acceptodds
Under review as a conference paper at ICLR 2027

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

38 views

Abstract

Unified multimodal models have recently demonstrated strong generative capabilities, yet whether and when generation improves understanding remains unclear. Existing benchmarks lack a systematic exploration of the specific tasks where generation facilitates understanding. To this end, we introduce UniG2U-Bench, a comprehensive benchmark categorizing generation-to-understanding (G2U) evaluation into 7 categories and 30 subtasks, requiring varying degrees of implicit or explicit visual transformations. Extensive evaluation of over 30 models reveals three core findings: Unified models generally underperform their base Vision-Language Models (VLMs), and Generate-then-Answer (GtA) inference typically degrades performance relative to direct inference. Consistent enhancements emerge in spatial intelligence, visual illusions, or multi-round reasoning subtasks, where enhanced spatial and shape perception, as well as multi-step intermediate image states, prove beneficial. Tasks with similar reasoning structures and models sharing architectures exhibit correlated behaviors, suggesting that generation-understanding coupling induces class-consistent inductive biases over tasks, pretraining data, and model architectures. These findings highlight the necessity for more diverse training data and novel paradigms to fully unlock the potential of unified multimodal modeling.

Primary area:
datasets and benchmarks
Keywords:
unify model, generation help understanding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3yavOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

UniG2U-Bench studies whether unified image generation and understanding models improve multimodal understanding, and whether generating an intermediate image at inference time helps answer visual questions. The benchmark contains 3,000 instances spanning seven categories and 30 subtasks, with evaluations of more than 30 models. It distinguishes two comparisons: unified models versus their selected base or proxy understanding models under Direct answering, and Generate-then-Answer versus Direct answering within unified models.

The paper reports that 12 of 16 paired unified models underperform their selected comparators under Direct answering, while six of eight applicable models underperform under Generate-then-Answer. It also identifies task-specific benefits, particularly in some spatial and iterative reasoning settings. The analysis includes a conceptual gradient-alignment framework, a likelihood-based account of commitment to generated intermediates, correlation analyses, failure categorization, and reference-intermediate interventions.

Critical strengths
  1. The benchmark separates two important questions that are often conflated. Improvements associated with unified training and improvements from explicitly generating visual intermediates at inference time are different phenomena. Evaluating them separately across a broad task taxonomy makes the benchmark useful for diagnosing where generation capabilities help understanding and where they introduce additional failure modes. The task-level results are more informative than a single aggregate comparison of unified models.

  2. The paper includes meaningful diagnostic controls and examines limitations of its own analysis. Appendix D controls for the shared-baseline effect in the correlation analysis by residualizing against base-model performance. Appendix N supplies manually checked reference intermediates, which helps distinguish failures of intermediate-image generation from failures to use an intermediate effectively. The separation of unsupported editing protocols into additional analyses is also appropriate. These controls make the empirical investigation more substantive than a broad leaderboard comparison alone.

Critical weaknesses
  1. The identities and matching rationale of the comparison models are not consistently specified, which affects the central performance deltas and grouping analyses. For example, Table 2 compares OneCAT-3B against Qwen2.5-VL-7B, while the original OneCAT paper identifies Qwen2.5-3B-Instruct as its language-model initialization. This requires a clear distinction between initialization and a chosen multimodal comparator. Using the already evaluated Qwen2.5-VL-3B as an alternative proxy changes the reported Direct deficit from −3.30 to −1.24, although its sign remains negative.

Similar ambiguity appears between Tables 2 and 13: the latter describes the backbone column as the paired base VLM, but lists Qwen-2.5 for BAGEL where Table 2 uses LLaVA-OneVision, and Qwen3-1.7B for Ovis-U1 where Table 2 uses Qwen2.5-VL-3B. Proxy comparisons can be informative, and the paper appropriately acknowledges that these are not causal estimates of unified training. Nevertheless, the exact initialization, visual architecture, selected comparator, and rationale need to be distinguished consistently. Otherwise, the magnitude of the reported deltas and the interpretation of “same-backbone” groupings are difficult to reproduce.

