ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

Feize Wu; Zhiyuan Yan; Yuhuan Hong; Yuwei Niu; Zongjian Li; Zheyuan Liu; Bin Lin; Jinbo Yan

cs.CVOct 14, 2026

ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

Authors: Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan,

Organizations: Peking University · Meituan · Tsinghua University · Sun Yat-sen University

Abstract

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.

Figures & tables

Figure 2: Overview of the ReasonUni architecture and method (left). Given an input image and reasoning-required editing instruction, the UMM infers and describes the target state, then generates candidate edited images conditioned on these descriptions. TreeGRPO formulates this process as a hierarchical branch–leaf structure: reasoning paths form branches, generated images form leaves, and visual rewards are aggregated and propagated to optimize understanding and generation jointly (right).

Figure 3: Flat vs. tree-structured sampling. Left (Flat): each reasoning path and its generated image form an independent sample. Right (Tree): each of

K

reasoning paths branches into

M

generated images, enabling group-relative comparisons across branch-level mean rewards and among image rewards within each branch. Even when a branch has a low mean reward, images scoring above that branch mean receive positive generation advantages.

Method	Understanding				Generation
Method	MMBench	MMVP	CV-Bench	MathVerse	DPGBench	WISE	GenEval ++
X-Omni-SFT	74.40	48.67	67.98	17.01	65.25	24.00	7.50
+ UAE	74.40	50.67	67.06	17.51	72.61	24.00	6.43
+ UAE †	75.26	51.33	66.87	17.64	65.49	26.00	6.43
+ Ours	75.60	52.67	68.85	17.77	78.59	37.00	23.21
Gain ( $\uparrow$ )	+1.20	+4.00	+0.87	+0.76	+13.34	+13.00	+15.71

Table 1: Quantitative comparison on multimodal understanding and generation. We benchmark our method against X-Omni and UAE Yan et al. (2025a) under the SFT and RL initialization settings. The gray

\dagger

row uses the same training data as ReasonUni, while the unmarked UAE rows use UAE’s original training data. Bold denotes the best results. Gain indicates the absolute improvement over the corresponding X-Omni baseline.

Method	General				Vision-Centric			Hallucination		OCR	Knowledge	Math
Method	MMB	SEED	MMStar	MMT	VStar	MMVP	CVB.	POPE	Hallu.	OCRB.	MMMU	MathV.
Show-o Xie et al. (2024)	47.9	–	–	–	–	50.0	–	73.8	–	–	25.1	–
Emu3 Wang et al. (2024b)	58.5	68.2	–	–	51.8	36.6	–	85.2	–	68.7	31.6	–
VILA-U Wu et al. (2024b)	66.6	57.1	–	–	38.7	22.0	–	85.8	–	–	32.2	–
Janus Pro Chen et al. (2025c)	79.2	72.1	–	–	–	–	–	87.4	–	59.5	41.0	–
MetaMorph Tong et al. (2025b)	75.2	71.8	–	–	44.0	48.3	–	–	–	–	41.8	–

Table 2: Comparison with state-of-the-art unified models. We conduct a comprehensive evaluation across six domains, including General (MMBench Liu et al. (2023) , SEED-Bench Li et al. (2023a) , MMStar Chen et al. (2024) , MMT Ying et al. (2024) ), Vision-Centric (VStar Wu & Xie (2024) , MMVP Tong et al. (2024b) , CV-Bench Tong et al. (2024a) ), Hallucination (POPE Li et al. (2023b) , HallusionBench Guan et al. (2024) ), OCR (OCRBench Liu et al. (2024b) ), Knowledge (MMMU) Yue et al. (2024) , and Math (MathVerse Zhang et al. (2024) ). – denotes results not reported in original papers.

Model	Perception			Reasoning			Overall
Model	CoarseP.	FP-Cross	FP-Single	AttributeR.	LogicalR.	RelationR.	Overall
X-Omni-RL	84.12	66.43	77.13	76.88	48.31	73.91	74.31
+UAE	84.46	67.13	79.18	77.39	51.69	73.91	75.43
+Ours	84.46	67.83	78.50	78.39	55.08	74.78	75.96
Gain( $\uparrow$ )	+0.34	+1.40	+1.37	+1.51	+6.77	+0.87	+1.65

Table 3: Fine-grained analysis on MMBench. We report detailed performance across Perception (Coarse-Perception, Fine-grained Cross-instance, Fine-grained Single-instance) and Reasoning (Attribute Reasoning, Logical Reasoning, Relation Reasoning) sub-tasks. Our method improves both perception and reasoning capabilities, and the gains in reasoning are more significant.

