arxiv: 2605.11400 · v1 · submitted 2026-05-12 · 💻 cs.MM

Recognition: 2 theorem links

· Lean Theorem

UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning

Hayes Bai , Yinyi Luo , Wenwen Wang , Qingsong Wen , Jindong Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 01:25 UTC · model grok-4.3

classification 💻 cs.MM

keywords unified multimodal modelscoordination pathsadaptive reasoningunderstanding and generationmultimodal reasoningpath selectioninterpretable behaviors

0 comments

The pith

Multimodal tasks benefit when models adaptively choose among multiple ways to coordinate understanding and generation instead of using one fixed pattern.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that unified multimodal models can improve by treating coordination between understanding and generation as a selectable path rather than a single fixed strategy. Different inputs naturally favor different sequences, such as direct answers, textual inference steps, visual thought construction, or hypothesis exploration. By training a path-conditioned executor on role-aligned trajectories and adding a lightweight planner for input-dependent selection, the approach exploits this diversity at inference time. A sympathetic reader would care because rigid coordination limits performance on varied tasks while adaptive selection yields gains and exposes interpretable intermediate steps. The core result is that performance rises over fixed baselines precisely when the model is allowed to pick the right path per input.

Core claim

Unified multimodal reasoning improves by representing task solving as the selection and execution of one coordination path drawn from a set that includes direct answering, textual inference, visual-thought construction, and hypothesis-based exploration. Role-aligned trajectories train a path-conditioned executor while a lightweight planner selects the path based on the specific input, replacing any single enforced pattern with input-dependent adaptation.

What carries the argument

Coordination-path diversity, implemented as a planner that selects among paths and a path-conditioned executor that runs the chosen sequence of understanding and generation steps.

If this is right

Performance gains appear on tasks where inputs vary in the coordination they require, because the planner routes each case to its favored path.
Intermediate behaviors become interpretable because the chosen path is explicit and can be inspected or logged.
Training remains compatible with existing unified architectures since only the planner and executor conditioning are added.
Inference cost can be modulated by choosing shorter or longer paths depending on input complexity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same path-selection idea could apply to other domains that mix multiple reasoning modes, such as code generation interleaved with explanation.
If paths are made composable, models might learn to create novel hybrid paths rather than only selecting from a fixed set.
The planner could be extended to output a short rationale for its choice, turning the selection step into an additional source of explanation.

Load-bearing premise

Multimodal tasks contain enough diversity in which coordination path works best that an input-dependent selector can reliably pick and execute the right one at inference time.

What would settle it

A controlled comparison on held-out multimodal benchmarks where the adaptive planner is replaced by the single best fixed path and the adaptive version shows no accuracy or efficiency gain.

Figures

Figures reproduced from arXiv: 2605.11400 by Hayes Bai, Jindong Wang, Qingsong Wen, Wenwen Wang, Yinyi Luo.

**Figure 1.** Figure 1: Coordination-path diversity in unified multimodal models. Different coordination paths exhibit complementary strengths across inputs. The large oracle gap over fixed strategies suggests that exploiting coordination-path diversity can significantly improve UMM performance. representations [Qin et al., 2025], or use explicit coordination patterns such as analyzing-drafting loops and interleaved reasoning-gen… view at source ↗

**Figure 2.** Figure 2: Overview of the training and inference process of UniPath. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Planner behavior across benchmarks and validation transfer. (a) Selected path distribution. (b) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: Accuracy versus average output-token cost on understanding benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: and [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Full instance-level path correctness on MMMU. Each column is an MMMU example and each row is a coordination path. Colored cells indicate that the corresponding path answers the example correctly, while gray cells indicate failure. The sparse and non-identical correctness patterns show that path complementarity also appears at the individual-question level, not only after aggregating by subject. B Training … view at source ↗

**Figure 7.** Figure 7: Selected path distributions of planner checkpoints. pl1 is the final planner used in our system. Lower-utility planner variants often collapse to a small subset of paths, while the final planner keeps a broader routing pattern. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

