acceptodds
Under review as a conference paper at ICLR 2027

Beyond Verbal Reasoning: Visual Structure Reasoning Language for Image Generation in Unified Multimodal Models

Abstract

Recent unified multimodal models (UMMs) have begun to introduce explicit natural-language reasoning before image generation. We revisit this default choice and propose Visual Structure Reasoning Language (VSRL), a structured reasoning language for image generation in UMMs. Given a prompt, the UMM autonomously generates VSRL comprising the global scene, objects, and their relations, and conditions image generation on it. We develop a VSRL data construction pipeline that converts paired prompts and target images into verified prompt–VSRL–image triplets. The resulting data are used to train the UMM to generate and use VSRL through supervised fine-tuning. We further introduce StruFlow-GRPO, a reinforcement learning framework that jointly optimizes reasoning and generation with programmatic structure rewards and image rewards. On BAGEL, VSRL outperforms both the base model and BAGEL’s native natural-language reasoning (Self CoT) across five generation benchmarks. Compared with Self CoT, VSRL improves GenEval Short from 79.0 to 88.9 and OneIG-Bench from 32.4 to 43.7. On Cheers, we train two separate models on VSRL and natural-language CoT data for a controlled comparison. VSRL outperforms both natural-language CoT and the base model across all evaluated generation benchmarks, including GenEval and TIIF-Bench. These results highlight reasoning representation as an important design dimension for UMM image generation and the potential of generation-oriented structured reasoning across UMMs.

Primary area:
generative models
Keywords:
unified multimodal models, image generation, structured reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_a959OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work replaces the open-ended textual reasoning that normally precedes UMM image generation with VSRL, a rigid structured representation covering the global scene, object IDs and attributes, plus their pairwise relations. The authors build triplets of prompt, VSRL and corresponding image, then supervise-finetune the UMM to output VSRL and draw images conditioned on it. They further train with StruFlow-GRPO, using programmatic structure rewards alongside image-level rewards. On the BAGEL benchmark suite, VSRL outperforms both the base model and vanilla Self CoT across five generation benchmarks; for instance, GenEval Short jumps from 79.0 to 88.9, and OneIG rises from 32.4 to 43.7. On Cheers, they train VSRL and regular natural-language CoT on identical underlying facts using the exact same training setup, and VSRL comes out ahead consistently on GenEval, TIIF and DPG.

Critical strengths

The Cheers control experiment is well conceived. It uses identical prompts, target images, initialization and training schedule; the natural-language chain-of-thought text is simply transcribed from the same factual content used for VSRL. This makes the performance gap difficult to write off as just better training data, and that is exactly why the representation argument lands for me.

Object IDs together with explicit relation anchors offer a practical fix for common CoT breakdowns: instances from the same category getting merged together and spatial descriptions drifting. The improvements seen on GenEval Short and benchmarks heavy on counting and relational reasoning align directly with this design choice, rather than being a generic boost from reinforcement learning.

Critical weaknesses

Credit assignment on BAGEL is unclear. The baseline Self CoT already went through its own training pipeline, while VSRL adds fresh SFT data plus StruFlow-GRPO on top. I would like to see ablations at matched training steps: VSRL with only SFT, then SFT plus image reward, then SFT plus both image and structure rewards. Without this breakdown, Table 1 risks overstating gains from the representation itself when they may simply come from extra optimization.

There is also very little evaluation of VSRL outputs at inference time. We do not know how frequently the model lists an object that never appears in the final image, or how often it states spatial relations backwards. If the structure reward only pushes for valid, complete parses, the model could produce clean VSRL schemas that the image generator still ignores. A small consistency check across prompt, VSRL and generated image would make the whole mechanism far more credible.

Questions for authors

We request ablations on a single backbone with identical update counts: VSRL under SFT alone, SFT plus only the image reward, and the full StruFlow-GRPO setup.

We also want breakdowns for three failure modes on test cases: (a) VSRL fails to include an object specified in the prompt, (b) the final image leaves out an object listed in VSRL, and (c) stated relations do not match the image. We need to see whether RL mainly cuts down on (b) and (c), or mostly cleans up (a).