Model	Color	Pos/Cnt	Col/Pos	Col/Cnt	Mul-Cnt	Pos/Size	Count	Overall
Janus-Pro Chen et al. (2025c)	45.0	7.5	30.0	12.5	12.5	35.0	30.0	24.60
T2I-R1 Jiang et al. (2025a)	67.5	7.5	35.0	20.0	30.0	25.0	32.5	31.10
BLIP3o Chen et al. (2025a)	25.0	12.5	60.0	12.5	22.5	57.5	25.0	30.70
OmniGen2 Wu et al. (2025b)	55.0	12.5	27.5	20.0	45.0	25.0	42.5	32.50
Bagel Deng et al. (2025)	32.5	25.0	32.5	25.0	37.5	47.5	60.0	37.10
X-Omni-RL	25.0	32.5	47.5	15.0	45.0	55.0	60.0	40.00

Table 4: Performance comparison on GenEval ++ . This benchmark evaluates fine-grained image generation capabilities across compositional attributes (e.g., Color, Position, Count). Our method achieves the highest overall score among the compared models.

	MM-P	MM-R	Und. Avg	Gen. Avg	Time
X-Omni-RL	75.89	66.36	52.66	42.50	—
Latent (Ours)	76.93	69.42	54.26	47.40	25 h
SigLIP-2	76.92	67.84	52.77	45.92	60 h
DINOv3	75.78	66.93	53.26	43.97	60 h
MSE	76.91	67.54	53.43	42.29	58 h
GPT-4.1	75.08	66.58	52.38	44.15	68 h

Table 5: (a) Reward design: latent reward vs. alternative formulations. (b) Sampling strategy:

K/M

sweep and flat GRPO comparison under matched budget. MM-P/R denote MMBench Perception/Reasoning. Und. Avg averages MMBench, MMVP, CV-Bench, and MathVerse; Gen. Avg averages WISE and GenEval++. Both are equally weighted means of percentage scores.

Figure 4: Comparison of reconstruction rewards. Across these examples, the latent reward produces the same preference ordering over candidate reconstructions as SigLIP2 and DINOv3, while operating directly on visual tokens without image decoding or external re-encoding.

Figure 5: Vehicle text-removal case. We compare the intermediate target-state descriptions and final reconstructions under the same edit instruction. The baseline retains the target text, whereas ReasonUni removes the largest text string on the vehicle’s front while preserving non-target details such as the license plate.

Figure 6: Qualitative results on WISE. For knowledge-focused prompts, ReasonUni generates more semantically accurate concepts than the baseline. The illustrated examples cover animal recognition, instrument categories, art-related knowledge, and historical technology.

Figure 7: Case studies across RL training steps. Starting from a baseline model, ReasonUni progressively improves instruction-following edits as TreeGRPO training proceeds (Step 200 $\rightarrow$ 800). Top: the model increasingly captures long-term wear patterns on boots. Bottom: the model more accurately removes the rider in the center of the scene together with the vehicle.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Visualization of synthesized reasoning editing samples . We construct the dataset by leveraging image-conversation pairs from LLaVA-NeXT-Data Liu et al. (2024a) . Left : The source image and the original multi-turn dialogue serving as context. Right : The pair of source and edited images, with the synthesized instruction displayed below. The instruction is derived from the dialogue to guide the editing process. Each entry forms a complete tuple $(I_{src},C_{hist},T_{instr},I_{edit})$ , where the editing command is semantically grounded in the conversation history.

Figure 9: Additional instruction-based editing cases. We compare baseline and ReasonUni target-state descriptions and reconstructions across floral, maritime, and street scenes. Across the three cases, ReasonUni better matches the displayed edited references while introducing fewer unrelated changes.

Figure 10: Qualitative results on GenEval++ . Compared with the baselines, ReasonUni better preserves compositional attributes, object counts, and spatial relations.

Figure 11: Qualitative results on WISE. For knowledge-focused prompts, ReasonUni generates more semantically accurate concepts than the baseline, including animal recognition, instrument category, art-related knowledge, and historical technology.

Method	Understanding				Generation
Method	MMBench	MMVP	CVBench	MathVerse	DPGBench	WISE	Geneval++
X-Omni-SFT	74.4	48.67	67.98	17.01	65.25	24	7.5
+TreeGRPO	75.51	50.00	69.20	16.75	77.24	34	20.36
+TreeGRPO +Reasoning	75.60	52.67	68.85	17.77	78.59	37.00	23.21
X-Omni-RL	74.31	50.67	68.26	17.39	82.29	45	40
+TreeGRPO	75.86	50.67	67.91	17.36	82.98	45	43.21

Appendix

Table 7: Ablation study of proposed components. We evaluate the effectiveness of TreeGRPO and reasoning augmentation across the SFT and RL settings.

Method	General				Vision-Centric			Hallucination		OCR	Knowledge	Math
Method	MMB	SEED	MMStar	MMT	VStar	MMVP	CVB.	POPE	Hallu.	OCRB.	MMMU	MathV.
Show-o Xie et al. (2024)	47.9	–	–	–	–	50.0	–	73.8	–	–	25.1	–
Emu3 Wang et al. (2024b)	58.5	68.2	–	–	51.8	36.6	–	85.2	–	68.7	31.6	–
VILA-U Wu et al. (2024b)	66.6	57.1	–	–	38.7	22.0	–	85.8	–	–	32.2	–
Janus Pro Chen et al. (2025c)	79.2	72.1	–	–	–	–	–	87.4	–	59.5	41.0	–
MetaMorph Tong et al. (2025b)	75.2	71.8	–	–	44.0	48.3	–	–	–	–	41.8	–

Appendix

Table 8: Comparison with SOTA unified models. We conduct a comprehensive evaluation, including General (MMBench Liu et al. (2023) , SEED-Bench Li et al. (2023a) , MMStar Chen et al. (2024) , MMT Ying et al. (2024) ), Vision-Centric (VStar Wu & Xie (2024) , MMVP Tong et al. (2024b) , CV-Bench Tong et al. (2024a) ), Hallucination (POPE Li et al. (2023b) , HallusionBench Guan et al. (2024) ), OCR (OCRBench Liu et al. (2024b) ), Knowledge (MMMU) Yue et al. (2024) , and Math (MathVerse Zhang et al. (2024) ). “–” denotes results not reported in original papers.