  1. The documented Direct/GtA protocols do not yet establish a controlled comparison of the contribution of generated images. Appendix K is labeled Non-CoT, but K.4 requests multiple generated images and substantially duplicates the visual-CoT prompts in L.10. The Direct object-motion prompt in K.1.2 uses camera-motion framing, whereas L.1.4 uses object-motion framing. The Direct and GtA MSR prompts also differ in whether they explicitly ask for multi-step spatial reasoning. These may be documentation errors; the PDF does not establish that they affected execution. If they were actually used, however, the comparison changes task framing and reasoning instructions in addition to image generation. The claimed budget matching is also not accompanied by sufficiently concrete per-model text-token, image-generation, and iteration limits. Because inference-time G2U is a principal research question, uncertainty about the executed prompts and resource controls is consequential.
Questions for authors
  1. [Clarification] Please provide an exact checkpoint and pairing table that separately identifies each model’s language initialization, visual component, selected understanding comparator, and matching rationale. Please also specify how the checkpoint labeled LLaDA-Instruct processes visual inputs and how backbone families are defined in the correlation analysis. How do the principal conclusions change on the subset with the most defensible matching?

  2. [Clarification] Which prompts were actually executed for Direct and GtA? Please resolve the generated-image requests in Appendix K.4 and the object-motion and MSR discrepancies between Appendices K and L. Are these documentation-only issues, or do any reported experiments require correction?

  3. [Assessment-relevant validation] Please report the actual text, image, and interaction budgets for both protocols. For the tasks where GtA changes reasoning scaffolding as well as adding an image, do the principal conclusions survive comparison with a text-only reasoning condition using equivalent task instructions and a clearly stated budget? Paired uncertainty for these comparisons would help establish how stable the reported effects are.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3yav (Part 1/3): Pairing of Unified Models and Comparators

We sincerely thank the reviewer for the very careful and detailed review, and for recognizing that the benchmark separates two questions that are often conflated, and that its diagnostic controls make the study more substantive than a leaderboard comparison. The reviewer's points on pairing and prompts were precise and correct, and addressing them has made the paper more reproducible. We answer each point below, and all changes are in the revised PDF (changes in blue).

W1 / Q1: "Please provide an exact checkpoint and pairing table that separately identifies each model's language initialization, visual component, selected understanding comparator, and matching rationale. [...] How do the principal conclusions change on the subset with the most defensible matching?"

Table 13 lists each unified model's language-model initialization, while Table 2 lists the model it is compared with; the two columns were labelled as if they were the same. The revision adds a pairing table (Appendix S) that gives, for each unified model, its language initialization, vision encoder, comparator, matching tier and the reason for the choice. Comparators follow the vision component and the training data, in three tiers:

  • Tier 1: the released VLM the unified model was initialized from (same vision encoder and understanding data);
  • Tier 2: a VLM with the same language initialization and a comparable vision encoder;
  • Tier 3: the closest available VLM in family and size, used only when neither exists.

MMaDA is compared with LLaDA-Instruct, the model it was initialized from, and is listed as Tier 2.

As the reviewer suggested, OmniGen2 is built on Qwen2.5-VL-3B, and OneCAT-3B is initialized from Qwen2.5-3B-Instruct, the language model of Qwen2.5-VL-3B; both are now compared with Qwen2.5-VL-3B (Tier 1 and Tier 2, respectively). On the subset with the most defensible matching, the principal conclusion holds: every Tier-1 model with Direct results scores below its comparator, and so does OneCAT-3B.

Table A. Direct gap to the comparator (points), Tier-1 models and OneCAT-3B (Appendix S).

ModelComparatorTierDirect gap
JavisGPTQwen2.5-VL-7B1−3.73
UAEQwen2.5-VL-3B1−1.45
OmniGen2Qwen2.5-VL-3B (was 7B)1−0.40 (was −2.46)
MammothModa2Qwen3-VL-8B1−7.78
OneCAT-3BQwen2.5-VL-3B (was 7B)2−1.24 (was −3.30)
AuthorsOpenReview ↗

Response to Reviewer 3yav (Part 2/3): Prompts Actually Executed

W2 / Q2: "Which prompts were actually executed for Direct and GtA? Please resolve the generated-image requests in Appendix K.4 and the object-motion and MSR discrepancies between Appendices K and L. Are these documentation-only issues, or do any reported experiments require correction?"