For the Cheers NL-CoT control, please verify that its reasoning template does not inject bounding boxes or coordinate information unavailable to VSRL, and that the maximum reasoning token budget and image decoding parameters are matched across both conditions.

4 comments
AuthorsOpenReview ↗

Q1/W1: RL Reward Ablation

Reviewer 3 (a959)

We appreciate your detailed and constructive questions and your help in identifying gaps in the experimental evaluation in the original manuscript. Following your suggestions, we conducted an additional RL reward ablation across multiple benchmarks, as well as a Prompt–VSRL–Image consistency analysis. Your suggestions have helped us better understand and evaluate the method. We respond to your comments below.

Q1/W1: RL Reward Ablation

Thank you for your suggestion regarding the attribution of performance gains on BAGEL. We agree that clarifying the contributions of different optimization factors helps us better assess VSRL's effectiveness.

Therefore, we conducted an additional 200-step RL ablation using only image rewards on BAGEL. The image-reward-only variant and full StruFlow-GRPO start from the same VSRL-SFT model and use the same trainable modules, number of training steps, and other RL settings, differing only in the reward components. Both RL variants are trained for 200 steps, while VSRL-SFT serves as their common initialization without additional RL updates.

We compared the methods on four generation benchmarks with five evaluation settings: GenEval, DPG, TIIF Short/Long, and OneIG. The results are as follows:

EvaluationVSRL-SFTImage-Reward-Only RL200 (vs. SFT)Full StruFlow-GRPO (vs. SFT)
GenEval88.4087.34 (−1.06)88.92 (+0.52)
DPG84.6284.96 (+0.34)85.20 (+0.58)
TIIF Short70.8871.23 (+0.35)71.90 (+1.02)
TIIF Long71.0371.65 (+0.62)72.38 (+1.35)
OneIG Mean43.2543.61 (+0.36)43.74 (+0.49)

We introduced structure rewards so that optimization considers both the quality of the final generated image and the structural quality of the VSRL generated by the model. Specifically, structure rewards evaluate aspects such as VSRL's structural validity, coverage of object categories required by the prompt, and structural richness. Together with image rewards, they contribute to trajectory returns and model parameter updates, allowing the model to use feedback on intermediate VSRL to optimize toward structures better suited to guiding image generation.

The results show that, under the same initialization and training budget, full StruFlow-GRPO outperforms image-reward-only RL in all five evaluation settings. This further indicates that the structural quality of VSRL is an important factor affecting final image generation, and that including it in optimization can lead to better generation performance.

We will add this RL reward ablation and its settings to the existing Appendix A.2, “Training Stages and Optimization Scope,” in the revised manuscript.

AuthorsOpenReview ↗

Q2/W2: Prompt–VSRL–Image Consistency Analysis and Q3: Information Consistency and Evaluation Settings for the Cheers Natural-Language Conversion

Q2/W2: Prompt–VSRL–Image Consistency Analysis

We appreciate your suggestion to evaluate VSRL reasoning quality, as this evaluation is important for assessing how effectively VSRL guides image generation. Therefore, following your suggestion, we conducted an additional Prompt–VSRL–Image consistency analysis.

We compared VSRL-SFT and VSRL-StruFlow-GRPO on GenEval Short and DPG, measuring the following three types of errors:

  • (a) Prompt-to-VSRL object omission: objects required by the prompt but omitted from VSRL.
  • (b) VSRL-to-image object omission: objects specified in VSRL but missing from the generated image.
  • (c) VSRL-to-image relation mismatch: relations specified in VSRL but not correctly realized in the generated image.

In this evaluation, each sample consists of the original prompt, the VSRL generated by the model, and the final generated image. We use Qwen3.8-27B as the judge and measure these three types of errors at the level of individual Prompt–VSRL–Image rollouts.

The results are shown below. Parentheses indicate relative reductions compared with SFT.