Method	DPGBench				WISE				GenEval++
Method	Rel.	Attr.	Other	Overall	Cult.	Time	Phys.	Overall	Pos/Cnt	Pos/Size	Cnt	Overall
Harmon Wu et al. (2025c)	-	-	-	-	38	48	44	41	-	-	-	-
Show-o Xie et al. (2024)	-	-	-	-	28	40	46	35	-	-	-	-
Emu3 Wang et al. (2024b)	90.22	86.84	83.15	80.60	-	-	-	-	-	-	-	-
Janus-Pro Chen et al. (2025c)	89.32	89.40	89.48	84.19	30	37	42	35	7.5	35.0	30.0	24.60
MetaQuery Pan et al. (2025)	-	-	-	82.05	56	55	63	55	-	-	-	-

Appendix

Table 9: Comparisons on generation benchmarks. We report performance on DPGBench Hu et al. (2024) , WISE Niu et al. (2025b) , and GenEval++ Ye et al. (2025) . Compared to the baseline and other unified models, our approach demonstrates favorable performance, showing consistent gains in attribute binding, world knowledge, and complex query alignment.

Model	Cultural	Time	Space	Biology	Physics	Chemistry	Overall
Harmon	38	48	52	37	44	29	41
Show-o	28	40	48	30	46	30	35
Janus Pro	30	37	49	36	42	26	35
MetaQuery-XL	56	55	62	49	63	41	55
UniWorld-V1	53	55	73	45	59	41	55
OmniGen2	42	52	64	43	50	34	47

Appendix

Table 10: Performance comparison on WISE. WISE assesses knowledge-grounded generation across six domains: Cultural, Time, Space, Biology, Physics, and Chemistry. ReasonUni improves over X-Omni-RL in every domain and achieves an overall score of 48.

Unified Multimodal Models

Unified multimodal models combine multiple data types in shared systems for cross-modal understanding, generation, reasoning, and interaction.

See all 287 papers

Reconstruction Alignment Improves Unified Multimodal Models
Sep 8, 2025Ji Xie, Trevor Darrell, Luke Zettlemoyer +1Unified Multimodal Models Cross-Modal Alignment
1UC Berkeley · University of Washington · 3Duke University
Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process
Jul 4, 2026Zican Hu, Xuyang Hu, Yiming Liu +10Unified Multimodal Models Multimodal Reasoning
Nanjing University · Shanghai AI Laboratory · Tsinghua University +5
UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models
Feb 9, 2026Cheng Yang, Chufan Shi, Bo Shui +7Unified Multimodal Models Image Generation
University of California San Diego · University of Southern California · University of Illinois Urbana-Champaign +1

Sep 8, 2025J/K move · Enter open · S save

Reconstruction Alignment Improves Unified Multimodal Models

Ji Xie, Trevor Darrell, Luke Zettlemoyer, XuDong Wang

1UC Berkeley · University of Washington · 3Duke University

Unified Multimodal Models Cross-Modal Alignment Multimodal Generation Multimodal Foundation Models Multimodal Understanding

Open paper AskPDF arXiv

Unified multimodal models (UMMs) unify visual understanding and generation within a single architecture. However, conventional training relies on image-text pairs (or sequences) whose captions are typically sparse and miss fine-grained visual details, even when they use hundreds of words to describe a simple image. We introduce Reconstruction Alignment (RECA), a resource-efficient post-training method that leverages visual understanding encoder embeddings as dense "text prompts", providing rich supervision without captions. Concretely, RECA conditions a UMM on its own visual understanding embeddings and optimizes it to reconstruct the input image with a self-supervised reconstruction loss, thereby realigning understanding and generation. Despite its simplicity, RECA is broadly applicable: across autoregressive, masked-autoregressive, and diffusion-based UMMs, it consistently improves generation and editing fidelity. With only 27 GPU hours, post-training with RECA substantially improves image generation performance on GenEval (0.73

\rightarrow

0.90) and DPGBench (80.93

\rightarrow

88.15), while also boosting editing benchmarks (ImgEdit 3.38

\rightarrow

3.75, GEdit 6.94

\rightarrow

7.27). Notably, RECA surpasses much larger open-source models and applies broadly across diverse UMM architectures, establishing it as an efficient and general post-training alignment strategy for UMMs.

Abstract

Unified Multimodal Models

Explore similar work

Reconstruction Alignment Improves Unified Multimodal Models

Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models