We thank the reviewer for checking the appendices so carefully. Both observations are correct. Appendix K.4 reproduces the visual-CoT prompt of L.10, and K.1.2 shows the camera-motion prompt instead of the object-motion one. These are documentation errors:

  • The Direct prompts actually executed for the Uni-MMMU tasks are text-only and do not ask for image generation. The jigsaw prompt, for example, asks the model to "mentally visualize placing each candidate into the bottom-right cell".
  • The executed object-motion prompt uses the object-motion wording.

Appendix U now gives every executed prompt, and K.4 and K.1.2 are corrected.

For MSR, the Direct prompt asks for step-by-step reasoning, while GtA gives this instruction in its generation stage. We reran MSR with a single shared instruction, so the conditions differ only in the intermediate.

Table B. MSR with one instruction shared by all conditions (accuracy, %, same input images).

ModelDirectText-CoTGtA
Ovis-U1322328
BAGEL302831
Janus-Pro232418
OmniGen2233227
AuthorsOpenReview ↗

Response to Reviewer 3yav (Part 3/3): Budget, Text-Only Control and Paired Uncertainty

Q3: "Please report the actual text, image, and interaction budgets for both protocols. For the tasks where GtA changes reasoning scaffolding as well as adding an image, do the principal conclusions survive comparison with a text-only reasoning condition using equivalent task instructions and a clearly stated budget? Paired uncertainty for these comparisons would help establish how stable the reported effects are."

Budgets. Appendix O lists the token caps, which are the same for Direct and the GtA answer stage within each task, the image-generation settings of each model, and the tokens and images actually used per item (see Tables D and E in our response to Reviewer AbYh).

Text-only control. To separate the extra step from the image, we added the Text-CoT and blank-image conditions (Appendix P). Both keep the GtA instructions, step structure and answer-stage cap. With the budget matched, GtA is below Text-CoT for all four models (significantly for Ovis-U1, Janus-Pro and OmniGen2 on 2,354 items), and for Ovis-U1 the blank-image pipeline is 2.8 points below Direct while the generated image accounts for a further 5.7. On all 3,000 items, including the four LLM-judged tasks, the principal conclusions hold (Table C).

Table C. Paired differences on all 3,000 items (points, bootstrap 95% CI).

ModelGtA − DirectGtA − Text-CoT
Ovis-U1−8.3 [−9.9, −6.7]−5.4 [−7.0, −3.8]
Janus-Pro−7.8 [−9.4, −6.2]−4.5 [−6.1, −2.9]
OmniGen2−0.8 [−2.0, +0.4]−1.1 [−2.5, +0.2]
BAGEL+0.3 [−0.9, +1.4]−0.9 [−2.3, +0.4]

Paired uncertainty. Every comparison between Direct, Text-CoT and GtA is now paired at the item level, with bootstrap 95% confidence intervals and exact McNemar tests, Holm-corrected across subtasks.

Table D. Examples of paired comparisons (Appendices P and T).

ComparisonModelDifference (points)Significance
GtA − Direct (original runs)TokenFlow−3.9p = 3×10⁻⁴
GtA − Direct (original runs)GPT-4o−1.7not significant
GtA − Text-CoT, IllusionBenchOvis-U1, Janus-Pro, BAGEL−11.8, −12.5, −4.5significant after Holm

Appendices P and T give the full tables.

We hope these responses address the reviewer's concerns. We would be sincerely grateful if the reviewer would take them into account in the final assessment, and we would be glad to provide any further clarification during the discussion period.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer 3yav,

Thank you again for the time and effort you have devoted to reviewing our paper. We have revised our responses to make them more precise, with each of your questions quoted and answered in turn. The responses cover the full pairing table with matching tiers, the prompts actually executed, the rerun of MSR with a shared instruction, and the budget-matched text-only control with paired uncertainty.