MethodGenEval a ↓GenEval b ↓GenEval c ↓DPG a ↓DPG b ↓DPG c ↓
VSRL-SFT0.0000%0.3948%3.5035%2.3308%1.8952%9.1005%
VSRL-StruFlow-GRPO0.0000% (unchanged)0.3130% (↓20.73%)3.4481% (↓1.58%)2.2627% (↓2.92%)1.5292% (↓19.31%)8.9957% (↓1.15%)

The results show that the improvements from RL are not limited to object coverage in VSRL relative to the prompt (a); we also observe reductions in the rates of missing objects in the generated images (b) and relation mismatches (c), the two issues you specifically raised. These results indicate that StruFlow-GRPO not only helps VSRL describe the objects required by the prompt more completely, but also improves how well the final image matches the objects and relations specified in VSRL, thereby improving VSRL's ability to guide image generation.

We will add a new subsection, “Prompt–VSRL–Image Consistency Analysis,” to Appendix A of the revised manuscript, including the full results, evaluation procedure, and statistical details.

Q3: Information Consistency and Evaluation Settings for the Cheers Natural-Language Conversion

We apologize for not explaining these experimental details clearly enough in the original manuscript and for the resulting confusion.

First, we confirm that the natural-language CoT conversion templates for Cheers do not add bounding boxes or coordinates that are not present in the original VSRL. As shown in the original Appendix C.1, “Control Construction,” and C.2, “Paired Supervision Example,” we use fixed templates and connecting words to convert the objects, attributes, and relations in VSRL into natural language, while preserving the object references at both endpoints of each relation and adding no visual information absent from the original VSRL.

Second, in the Cheers comparison between VSRL and natural-language CoT, all inference settings used in the evaluation are kept the same, including the maximum number of reasoning tokens and the image decoding settings you mentioned. We apologize for not making this clear in the original manuscript. We will clarify this in the existing Appendix D, “Supervised Fine-Tuning and Inference Configurations,” and explicitly report the corresponding settings.

Thank you again for your careful review. Your suggestions have not only helped us add the necessary ablation and consistency experiments, but also encouraged us to explain more clearly how VSRL is optimized and how it affects final image generation. This has helped us improve the paper.

Great, I tend to accept this paper due to authors meaningful feedback and solution.

Thank you for your encouraging feedback and positive assessment of our work. We are delighted that our additional experiments and clarifications have helped address your concerns, and we are truly encouraged by your support.Your constructive suggestions have also helped us strengthen the experimental evidence and clarify our method and experimental setup. We will incorporate these additional results and details into the revised manuscript.

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

The paper proposes VSRL, a structured scene description that a UMM generates before producing an image. It combines supervised training with StruFlow-GRPO and evaluates the approach on BAGEL and Cheers.

Critical strengths
  1. VSRL improves GenEval Short from 80.00 to 84.30 and TIIF Short from 64.00 to 67.13. This makes a convincing case that the representation itself can help, rather than just the addition of more descriptive information.

  2. On BAGEL, VSRL with supervised training alone raises GenEval Short from about 77.9 to 88.4, with large improvements in spatial positioning and color attribution. The model generates its own structure without an external planner. Together with the Cheers results, this makes the approach worth considering beyond a single model or benchmark.

Critical weaknesses
  1. The BAGEL comparison does not isolate the effect of VSRL. It also adds 278,970 training examples, bounding box supervision, and RL. These changes could explain part of the gains. The Cheers comparison helps, but it uses much less data and no boxes. It remains unclear how much VSRL itself improves BAGEL.
Questions for authors
  1. Do you have an existing BAGEL comparison in which natural language reasoning receives the same training examples and visual information, including boxes? Such a result would directly affect my assessment of the representation's contribution.
3 comments
AuthorsOpenReview ↗

Contribution of the VSRL Reasoning Representation on BAGEL and a Same-Data Natural-Language Comparison (W1, Q1)(Part 1)

Reviewer 2 (JeDc)

We thank you for your positive assessment of our work and for your detailed and constructive suggestions. Your questions help us clarify the contribution of the VSRL reasoning representation itself.

Contribution of the VSRL Reasoning Representation on BAGEL and a Same-Data Natural-Language Comparison (W1, Q1)

We initially used BAGEL's native Self CoT as the natural-language reasoning baseline because BAGEL had already learned natural-language reasoning from approximately 500k reasoning-enhanced samples. For VSRL, the model needed to learn a different representation, including structured fields, object identities, and relation references, from 278,970 additional structured samples. It also needed to learn how to use these structures to guide image generation. Therefore, we compared the native Self CoT baseline with the VSRL-trained model, particularly given that VSRL uses dedicated trigger markers and a reasoning representation that are entirely different from those of native Self CoT, to examine the generation performance of this structured reasoning approach.

