PaperlayerView exact revision on arXiv

Source-linked AI summary

ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan

arXiv:2610.20781v1cs.CVcs.AI

TL;DR

Unified multimodal models can see conflicts between understanding and generation, and architectural unification alone does not establish mutual gains. ReasonUni adds reasoning to reconstructive post-training and uses TreeGRPO to jointly optimize the two modules; experiments report consistent gains across both capabilities and understanding benchmarks.

  • Problem

    Naively coupling understanding and generation can yield suboptimal results, and architectural unification alone does not establish mutual gains between the tasks.

    Source: §1 · p. 2

  • Method

    ReasonUni reasons about an instructed image edit to describe the target state, then generates the target image, while TreeGRPO jointly optimizes both modules using tree-structured reward comparisons.

    Source: §1 · pp. 2–3

  • Results

    ReasonUni reports consistent gains in understanding and generation across benchmarks, with improvements on all 12 evaluated understanding benchmarks.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    The findings support reasoning-augmented reconstruction as an effective approach to using generation feedback for mutual gains in unified multimodal models.

    Source: §5 · p. 10

Abstract

from arXiv · show

Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation. We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output. To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.

1 INTRODUCTION

Unified multimodal models can lose understanding performance when generation and understanding are optimized together, while prior reconstruction methods provide limited supervision for higher-order visual reasoning. ReasonUni extends reconstructive post-training to reasoning and jointly optimizes understanding and generation; experiments report gains across both capabilities.

1 INTRODUCTION

  • Naively coupling understanding and generation can yield suboptimal results, and architectural unification alone does not establish mutual gains between the tasks.Prior work reports that optimizing diffusion-based generation can impair understanding performance and learned representations, and vice versa.Source: §1 · p. 2
  • Prior reconstruction frameworks recognize what is present but provide limited supervision for reasoning about scene structure, intent, or causal relations.The paper identifies this as a limitation for generalization to complex, open-ended visual reasoning tasks.Source: §1 · p. 2
  • ReasonUni extends reconstructive post-training beyond perception by reasoning about an edit and describing the target state before generating the conditioned target image.The input is an image and editing instruction; the generation module reconstructs the target from the description.Source: §1 · p. 2
  • TreeGRPO samples multiple images for each reasoning path, aggregates their rewards for branch-level understanding updates, and uses within-branch normalization to guide generation.Applying GRPO at both levels jointly optimizes the two modules under a unified objective.Source: §1 · p. 3
  • ReasonUni gains on both understanding and generation across benchmarks, including gains on all 12 evaluated understanding benchmarks.Compared with UAE, it leads by up to 5.98 points on DPGBench, 13.00 on WISE, and 16.78 on GenEval++; logical reasoning on MMBench gains 6.77 points over X-Omni versus 3.38 for UAE.Source: §1 · p. 3

2 METHOD

ReasonUni casts reconstruction as instruction-driven editing: the model describes the target state before generating it. TreeGRPO organizes multiple reasoning paths and image candidates to jointly optimize understanding and generation.

2.1 PRELIMINARIES: UNIFIED DISCRETE AUTOREGRESSIVE MODELING

  • A shared visual and text token space supports understanding and generation policies in one unified discrete autoregressive model.The model uses standard next-token autoregressive likelihood; visual inputs and outputs share a semantic codebook.Source: §2.1 · p. 3

2.2 REASONUNI: REASONING-AUGMENTED RECONSTRUCTIVE FRAMEWORK FOR UMMS

  • ReasonUni addresses perception-centric reconstruction by making the model interpret an edit instruction and describe the desired target state before image generation.The target-state description integrates the intended transformation with source-scene content that should remain unchanged; generation is conditioned on that description.Source: §2.2 · pp. 3–4

2.3 TREEGRPO: TREE-BASED JOINT GROUP RELATIVE POLICY OPTIMIZATION

  • Branch-average rewards supervise understanding, while within-branch image comparisons guide generation under a fixed target-state description.Two-level normalization compares each module against alternatives at its own level of the hierarchy.Source: §2.3 · pp. 4–5
  • TreeGRPO samples K reasoning paths and M edited-image candidates per path, separating uncertainty about the reasoning path from image rendering quality.The resulting tree has K branches and K × M leaves, with multiple images per path providing a robust estimate of path quality.Source: §2.3 · p. 4
  • TreeGRPO jointly optimizes both modules under one objective, using rewards computed directly in X-Omni’s semantically aligned visual token space.Unlike UAE, which uses CLIP features of decoded images, this latent reward uses mean-pooled codebook embeddings and cosine similarity to the target image.Source: §2.3 · pp. 4–5

3 EXPERIMENTS

Experiments test whether ReasonUni improves understanding and generation jointly across initialization settings and benchmark families, with fine-grained analyses, ablations, and qualitative probes. Results show gains across both capabilities, while the analyses examine reasoning, generation control, reward efficiency, sampling, and editing behavior.