**Figure 8.** Figure 8: Planner feature-space visualization on correctly solved sampled examples. Panels (a)–(b) show the global PCA–UMAP embedding colored by dataset/domain and by one randomly sampled oracle-correct path label for multi-path examples. Panels (c)–(e) show local UMAP projections within three representative buckets, colored by the planner-selected successful path. The global space shows stronger dataset/domain clus… view at source ↗

**Figure 9.** Figure 9: Smoothed training losses for the staged LoRA executor. Panels are arranged in one row and use the same stage names as [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗

read the original abstract

Unified multimodal models (UMMs) aim to integrate understanding and generation within a single architecture. However, it remains underexplored how to effectively coordinate these two capabilities for more effective and efficient reasoning. Existing coordination approaches either perform coupling during training, without explicit inference-time coordination, or impose a fixed coordination pattern for all inputs. In this work, we show that multimodal tasks exhibit substantial coordination-path diversity: different inputs favor different coordination paths. This suggests that exploiting such diversity is key to improving performance. We propose UniPath, a framework for adaptively modeling and exploiting coordination-path diversity. Instead of enforcing a single coordination pattern, we represent task solving as the selection and execution of a path, ranging from direct answering to textual inference, visual-thought construction, and hypothesis-based exploration. We construct role-aligned trajectories to train a path-conditioned executor and introduce a lightweight planner mechanism to enable input-dependent path selection. Experiments show that leveraging coordination-path diversity improves performance over fixed coordination strategies while providing interpretable intermediate behaviors. The code is available at:https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/unipath.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

UniPath's adaptive path selection for multimodal coordination is a clear idea, but gains may trace to multi-path training data rather than the planner.

read the letter

The key takeaway is that UniPath tries to improve unified multimodal reasoning by letting a planner pick different coordination paths at inference time based on the input. This could be useful if the diversity in paths really matters, but the evidence so far is thin. They represent solving a task as selecting one of several paths: direct answering, textual inference, visual-thought construction, or hypothesis exploration. They create role-aligned trajectories for training a path-conditioned executor and train a lightweight planner to choose the path. This setup is presented as a way to exploit coordination-path diversity instead of forcing one pattern for everything. What the paper does well is to make the coordination explicit and selectable. Previous approaches either baked it into training without choice or used a fixed pattern. Adding the planner and the trajectories gives a mechanism for adaptation and some interpretability through the path choices. Releasing the code at the GitHub link is a plus for anyone who wants to reproduce or extend it. The main soft spot is around the experiments and the source of the gains. The abstract claims better performance over fixed strategies, but gives no specifics on the datasets, metrics, or how large the improvements are. The stress-test concern is on point here: the fixed baselines might have been trained with less diverse data. If the advantage disappears when all models see the same multi-path trajectories, then the planner's adaptivity isn't doing the heavy lifting. The paper should demonstrate that the selection itself adds value beyond richer supervision. This work is for people in multimodal AI who are thinking about how to blend understanding and generation more dynamically. A reader working on vision-language models or inference-time methods could pick up ideas from the path framework. It has a clear proposal and an implementation, so it deserves a serious referee to check the details and controls. I would recommend putting it through peer review, with the reviewers paying close attention to the baseline training setup and any ablation on the planner.

Referee Report

2 major / 2 minor

Summary. The paper introduces UniPath, a framework for unified multimodal models (UMMs) that models coordination between understanding and generation as selection among diverse paths (direct answering, textual inference, visual-thought construction, hypothesis exploration). It constructs role-aligned trajectories to train a path-conditioned executor and a lightweight planner for input-dependent path selection at inference time, claiming that exploiting coordination-path diversity yields better performance than fixed coordination strategies while producing interpretable intermediate behaviors. Code is released.