However, we also recognize that BAGEL's native Self CoT and VSRL differ in their supervision sources and training histories, so this comparison cannot fully isolate the contribution of the reasoning representation itself. For this reason, in the original manuscript we also selected Cheers as another backbone and trained separate models from the same initialization, using VSRL and natural-language CoT supervision containing the same visual information. This allows us to make a rigorous comparison and examine the effect of the intermediate reasoning representation more directly.

If we directly convert all VSRL supervision data into natural-language CoT and continue training on BAGEL, the additional training examples and visual information can be kept the same, but this comparison still cannot remove the influence of BAGEL's previous training history: a model trained with natural-language CoT can build on the reasoning format BAGEL has already learned and continue training with approximately 279k additional examples containing visual information; by contrast, a model trained with VSRL must learn new structured representation rules from the same examples, as well as how to use this representation to guide image generation. Thus, although both models start from the same checkpoint, the pretrained model has already learned natural-language reasoning but still needs to learn the VSRL format.

Nevertheless, the experiment you suggested offers a challenging and valuable comparison for VSRL that can provide further evidence for the contribution of the reasoning representation. Therefore, following your suggestion, we conducted this additional experiment on BAGEL.

Specifically, we converted all 278,970 examples used for VSRL training into natural-language CoT using fixed templates, fully preserving the scene, objects, attributes, relations, and bounding-box information in the original VSRL. We then trained the corresponding BAGEL + NL CoT-SFT model starting from the same native BAGEL checkpoint.

In this experiment, we focus on comparing NL CoT-SFT with VSRL-SFT, without introducing the additional StruFlow-GRPO stage, to exclude the effect of RL on the comparison and examine the contribution of the reasoning representation itself more directly.

The results are as follows:

MethodGenEval ↑TIIF-S ↑TIIF-L ↑OneIG Mean ↑CVTG Word Acc. (×100) ↑
BAGEL + NL CoT-SFT87.0969.1670.7541.5336.26
BAGEL + VSRL-SFT88.40 (+1.31)70.88 (+1.72)71.03 (+0.28)43.25 (+1.72)38.57 (+2.31)

The other CVTG-2K metrics are as follows:

MethodNED ↑CLIPScore ↑VQAScore ↑Aesthetic ↑
BAGEL + NL CoT-SFT0.66810.79110.85654.7980
BAGEL + VSRL-SFT0.6821 (+0.0140)0.7970 (+0.0059)0.8739 (+0.0174)4.9663 (+0.1683)

The results show that, even though the model needs to learn the new VSRL representation format in this comparison, VSRL-SFT achieves better performance in all the reported evaluation settings across GenEval, TIIF, OneIG, and CVTG. These results further indicate that VSRL's gains do not depend entirely on additional training information or RL optimization, and that the structured reasoning representation itself is also an important factor affecting final image-generation performance.

AuthorsOpenReview ↗

Contribution of the VSRL Reasoning Representation on BAGEL and a Same-Data Natural-Language Comparison (W1, Q1)(Part 2)

We will add the same-data natural-language comparison on BAGEL to the existing Appendix A.1, “Reasoning Representation Comparison,” alongside the original comparison on Cheers. We will also add the BAGEL natural-language conversion procedure and paired supervision examples to Appendix C, “Natural-Language CoT Construction,” and describe the corresponding training configurations in Appendix D, “Supervised Fine-Tuning and Inference Configurations.”

Thank you again for your careful review and suggestions, which have helped us strengthen the experimental evidence for the contribution of the reasoning representation on BAGEL.

AuthorsOpenReview ↗

Follow-up on Our Response and Additional Experiments

Thank you again for your careful review and constructive suggestions.

As the author–reviewer discussion period is drawing to a close, we sincerely hope that our detailed response and additional experiments, particularly the information-matched comparison with natural-language CoT on BAGEL, have helped clarify our work and address your concerns.