3.1 MAIN QUANTITATIVE RESULTS

  • Under SFT, ReasonUni gains 13.34 points on DPGBench, 13.00 on WISE, and 15.71 on GenEval++ over X-Omni; under RL, it gains 1.65 on MMBench, 2.00 on MMVP, 1.75 on CV-Bench, and 6.79 on GenEval++.In the matched-data UAE† control, the advantage holds across all seven metrics.Source: §3.1 · p. 6
  • ReasonUni improves over X-Omni on every metric across 12 understanding metrics and six benchmark families, with largest gains on MMMU (+2.67), MMVP (+2.00), and CV-Bench (+1.75).Subsequent experiments use the RL setting.Source: §3.1 · p. 6, Table 2 · p. 6

3.2 FINE-GRAINED QUANTITATIVE ANALYSIS

  • On MMBench, logical reasoning improves from 48.31 to 55.08 (+6.77) over X-Omni, while comparisons with UAE show perception remains comparable and reasoning gains persist.ReasonUni improves all six dimensions; compared with UAE, overall MMBench rises from 75.43 to 75.96.Source: §3.2 · p. 7, Table 3 · p. 7
  • GenEval++ reaches 46.79 overall, +6.79 over X-Omni-RL, with the largest gains on Pos/Size (+25.0) and Color (+17.5).Pos/Count, Color/Position, and Multi-Count each improve by +2.5.Source: §3.2 · p. 7, Table 4 · p. 7

3.3 ABLATION STUDY

  • Latent rewards achieve the highest Und. Avg (54.26) and Gen. Avg (47.40) among compared rewards while reducing training time by about 60% (25h vs. 58–68h on 8×H100).Rewards are computed in SigLIP-VQ token space without diffusion decoding or external re-encoding; Figure 4 reports the same preference ordering as SigLIP2 and DINOv3.Source: §3.3 · p. 7, Table 5 · p. 8, Figure 4 · p. 8
  • With 16 candidate images, TreeGRPO’s K=4, M=4 configuration raises Und. Avg from 53.90 to 54.26 and Gen. Avg from 45.40 to 47.40 versus flat GRPO, while reducing training time from 30h to 25h.Sharing each description across four images requires four target-state descriptions rather than sixteen; K=8, M=8 further raises Gen. Avg to 49.02.Source: §3.3 · pp. 7–8, Table 5 · p. 8

3.4 QUALITATIVE CASE STUDY

  • Under the same vehicle text-removal instruction, ReasonUni removes the largest front text while preserving the license plate, unlike X-Omni.Both models receive the same prompt and instruction.Source: §3.4 · p. 8
  • Later TreeGRPO checkpoints show more boot wear and remove the central rider and vehicle, alongside rising reconstruction reward and longer target-state descriptions.The reconstruction reward rises rapidly early in training and gradually levels off.Source: §3.4 · p. 8

4 RELATED WORK

Related work covers unified multimodal architectures and post-training approaches that connect understanding with generation. Existing reconstruction methods couple the tasks through extracted semantics, but their supervision largely remains perception-level rather than explicitly addressing visual reasoning.