We would be grateful if you could take a look at the updated responses at your convenience and let us know whether they address your concerns. If any question remains, we would be very happy to discuss it further during the discussion period. If our responses have resolved your concerns, we would sincerely appreciate it if you would consider updating your assessment.

Thank you again for your effort and your constructive feedback.

Best regards,

The Authors

Reviewer_sHgYOpenReview ↗
Rating
4(+1)
Confidence
3(−1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces UniG2U-Bench, a benchmark designed to systematically investigate whether image generation capabilities in unified multimodal models can improve visual understanding, and under which circumstances such benefits emerge.

The authors curate 3,000 samples across 7 reasoning categories and 30 subtasks from existing datasets and evaluate over 30 models. By comparing unified models with their corresponding base VLMs, as well as Direct and Generate-then-Answer inference, the authors find that unified training is often associated with degraded understanding performance, while explicit visual generation does not consistently improve reasoning. However, benefits emerge in specific tasks involving spatial reasoning, visual illusions, and multi-step state transformations. The paper further provides a theoretical analysis based on gradient alignment and intermediate visual states to explain these observations.

Critical strengths
  1. Well-motivated research question. Existing unified multimodal benchmarks typically evaluate generation and understanding as separate capabilities. In contrast, this paper directly investigates whether generation can improve understanding and distinguishes the effects associated with unified training from those of explicit intermediate visual generation. This perspective addresses an important research question.

  2. Comprehensive and insightful empirical analysis. The paper evaluates a broad range of models across diverse reasoning tasks. It not only reveals potential degradation in understanding performance after multimodal unification but also identifies specific tasks where visual generation can provide meaningful benefits. The correlation analysis across tasks and model families further offers valuable insights for future research on unified multimodal models.

Critical weaknesses

Insufficient validation of the proposed mechanistic explanations. The paper uses gradient alignment and intermediate visual states to explain the observed performance differences, but the key theoretical quantities are not directly measured. Furthermore, Generate-then-Answer and Direct inference differ not only in whether intermediate images are generated but also in their additional reasoning steps and task-specific prompting. Without systematic comparisons against text-only CoT under matched inference budgets, the independent contribution of visual generation remains unclear.

Questions for authors

Could the authors compare Generate-then-Answer with text-only CoT under matched inference budgets and further examine the relationship between gradient alignment or intermediate visual quality and performance changes? Such experiments would help distinguish the contribution of visual generation itself from the benefits of additional inference computation.

If the authors make a convincing effort to address these concerns, I would consider raising my rating.

2 comments

W1: Direct evidence for the mechanism

The revision measures both quantities.

Gradient sharing. On three unified models with different designs, we compute the gradients of the understanding loss and of the model's own image-generation loss on the parameters the two losses share, using 30 items per category drawn with a fixed seed. For editing we use the ground-truth next state as the target.

ModelHow generation and understanding share parametersGeneration / understanding gradient normEditing / text-to-image
Janus-Pro-7Bwhole backbone shared0.090– (no editing)
Ovis-U1LLM conditions a separate generator0.00491.0×
BAGELseparate experts; the input image is encoded by the understanding expert0.00196.8×

These measurements match the relation in Section 3. The more parameters understanding and generation share, the more strongly generation training acts on understanding. Editing shares more with understanding than text-to-image generation, because an edit is conditioned on an input image that the understanding pathway encodes: in BAGEL the editing gradient on the shared parameters is about seven times the text-to-image one, and in Ovis-U1, where both kinds of generation pass through the same conditioning, they are equal. The direction of the interaction depends on the task: in Janus-Pro-7B the two gradients point the same way on multi-step spatial reasoning and conflict on IllusionBench, in line with the task-dependent gains in Table 2 (Appendix R).

Intermediate quality. Three annotators rated 207 intermediates. We drew 210 (70 per model) with a fixed seed, stratified over task groups and without looking at correctness, and used three as calibration examples; annotators did not see model names or answers (Krippendorff's α = 0.61).

ModelHuman fidelity (1–5)RA (1–5)GtA − Direct (points)
BAGEL3.082.37+3
Ovis-U11.871.80−17
Janus-Pro1.401.14−11

The model with the best-rated intermediates is the one that benefits from GtA. Re-scoring the same images with the paper's RA prompt gives a Spearman correlation of 0.64 with the human ratings, close to the agreement between annotators, so RA tracks image quality reliably (Appendix Q).