If you have any remaining questions, or if any aspect of our method or experiments would benefit from further clarification, please do not hesitate to let us know. We would be very happy to respond as promptly as possible during the remaining discussion period.

We sincerely appreciate your time and consideration, and hope that the additional evidence has further strengthened your confidence in our work.

Reviewer_uM7vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Visual Structure Reasoning Language (VSRL), an intermediate representation for image generation in unified multimodal models. VSRL organizes the target scene into global information, explicitly identified objects and their attributes, and relations referencing those objects. The model first generates this structure and then conditions image generation on it. The authors construct prompt–structure–image supervision for supervised fine-tuning and introduce StruFlow-GRPO, which combines image and programmatic structure rewards while updating language and image-generation modules through an image-policy objective. Experiments on BAGEL and Cheers report improvements across multiple generation benchmarks. A controlled comparison on Cheers uses the same scene information and training settings to compare VSRL with natural-language supervision.

Critical strengths
  1. Controlled representation comparison. The Cheers experiment matches scene information and training conditions between VSRL and natural-language CoT, providing useful evidence that the intermediate representation affects generation quality.

  2. Consistent practical gains. VSRL improves generation across BAGEL and Cheers, with particularly strong gains in spatial relations and attribute binding while retaining the native generation pipeline.

Critical weaknesses
  1. Unclear optimization of structured reasoning. The RL objective only includes image-policy ratios conditioned on sampled VSRL (Eqs. 10–12). Updating language-side conditioning representations does not establish that the model learns to generate better structures.

  2. Limited evidence on reasoning-intensive generation. OneIG Reasoning improves only from 20.60 to 20.79 over Self CoT (Table 7). Evaluation on reasoning-intensive benchmarks such as GIR-Bench-T2I[1] and T2I-CoReBench[2] would better establish benefits beyond organizing and following explicit scene constraints.

[1] GIR-Bench: Versatile Benchmark for Generating Images with Reasoning [2] Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play?

Questions for authors
  1. Does StruFlow-GRPO include a text-policy objective omitted from the paper?

  2. Could the authors compare the correctness of generated VSRL before and after RL, using paired examples and, ideally, quantitative statistics?

5 comments
AuthorsOpenReview ↗

1.How StruFlow-GRPO Optimizes VSRL and Quantitative Analysis of VSRL Correctness Before and After RL (W1, Q1, Q2)(Part 1)

Reviewer 1 (uM7v)

We thank you for your positive assessment of our work and for your thoughtful and constructive questions. We respond to your comments below.

1. How StruFlow-GRPO Optimizes VSRL and Quantitative Analysis of VSRL Correctness Before and After RL (W1, Q1, Q2)

We apologize for not sufficiently explaining this optimization mechanism in the original manuscript. As you correctly pointed out, StruFlow-GRPO does not include an additional text-policy loss. In our framework, the generated VSRL is included in joint optimization as follows: The structure reward in Eq. (5) contributes to the total trajectory reward in Eq. (6), which is used to compute the group-relative advantage in Eq. (9). This advantage then enters the Flow-GRPO objective in Eq. (12).

The policy ratios in Eqs. (10)–(11) are computed from the image denoising process conditioned on VSRL. Gradients from the image-policy objective are propagated through the VSRL conditioning representations, allowing the parameters of both the language and image-generation modules to be updated (§3.4 and Appendix E.2).

We designed the method this way because VSRL is intended to support image generation. Therefore, rather than introducing an independent text-policy objective, we aim to use feedback from the image-generation task to jointly adapt the structural representations and image-generation process, so that the model learns conditioning representations better suited to subsequent image synthesis. We will further clarify this optimization mechanism and its gradient propagation path in Appendix E.2, “Image-Policy Objective and Joint Updates,” of the revised manuscript.

However, we also agree with your key concern: updating the language-side conditioning representations alone does not directly prove that the model generates more correct VSRL structures after RL. Therefore, to further examine the effect of this optimization mechanism on actual generation and address your question about changes in VSRL correctness before and after RL, we conducted an additional Prompt–VSRL–Image consistency analysis.