Architecture Design of Unified Multimodal Models

  • UMMs use hybrid AR–diffusion or fully autoregressive designs, with some fully AR models sharing semantically aligned visual tokens across understanding and generation.The described hybrid designs include dedicated generative components or adapters to external image generators.Source: §4 · pp. 8–9

Post-Training in Unified Multimodal Models

  • Generation-oriented post-training methods improve generation or editing, while T2I-R1 reasons over text prompts; ReasonUni instead reasons over visual inputs and uses generation feedback for visual reasoning.Source: §4 · p. 10
  • Reconstruction approaches condition image reconstruction on extracted semantics, but their supervision remains largely perception-level and offers limited explicit visual-reasoning supervision.The cited work frames this limitation as constraining generalization to tasks requiring reasoning beyond perception.Source: §4 · p. 10

5 CONCLUSION

ReasonUni extends reconstructive post-training to visual reasoning, and TreeGRPO jointly optimizes understanding and generation through tree-structured reward comparisons. Across initialization settings, it improves generation benchmarks and all 12 evaluated understanding benchmarks, supporting reasoning-augmented reconstruction as an approach to mutual gains in UMMs.

5 CONCLUSION

  • ReasonUni extends reconstructive post-training to visual reasoning, with TreeGRPO jointly optimizing understanding and generation through tree-structured reward comparisons.Source: §5 · p. 10
  • Across both initialization settings, ReasonUni gains on generation benchmarks and improves all 12 evaluated understanding benchmarks, including logical reasoning on MMBench.Source: §5 · p. 10
  • The findings support reasoning-augmented reconstruction as an effective approach to using generation feedback for mutual gains in unified multimodal models.Source: §5 · p. 10

A DATASET SYNTHESIS AND SAMPLES

The dataset pipeline creates reasoning-driven editing examples from image contexts, with instructions that require relational, logical, or commonsense identification rather than direct visual description. It pairs these instructions with explicit target-state descriptions and edited images to provide challenging, relatively deterministic training examples.

A DATASET SYNTHESIS AND SAMPLES

  • The synthesis pipeline uses image-conversation pairs as context and produces tuples containing the source image, conversation history, synthesized instruction, and edited image.Figure 8 depicts the context and resulting source/edit pair, with the instruction grounded in the dialogue.Source: A DATASET SYNTHESIS AND SAMPLES · p. 16, Figure 8 · p. 16

A.1 GENERATION OF REASONING INSTRUCTIONS

  • Gemini-3-Pro mines object relationships, functional attributes, spatial distributions, and physical laws from LLaVA-NeXT-Data contexts to support task construction.Source: §A.1 · p. 16
  • Instructions identify targets through functional or relational context rather than direct descriptions, spanning twelve reasoning categories including spatial relations, quantitative comparison, causal prediction, and physical attributes.The pipeline also generates an explicit reference effect describing the post-edit state to reduce ambiguity in the reconstruction target.Source: §A.1 · pp. 16–17

A.2 INSTRUCTION-GUIDED IMAGE EDITING AND EXECUTION

  • The editing stage uses the reasoning instruction and reference description to generate the target image, aiming to preserve plausible background completion and consistent lighting.The authors plan to open-source 30,000 synthesized samples as a training and evaluation resource.Source: §A.2 · p. 17

B.1 BASELINES AND ABLATION STUDY IMPLEMENTATION DETAILS

The implementation uses X-Omni checkpoints and a mixed editing and reasoning-data training setup. Baseline and ablation experiments compare UAE with ReasonUni and reconstruction-only TreeGRPO with the full reasoning-augmented method.

Base Model and Main Experiments

  • Main experiments use X-Omni, initialized from its official SFT and RL checkpoints.X-Omni is a pure autoregressive model that represents images with semantically aligned SigLIP-VQ visual tokens.Source: §B.1 · p. 17

Training Data and Configuration

  • Training uses one epoch of mixed editing and synthetic reasoning-driven data, with TreeGRPO sampling 4 reasoning paths and 4 images per path.The setup uses 8 H100 GPUs and a peak learning rate of 1 × 10−5.Source: §B.1 · p. 17

Baseline Implementation

  • UAE is re-implemented on X-Omni for comparisons, with a matched-data control using the same initialization, samples, training steps, and optimization hyperparameters as ReasonUni.Unmarked UAE results use its official training data.Source: §B.1 · pp. 17–18

Reconstruction-only Ablation

  • The reconstruction-only ablation retains TreeGRPO but reconstructs the source image rather than an edited target state.Its understanding module describes the source image without an editing instruction.Source: §B.1 · p. 18

B.2 EVALUATION DETAILS

Visual understanding is evaluated with VLMEvalKit and fixed deterministic inference settings. Image-generation evaluation follows the official X-Omni configuration, with designated pipelines for three benchmarks.