W2/Q: Comparison with text-only CoT under the same budget

We ran both controls. They keep the GtA instructions, step structure and token cap: in Text-CoT the model writes out the intermediate state instead of drawing it, and in GtA-blank the generated image is replaced by a blank one. Accuracy (%) on 2,354 items (all tasks except the four scored by an LLM judge: Geometry3K, AuxSolidMath, BabyVision and PhyX):

ModelDirectText-CoTGtA-blankGtA
Ovis-U136.333.533.527.8
Janus-Pro31.527.6–21.6
OmniGen233.033.433.331.5
BAGEL38.940.539.539.1

Given the same extra step, writing does at least as well as drawing for every model and clearly better for Ovis-U1 and Janus-Pro. For Ovis-U1, the step itself with a blank image costs 2.8 points (36.3 to 33.5), and its own generated image costs a further 5.7 (to 27.8). The size of the drop follows image quality: it is small for BAGEL, whose intermediates are rated highest, and large for Ovis-U1 and Janus-Pro. Raising the Text-CoT cap from 512 to 4,096 tokens leaves Ovis-U1's answers unchanged, so the text budget is not a constraint. Appendix P gives per-subtask results with paired confidence intervals.

Thank you for your reply. This resolves my question, and I will raising my score.

Reviewer_AbYhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces UniG2U-Bench, comprising 3,000 examples across seven categories and 30 subtasks, to study whether visual generation improves multimodal understanding. It compares unified models with paired base VLMs, evaluates direct versus generate-then-answer inference, and provides a likelihood-based interpretation. The results reveal overall performance degradation alongside selective gains on spatial and transformation-intensive tasks, while additional analyses identify backbone-related behavioral correlations and examine the effects of verified reference intermediates.

Critical strengths
  • The research question is relevant, and the benchmark provides broad coverage across models and reasoning tasks.
  • The paired comparisons and dual inference protocols offer useful diagnostic perspectives, with appropriate acknowledgment of remaining confounds.
  • The baseline-controlled correlation analysis and reference-intermediate experiments strengthen the empirical study beyond aggregate performance comparisons.
Critical weaknesses
  • RealUnify already compares direct and generation-then-understanding evaluation and reports degradation, while ROVER explicitly evaluates visually augmented answering. The distinction drawn in Section 2 is therefore insufficiently justified.
  • The proposed benefit of transformation-intensive tasks does not adequately explain the contrasting outcomes in Table 14, such as BAGEL’s improvements versus Ovis-U1’s degradation. The study does not yet establish sufficiently precise conditions for when generation helps.
Questions for authors
  • Could the authors clarify what “matched inference budgets” means in Sections 4.2 and 5.2? Please specify the token limits and whether they apply to per call or to the full pipeline, how image generation and re-encoding costs are accounted for, and the maximum rounds and stopping rules for iterative GtA tasks.
3 comments
AuthorsOpenReview ↗

Response to Reviewer AbYh (Part 1/2): Positioning and When Generation Helps

We sincerely thank the reviewer for the careful review, and for recognizing the relevance of the research question, the broad coverage of the benchmark, and the diagnostic value of the paired comparisons, the baseline-controlled correlation analysis and the reference-intermediate experiments. We answer each point below, and all new results are in the revised PDF (changes in blue).

W1: "RealUnify already compares direct and generation-then-understanding evaluation and reports degradation, while ROVER explicitly evaluates visually augmented answering. The distinction drawn in Section 2 is therefore insufficiently justified."

We thank the reviewer for pointing us to this comparison. RealUnify and ROVER both observe that generating an image before answering often lowers accuracy. In both, generation-to-understanding is one part of a broader evaluation; UniG2U is built for this question alone.

Table A. G2U components of RealUnify, ROVER and UniG2U (Appendix T).