Significance. If the reported gains are shown to stem specifically from inference-time adaptive selection rather than from multi-path training data alone, the approach could offer a practical way to improve reasoning efficiency and flexibility in UMMs without committing to a single coordination pattern. The public code release supports reproducibility and follow-up work.

major comments (2)

[Experiments] Experiments section (and abstract): the fixed-coordination baselines are described as using 'a single coordination pattern,' but it is not stated whether these baselines were trained on the identical role-aligned multi-path trajectory set (with path forced) or on single-pattern data only. This distinction is load-bearing for the central claim that adaptive selection, rather than exposure to diverse supervision, drives the improvement.
[Method / Experiments] §3 (Method) and Experiments: the claim that 'multimodal tasks exhibit substantial coordination-path diversity' is asserted but lacks a quantitative breakdown (e.g., a table of per-input path selection frequencies, oracle path performance, or ablation showing planner accuracy). Without this, it is difficult to assess whether the planner's selections are meaningfully input-dependent or merely defaulting to a high-performing path.

minor comments (2)

[Abstract / Experiments] Abstract and §4: dataset names, sizes, exact metrics, and statistical significance tests are not mentioned, making it hard to contextualize the reported improvements.
[Method] Notation: the distinction between the 'path-conditioned executor' and the base UMM is introduced without an explicit equation or diagram showing how path conditioning is injected (e.g., via prefix tokens, adapter layers, or conditioning vector).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight key aspects needed to strengthen our claims. We address each major point below and will revise the manuscript to provide the requested clarifications and quantitative details.

read point-by-point responses

Referee: [Experiments] Experiments section (and abstract): the fixed-coordination baselines are described as using 'a single coordination pattern,' but it is not stated whether these baselines were trained on the identical role-aligned multi-path trajectory set (with path forced) or on single-pattern data only. This distinction is load-bearing for the central claim that adaptive selection, rather than exposure to diverse supervision, drives the improvement.

Authors: We agree that this distinction is essential to isolate the benefit of adaptive path selection. In our experiments, each fixed-coordination baseline was trained only on data corresponding to its single fixed pattern (i.e., single-pattern data), while the role-aligned multi-path trajectories were used exclusively to train the path-conditioned executor and the planner in UniPath. To eliminate ambiguity, we will revise the Experiments section to explicitly state the training data for all baselines and add an ablation comparing UniPath to a multi-path-trained model that uses fixed (non-adaptive) path selection at inference time. revision: yes
Referee: [Method / Experiments] §3 (Method) and Experiments: the claim that 'multimodal tasks exhibit substantial coordination-path diversity' is asserted but lacks a quantitative breakdown (e.g., a table of per-input path selection frequencies, oracle path performance, or ablation showing planner accuracy). Without this, it is difficult to assess whether the planner's selections are meaningfully input-dependent or merely defaulting to a high-performing path.

Authors: We acknowledge that quantitative evidence is needed to support the diversity claim and to show the planner's input-dependence. We will add to the Experiments section a table reporting per-task and per-input path selection frequencies by the planner, oracle upper-bound performance (best path per input), and planner accuracy relative to oracle or preferred paths. These additions will demonstrate that selections vary meaningfully with input rather than defaulting to a single high-performing path. revision: yes

Circularity Check

0 steps flagged

No circularity: new framework with independent training and selection components

full rationale

The paper defines UniPath as a new training and inference mechanism: role-aligned trajectories are constructed to train a path-conditioned executor, and a separate lightweight planner is introduced for input-dependent path selection at inference time. This structure does not reduce any claimed prediction or result to a fitted parameter by construction, nor does it rely on self-citation chains, uniqueness theorems imported from prior author work, or ansatzes smuggled via citation. The observation of coordination-path diversity motivates the design but is not presupposed in the method's equations or training objective. Experimental comparisons to fixed coordination strategies are presented as external evaluations rather than tautological outcomes of the same inputs. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Abstract-only review limits visibility into parameters; the core assumption of path diversity is treated as a domain observation rather than derived.