Visual Understanding Evaluation

  • Visual understanding evaluation uses VLMEvalKit, with GPT-4o for external API operations and inference temperature set to 0.Inference implementation and system prompts follow the official X-Omni codebase.Source: §B.2 · p. 18

Image Generation Evaluation

  • Image generation uses the official X-Omni inference configuration at 1152 × 1152 resolution, with temperature 1.0 and CFG scale 1.0.Sampling also uses min-p 0.03 and top-p 1.0.Source: §B.2 · p. 18
  • DPG-Bench, WISE, and GenEval++ are evaluated with their official pipelines and designated scoring models.The benchmarks are cited as Hu et al. (2024), Niu et al. (2025b), and Ye et al. (2025), respectively.Source: §B.2 · p. 18

C QUALITATIVE ANALYSIS AND VISUALIZATIONS

Qualitative examples show ReasonUni’s edited descriptions and reconstructions matching references more closely while preserving non-target content. Additional comparisons highlight compositional prompt following and knowledge-focused generation.

Figure 10

Figure 9

  • Across floral, maritime, and street scenes, ReasonUni better matches edited references and introduces fewer unrelated changes.Its target-state descriptions also better match the displayed references, with more consistent preservation of non-target content.Source: Figure 9 · p. 19, C QUALITATIVE ANALYSIS AND… · p. 19

Figure 11

  • For knowledge-focused prompts, ReasonUni generates more semantically accurate concepts than the baseline across several knowledge categories.Examples include animal recognition, instrument category, art-related knowledge, and historical technology.Source: Figure 11 · p. 21

D EXTENDED RESULTS AND ABLATION STUDIES

Extended results describe stronger fine-grained prompt following and improvements in compositional faithfulness, knowledge-grounded generation, and position/count control. The section also compares full reasoning augmentation with reconstruction-only TreeGRPO across training settings.

Extended quantitative results

  • ReasonUni shows stronger DPGBench prompt following than X-Omni on overall score and most sub-metrics, including fine-grained prompt measures.Table 6 breaks performance down into Global, Entity, Attribute, Relation, and Other metrics.Source: D EXTENDED RESULTS AND ABLATION STUDIES · p. 22, Table 6 · p. 22
  • Generation results highlight improved compositional faithfulness, knowledge-grounded generation, and position/count control across DPGBench, WISE, and GenEval++.The section reports these results in Table 9.Source: D EXTENDED RESULTS AND ABLATION STUDIES · p. 22

Extended results and ablation studies

  • The multimodal understanding comparison reports the largest gains in vision-centric evaluation while maintaining stable hallucination and OCR behavior.Table 7 compares reconstruction-only TreeGRPO with the full reasoning-augmented framework across SFT and RL settings for understanding and generation.Source: D EXTENDED RESULTS AND ABLATION STUDIES · p. 22, Table 7 · p. 22

D.1 WISE RESULTS AND ANALYSIS

WISE results show gains in both SFT and RL settings, including improvements across all six evaluated domains in RL. The prompt templates frame editing as a reasoning task and specify how to describe and generate the target state.

D.1 WISE RESULTS AND ANALYSIS

  • WISE rises from 24 to 37 in SFT and from 45 to 48 in RL; in RL, ReasonUni improves over X-Omni-RL in all six domains.Table 10 reports the six domains as Cultural, Time, Space, Biology, Physics, and Chemistry.Source: §D.1 · p. 22, Table 10 · p. 23
  • The largest gains on Space, Physics, and Chemistry accompany stronger generation of intended concepts rather than plausible substitutes.The examples span animal taxonomy, instruments, art, and historical technology; the authors connect the domain gains to knowledge-grounded compositional generation.Source: §D.1 · p. 22

E PROMPT TEMPLATES

  • The templates pose editing as an implicit-target reasoning challenge, using spatial relations, common sense, and other reasoning types to identify what to change.Instructions are meant to point to a target through relationships or function rather than explicitly naming its appearance.Source: E PROMPT TEMPLATES · p. 25
  • Editing instructions should be self-contained, explicit, plausible, and deterministic, while target descriptions specify the edited scene for image generation.The description prompt asks for scene layout, entities, environmental details, and any visible text, and directs the model to output the edited result description rather than its reasoning process.Source: E PROMPT TEMPLATES · pp. 24–25

Related papers

Loading 2610.20781v1…