RealUnify (GEU)ROVER-TGUniG2U
Itemsabout 4004043,000
Subtasks4630 (7 categories)
Intermediate diagnosticsstepwise evaluationIR, AlignRA, AL, reference intermediates, human ratings
Text-only control, same budgetnonoyes
Blank-image controlnonoyes

This scale lets us test the effect item by item and find the subtasks where generation helps, rather than only report an overall drop. With the two new controls, describing the intermediate state in text scores at least as high as drawing it for every model tested, on all 3,000 items (Appendix P). Section 2 and Appendix T now state this distinction explicitly.

W2: "The proposed benefit of transformation-intensive tasks does not adequately explain the contrasting outcomes in Table 14, such as BAGEL's improvements versus Ovis-U1's degradation. The study does not yet establish sufficiently precise conditions for when generation helps."

The two models move in opposite directions because of the intermediate images they produce. Maze and sliding puzzle show this most clearly.

Table B. Frame accuracy (%), Direct → GtA.

ModelMazeSliding puzzle
BAGEL2.5 → 24.30.0 → 27.8
Ovis-U116.1 → 0.016.6 → 0.8

BAGEL's step images keep the layout of the maze and the 3×3 tile grid of the puzzle, so each step gives the answer stage a usable state. On the same items, Ovis-U1 draws an empty grid at every maze step and a sketch unrelated to the puzzle, and the answer stage follows them.

Human ratings show the same pattern across tasks. Three annotators rated 207 intermediates, drawn with a fixed seed across task groups and shown without model names (Appendix Q).

Table C. Intermediate quality and the effect of GtA.

ModelHuman fidelity (1–5)RA (1–5)GtA − Direct (points)
BAGEL3.082.37+3
Ovis-U11.871.80−17
Janus-Pro1.401.14−11

The condition is therefore precise: generation helps when the model can draw the task state correctly. Figure 23 (Appendix Q) shows, side by side, all four maze and sliding-puzzle items that both models answered in the rated sample.

AuthorsOpenReview ↗

Response to Reviewer AbYh (Part 2/2): Matched Inference Budgets

Q: "Could the authors clarify what "matched inference budgets" means in Sections 4.2 and 5.2? Please specify the token limits and whether they apply to per call or to the full pipeline, how image generation and re-encoding costs are accounted for, and the maximum rounds and stopping rules for iterative GtA tasks."

Within each task, Direct and the answer stage of GtA use the same cap on newly generated text tokens. The cap applies to each answer call: the single Direct call, and the answer call of GtA after its generation stage.

Table D. Per-task cap on newly generated text tokens (Appendix O).

TasksToken cap
ChartQA16
IllusionBench64
RealUnify, VSP, PhyX1,024
MMSI, BabyVision, jigsaw2,048
VisualPuzzles, maze, sliding puzzle4,096
Geometry3K, AuxSolidMath16,384

The generation stage of GtA produces one image per step, and iterative GtA on maze and sliding puzzle produces one image per move. Each image is 1024×1024 with 50 sampling steps for BAGEL, Ovis-U1 and OmniGen2, and 384×384 for Janus-Pro. We account for the cost of generation by measuring what each protocol actually uses per item.

Table E. Measured mean generated text tokens (all stages) and generated images per item.

ModelDirect tokensGtA tokensGtA imagesText-CoT tokens
Ovis-U169831.29239
BAGEL431621.29401
Janus-Pro721140.87298

To separate the cost of the extra step from the content of the image, we added a blank-image condition with the same pipeline and budget.

Table F. Accuracy (%) on 2,354 items.

ModelDirectGtA-blankGtA
Ovis-U136.333.527.8
Janus-Pro31.527.221.6
BAGEL38.939.539.1
OmniGen233.033.331.5

For Ovis-U1, the step with a blank image costs 2.8 points and its own generated image a further 5.7, and Janus-Pro shows the same pattern; for BAGEL, neither lowers accuracy. The same holds on all 3,000 items (Appendix P, Table 22).

We hope these responses address the reviewer's concerns. We would be sincerely grateful if the reviewer would take them into account in the final assessment, and we would be glad to provide any further clarification during the discussion period.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer AbYh,

Thank you again for the time and effort you have devoted to reviewing our paper. We have revised our responses to make them more precise, with each of your questions quoted and answered in turn. The responses cover the comparison with RealUnify and ROVER, the conditions under which generation helps, and the exact token caps and measured generation costs.