axioms (1)

domain assumption Multimodal tasks exhibit substantial coordination-path diversity where different inputs favor different paths
Explicitly stated as the key observation motivating the framework

invented entities (2)

path-conditioned executor no independent evidence
purpose: Executes the selected coordination path during inference
New component introduced to handle variable paths
lightweight planner no independent evidence
purpose: Selects input-dependent coordination path
New mechanism for adaptive selection

pith-pipeline@v0.9.0 · 5511 in / 1189 out tokens · 30358 ms · 2026-05-13T01:25:02.697634+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We propose UniPath, a framework for adaptively modeling and exploiting coordination-path diversity... planner-executor system that selects paths per input
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

role-weighted language modeling loss... aligned visual thought supervision

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 12 internal anchors

[1]

Qwen3-VL Technical Report

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset

Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset.arXiv preprint arXiv:2505.09568, 2025a. Junying Chen, Zhenyang Cai, Pengcheng Chen, Shunian Chen, Ke Ji, Xidong Wang, Yunjin Yang, ...

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling

Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811, 2025c. Zihui Cheng, Qiguang Chen, Xiao Xu, Jiaqi Wang, Weiyun Wang, Hao Fei, Yidong Wang, Alex Jinpeng Wang, Zhi Chen, Wanxiang Che, et ...

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Emerging Properties in Unified Multimodal Pretraining

Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025a

Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Shilin Yan, Hao Tian, Xingyu Zeng, Rui Zhao, Jifeng Dai, et al. Got: Unleashing reasoning capability of multimodal large language model for visual generation and editing.arXiv preprint arXiv:2503.10639, 2025a. Rongyao Fang, Aldrich Yu, Chengqi Duan, Linjiang Huang, Shuai Bai, Yuxuan Cai, Kun Wa...

work page arXiv
[6]

Interleaving reasoning for better text-to-image generation.arXiv preprint arXiv:2509.06945, 2025

Wenxuan Huang, Shuang Chen, Zheyong Xie, Shaosheng Cao, Shixiang Tang, Yufan Shen, Qingyu Yin, Wenbo Hu, Xiaoman Wang, Yuntian Tang, et al. Interleaving reasoning for better text-to-image generation. arXiv preprint arXiv:2509.06945,

work page arXiv
[7]

T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703,

Dongzhi Jiang, Ziyu Guo, Renrui Zhang, Zhuofan Zong, Hao Li, Le Zhuo, Shilin Yan, Pheng-Ann Heng, and Hongsheng Li. T2i-r1: Reinforcing image generation with collaborative semantic-level and token-level cot.arXiv preprint arXiv:2505.00703,

work page arXiv
[8]

LLaVA-OneVision: Easy Visual Task Transfer

Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

Imagine while reasoning in space: Multimodal visualization-of-thought, 2025b.https://arxiv.org/abs/2501.07542

Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vuli´c, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought.arXiv preprint arXiv:2501.07542,

work page arXiv
[10]

UniWorld-V1: High-Resolution Semantic Encoders for Unified Visual Understanding and Generation

Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, et al. Uniworld-v1: High-resolution semantic encoders for unified visual understanding and generation.arXiv preprint arXiv:2506.03147,

work page internal anchor Pith review Pith/arXiv arXiv
[11]

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

work page internal anchor Pith review Pith/arXiv arXiv
[12]

TorchUMM: A Unified Multimodal Model Codebase for Evaluation, Analysis, and Post-training

Yinyi Luo, Wenwen Wang, Hayes Bai, Hongyu Zhu, Hao Chen, Pan He, Marios Savvides, Sharon Li, and Jindong Wang. Torchumm: A unified multimodal model codebase for evaluation, analysis, and post-training.arXiv preprint arXiv:2604.10784,

work page internal anchor Pith review Pith/arXiv arXiv
[13]

Wise: A world knowledge-informed semantic evaluation for text-to-image generation.arXiv preprint arXiv:2503.07265,

