REWARD FUNCTIONS FOR GRPO: Navigation rewards measure the longest continuous prefix of correct actions, encouraging trajectories to become correct incrementally from the starting point.The reward uses the maximum k for which all actions through k match the ground truth.
  • A.1.3. REWARD FUNCTIONS FOR GRPO: TSP rewards first require the predicted and ground-truth coordinate sets to match, then reward matching total path length within ϵ = 1 × 10^-4.This tiered structure prioritizes identifying all required cities before optimizing visitation order.
  • A.1.3. REWARD FUNCTIONS FOR GRPO: Sudoku rewards compute the proportion of correctly filled cells only when the extracted prediction contains the required 81 digits.The formulation compares each predicted digit with its corresponding ground-truth cell.
  • A.1.3. REWARD FUNCTIONS FOR GRPO: Jigsaw rewards measure the proportion of image patches placed at their correct absolute positions when the predicted sequence contains all n patches.Predicted and ground-truth sequences are normalized by removing extraneous whitespace before matching.

B. Limitations and Future Work

DiffThinker performs strongly on targeted vision-centric reasoning, but its out-of-distribution generalization remains constrained by the generative foundation model beneath it. The paper also identifies complementary strengths between DiffThinker and MLLMs, motivating future collaboration.

  • B. Limitations and Future Work: DiffThinker’s out-of-distribution generalization remains constrained by the limited zero-shot reasoning proficiency of current generative foundation models.Because reasoning is directly modeled as generation, performance on unseen complex scenarios depends heavily on the representational depth of the underlying base model.
  • B. Limitations and Future Work: The paper focuses primarily on vision-centric challenges, while MLLMs retain a clear advantage in text-centric domains such as complex mathematical problems.
  • B. Limitations and Future Work: Future work should investigate collaboration between DiffThinker’s visual precision and MLLMs’ linguistic and symbolic capabilities.The authors frame the paradigms as complementary rather than mutually exclusive.

C. Qualitative Analysis

The qualitative analysis compares DiffThinker and Gemini-3-Pro through task-specific success and failure cases, highlighting where each system succeeds or encounters distinct reasoning limitations. Cases span VSP, VSP-Super, Maze, TSP, Sudoku, Jigsaw, and VisPuzzle.

  • C. Qualitative Analysis: The analysis evaluates Gemini-3-Pro on the same problem instances where DiffThinker succeeded, using paired success and failure cases across seven tasks.The comparison includes reasoning processes and measured reasoning duration for Gemini-3-Pro.
  • C. Qualitative Analysis: DiffThinker failures include excessive parallel reasoning that loses a unique trajectory in VSP and an obstructed near-correct path in VSP-Super.
  • C. Qualitative Analysis: The figures also document successful DiffThinker cases for VSP, VSP-Super, Maze, TSP, Sudoku, Jigsaw, and VisPuzzle, alongside a successful Gemini-3-Pro case for VisPuzzle.
  • C. Qualitative Analysis: On Maze, DiffThinker may fail to sustain deep reasoning over long start-to-goal distances, producing only a preliminary trajectory.
  • C. Qualitative Analysis: On TSP, DiffThinker can find a feasible closed loop without finding the shortest path, while Gemini-3-Pro provides a correct solution in the shown case.
  • C. Qualitative Analysis: DiffThinker’s Sudoku and Jigsaw failures show that mostly completed outputs can still contain errors, especially when randomly generated regions are difficult to distinguish.The Jigsaw case is globally reasonable but incorrect in fine details; Gemini-3-Pro succeeds in the shown Sudoku and Jigsaw cases.
Loading 2512.24165v1…