We would be grateful if you could take a look at the updated responses at your convenience and let us know whether they address your concerns. If any question remains, we would be very happy to discuss it further during the discussion period. If our responses have resolved your concerns, we would sincerely appreciate it if you would consider updating your assessment.

Thank you again for your effort and your constructive feedback.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions

Dear Reviewers and Area Chair,

We sincerely thank all three reviewers for their detailed and constructive comments. We are encouraged that the reviewers found the research question relevant and well motivated (AbYh, sHgY), the separation of unified training from inference-time generation useful for diagnosing where generation helps understanding (3yav, sHgY), and the empirical analysis broad and insightful (AbYh, sHgY, 3yav).

We have uploaded a revised PDF; changes are in blue, and the new analyses are in Appendices O–U. All paired comparisons are reported there with bootstrap 95% confidence intervals and McNemar tests with Holm correction.

1. Text-only and blank-image controls (sHgY, 3yav, AbYh). Both new conditions keep the GtA instructions, step structure and token cap. Text-CoT writes the intermediate state in words; GtA-blank replaces the generated image with a blank one. Accuracy (%) on 2,354 items (all tasks except the four scored by an LLM judge: Geometry3K, AuxSolidMath, BabyVision and PhyX):

ModelDirectText-CoTGtA-blankGtA
Ovis-U136.333.533.527.8
Janus-Pro31.527.627.221.6
OmniGen233.033.433.331.5
BAGEL38.940.539.539.1

Text-CoT scores at least as high as GtA for every model, and clearly higher for Ovis-U1 and Janus-Pro. For Ovis-U1, adding the step with a blank image lowers accuracy from 36.3 to 33.5, and putting its own generated image in that step lowers it further to 27.8; Janus-Pro shows the same pattern (31.5, 27.2 and 21.6) (Appendix P). On all 3,000 items, including the four LLM-judged tasks, Text-CoT remains at least as high as GtA for every model (Appendix P, Table 22).

2. When generation helps (AbYh, sHgY). Three annotators rated 207 intermediates, drawn with a fixed seed across task groups and shown without model names (Krippendorff's α = 0.61):

ModelHuman fidelity (1–5)RA (1–5)GtA − Direct (points)
BAGEL3.082.37+3
Ovis-U11.871.80−17
Janus-Pro1.401.14−11

The model whose intermediates preserve the task state is the one that gains from GtA. The paper's RA metric agrees with the annotators (Spearman ρ = 0.64).

3. Gradient sharing (sHgY).

ModelHow generation and understanding share parametersGeneration / understanding gradient normEditing / text-to-image
Janus-Pro-7Bwhole backbone shared0.090– (no editing)
Ovis-U1LLM conditions a separate generator0.00491.0×
BAGELseparate experts; the input image is encoded by the understanding expert0.00196.8×

The more parameters understanding and generation share, the more strongly generation training acts on understanding, and editing shares more with understanding than text-to-image generation when the input image passes through the understanding pathway (Appendix R).

4. Budget (AbYh, 3yav). Direct and the GtA answer stage share the same token cap within each task; Appendix O lists the caps and the tokens and images actually used.

5. Prompts (3yav). Appendix U gives the exact prompts executed. K.4 and K.1.2 were documentation errors and are corrected. MSR was rerun with one instruction shared by all conditions.

6. Pairing (3yav). Appendix S gives a full pairing table with matching tiers. OmniGen2 (built on Qwen2.5-VL-3B) and OneCAT-3B (initialized from its language model, Qwen2.5-3B-Instruct) are now compared with Qwen2.5-VL-3B. Every Tier-1 model with Direct results remains below its comparator.

7. Positioning (AbYh). Appendix T compares UniG2U with RealUnify and ROVER.

8. Data. The human ratings, images and analysis script: https://anonymous.4open.science/r/unig2u-human-ratings-046B

We are grateful for the reviewers' suggestions, which have clearly strengthened the paper, and we warmly welcome any further questions during the discussion period.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.