Yuwei Niu, Munan Ning, Mengren Zheng, Weiyang Jin, Bin Lin, Peng Jin, Jiaqi Liao, Chaoran Feng, Kunpeng Ning, Bin Zhu, et al. Wise: A world knowledge-informed semantic evaluation for text-to-image generation.arXiv preprint arXiv:2503.07265,

work page arXiv
[14]

Training vision-language pro- cess reward models for test-time scaling in multimodal rea- soning: Key insights and lessons learned.arXiv preprint arXiv:2509.23250, 2025

Brandon Ong, Tej Deep Pala, Vernon Toh, William Chandra Tjhi, and Soujanya Poria. Training vision- language process reward models for test-time scaling in multimodal reasoning: Key insights and lessons learned.arXiv preprint arXiv:2509.23250,

work page arXiv
[15]

Uni-cot: Towards unified chain-of-thought reasoning across text and vision.arXiv preprint arXiv:2508.05606,

Luozheng Qin, Jia Gong, Yuqing Sun, Tianjiao Li, Mengping Yang, Xiaomeng Yang, Chao Qu, Zhiyu Tan, and Hao Li. Uni-cot: Towards unified chain-of-thought reasoning across text and vision.arXiv preprint arXiv:2508.05606,

work page arXiv
[16]

Unigame: Turning a unified multimodal model into its own adversary.arXiv preprint arXiv:2511.19413,

Zhaolong Su, Wang Lu, Hao Chen, Sharon Li, and Jindong Wang. Unigame: Turning a unified multimodal model into its own adversary.arXiv preprint arXiv:2511.19413,

work page arXiv
[17]

Gemini: A Family of Highly Capable Multimodal Models

Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

work page internal anchor Pith review Pith/arXiv arXiv
[18]

Deepgen 1.0: A lightweight unified multimodal model for advancing image generation and editing.arXiv preprint arXiv:2602.12205, 2026

Dianyi Wang, Ruihang Li, Feng Han, Chaofan Ma, Wei Song, Siyuan Wang, Yibin Wang, Yi Xin, Hongjian Liu, Zhixiong Zhang, et al. Deepgen 1.0: A lightweight unified multimodal model for advancing image generation and editing.arXiv preprint arXiv:2602.12205,

work page arXiv
[19]

Ovis-u1 technical report

Guo-Hua Wang, Shanshan Zhao, Xinjie Zhang, Liangfu Cao, Pengxin Zhan, Lunhao Duan, Shiyin Lu, Ming- hao Fu, Xiaohao Chen, Jianshan Zhao, et al. Ovis-u1 technical report.arXiv preprint arXiv:2506.23044,

work page arXiv
[20]

Emu3: Next-Token Prediction is All You Need

Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need.arXiv preprint arXiv:2409.18869,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

OmniGen2: Towards Instruction-Aligned Multimodal Generation

Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12966–12977, 2025a. Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shi...

work page internal anchor Pith review Pith/arXiv arXiv
[22]

Openuni: A simple baseline for unified multimodal understanding and generation.arXiv preprint arXiv:2505.23661, 2025c

Size Wu, Zhonghua Wu, Zerui Gong, Qingyi Tao, Sheng Jin, Qinyue Li, Wei Li, and Chen Change Loy. Openuni: A simple baseline for unified multimodal understanding and generation.arXiv preprint arXiv:2505.23661, 2025c. Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Zhonghua Wu, Qingyi Tao, Wentao Liu, Wei Li, and Chen Change Loy. Harmonizing visual representati...

work page arXiv
[23]

Reconstruction alignment improves unified multimodal models.arXiv preprint arXiv:2509.07295, 2025a

Ji Xie, Trevor Darrell, Luke Zettlemoyer, and XuDong Wang. Reconstruction alignment improves unified multimodal models.arXiv preprint arXiv:2509.07295, 2025a. Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimod...

work page arXiv
[24]

Show-o2: Improved Native Unified Multimodal Models

Jinheng Xie, Zhenheng Yang, and Mike Zheng Shou. Show-o2: Improved native unified multimodal models. arXiv preprint arXiv:2506.15564, 2025b. Zhiyuan Yan, Kaiqing Lin, Zongjian Li, Junyan Ye, Hui Han, Zhendong Wang, Hao Liu, Bin Lin, Hao Li, Xue Xu, et al. Can understanding and generation truly benefit together–or just coexist?arXiv e-prints, pages arXiv–2509,

work page internal anchor Pith review Pith/arXiv arXiv
[25]

Mmada: Multimodal large diffusion language models

Ling Yang, Ye Tian, Bowen Li, Xinchen Zhang, Ke Shen, Yunhai Tong, and Mengdi Wang. Mmada: Multimodal large diffusion language models.arXiv preprint arXiv:2505.15809,

work page arXiv
[26]

Echo-4o: Harnessing the power of gpt-4o synthetic images for improved image generation.arXiv preprint arXiv:2508.09987, 2025

Junyan Ye, Dongzhi Jiang, Zihao Wang, Leqi Zhu, Zhenghao Hu, Zilong Huang, Jun He, Zhiyuan Yan, Jinghua Yu, Hongsheng Li, et al. Echo-4o: Harnessing the power of gpt-4o synthetic images for improved image generation.arXiv preprint arXiv:2508.09987,

work page arXiv
[27]

Adaptive chain-of-focus reasoning via dynamic visual search and zooming for efficient vlms.arXiv preprint arXiv:2505.15436, 2025

Xintong Zhang, Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaowen Zhang, Yang Liu, Tao Yuan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, et al. Adaptive chain-of-focus reasoning via dynamic visual search and zooming for efficient vlms.arXiv preprint arXiv:2505.15436,

work page arXiv
[28]

thinking with images

14 Shanshan Zhao, Xinjie Zhang, Jintao Guo, Jiakui Hu, Lunhao Duan, Minghao Fu, Yong Xien Chng, Guo- Hua Wang, Qing-Guo Chen, Zhao Xu, et al. Unified multimodal understanding and generation models: Advances, challenges, and opportunities.arXiv preprint arXiv:2505.02567,

work page arXiv
[29]

3.1: (A) corresponds to pA, (U, A) to pU, (U, R, A) top R,(U, R, C, R, A)top C, and(U, R, H, R, A)top H

The row labels use the role-sequence notation from Sec. 3.1: (A) corresponds to pA, (U, A) to pU, (U, R, A) top R,(U, R, C, R, A)top C, and(U, R, H, R, A)top H. Figure 5:Full subject-level path affinity on MMMU.Each column corresponds to an MMMU subject, and each row reports the accuracy of one coordination path. The overall pattern shows that subject dom...

work page 2017
[30]

Table 9: Routed results with the Harmon-1.5B backbone

This suggests that the coordination paths remain useful across backbones, while converting path complementarity into reliable routed gains becomes harder with a weaker executor. Table 9: Routed results with the Harmon-1.5B backbone. Scores are accuracies in percent. Method MMMU MathVista MMStar MMB-EN MMB-CN Average Raw 34.33 24.50 37.47 62.72 54.65 42.73...

work page 1980
[31]

thought CE

Executor training is organized into four staged splits that activate different links of the path: text-only understanding (S1), understanding with aligned visual thoughts (S2), final image generation without aligned visual-thought supervision (S3), and final image generation with construction or hypothesis supervision (S4). The four stages form a strict c...

work page 2048
[32]

Table 13: Staged executor split statistics

They describe the executor-stage data rather than the additional path-outcome runs used to build planner supervision. Table 13: Staged executor split statistics. Stage Split name Train Val Total S1understanding_text12,733 164 12,897 S2understanding_visual5,232 68 5,300 S3image_answer_plain5,380 92 5,472 S4image_answer_visual6,282 115 6,397 Total – 29,627 ...

work page 2022