We compared VSRL-SFT and VSRL-StruFlow-GRPO on GenEval Short and DPG. Each evaluation sample consists of the original prompt, the VSRL generated by the model, and the final generated image. We use Qwen3.8-27B as the judge to evaluate three types of errors:

  • (a) Prompt-to-VSRL object omission: objects required by the prompt but omitted from VSRL.
  • (b) VSRL-to-image object omission: objects specified in VSRL but missing from the generated image.
  • (c) VSRL-to-image relation mismatch: relations specified in VSRL but not correctly realized in the generated image.

The results are shown below. Parentheses indicate relative reductions compared with SFT.

MethodGenEval a ↓GenEval b ↓GenEval c ↓DPG a ↓DPG b ↓DPG c ↓
VSRL-SFT0.0000%0.3948%3.5035%2.3308%1.8952%9.1005%
VSRL-StruFlow-GRPO0.0000% (unchanged)0.3130% (↓20.73%)3.4481% (↓1.58%)2.2627% (↓2.92%)1.5292% (↓19.31%)8.9957% (↓1.15%)

As the results show, after RL optimization, VSRL improves in both object coverage relative to the prompt and its ability to guide subsequent image generation. On GenEval Short, VSRL maintains zero object omissions, while on DPG the object omission rate decreases further.

The main improvement is in how effectively VSRL guides image generation: on both benchmarks, generated images follow the object and relation constraints specified in VSRL more closely. In particular, the object omission rate in generated images decreases by 20.73% and 19.31% on GenEval Short and DPG, respectively, while the relation mismatch rate also decreases.

AuthorsOpenReview ↗

2. Additional Evaluation on T2I-CoReBench (W2)(Part 2)

To illustrate the change in VSRL correctness before and after RL more directly, we selected a paired example from GenEval. Because the complete VSRL is long, we show only obj_1, the object with the most relevant changes, while preserving the original fields and values generated by the model.

For the same prompt:

“a photo of a cow left of a stop sign”

VSRL-SFT (obj_1):

{
  "id": "obj_1",
  "name": "cow",
  "position": "right foreground",
  "color": "black and white",
  "pose": "standing facing the camera",
  "head": "with curved horns",
  "bbox": [
    456,
    272,
    907,
    997
  ]
}

VSRL-StruFlow-GRPO (obj_1):

{
  "id": "obj_1",
  "name": "cow",
  "position": "left foreground",
  "color": "black and white",
  "pose": "standing facing the camera",
  "state": "calm",
  "bbox": [
    32,
    298,
    598,
    997
  ]
}

Before RL, VSRL describes the cow as being in the right foreground, whereas after RL, its position is corrected to the left foreground, and the corresponding bounding box also moves from the right side of the image to the left.

This paired example directly shows an improvement in how the generated VSRL describes spatial position after RL, making the description more consistent with the original prompt.

We will add a new subsection, “Prompt–VSRL–Image Consistency Analysis,” to Appendix A of the revised manuscript, reporting the quantitative analysis, metric definitions and aggregation procedure, and paired examples before and after RL.

2. Additional Evaluation on T2I-CoReBench (W2)

Thank you for suggesting that we evaluate VSRL on tasks that require more reasoning.

VSRL was originally designed as a visual reasoning language for image generation, so the evaluations in our initial submission mainly focused on visual scene organization and how well generated images satisfy the specified constraints. Benchmarks such as GIR-Bench-T2I and T2I-CoReBench place more emphasis on complex semantic reasoning, and we therefore did not include them in our initial evaluation.

However, we believe that visual structure reasoning for image generation and semantic reasoning are not in conflict; instead, they can be combined to support more complex image-generation tasks. Your suggestion has encouraged us to evaluate VSRL on tasks that require more semantic reasoning.

Following your suggestion, we also evaluated VSRL on T2I-CoReBench to examine its performance on such reasoning tasks. The results are as follows:

MethodC Mean ↑R Mean ↑Overall ↑
BAGEL51.8136.6441.70
BAGEL Self CoT44.9145.4245.25
VSRL-StruFlow-GRPO51.5945.5647.57

The results show that VSRL-StruFlow-GRPO achieves an Overall score of 47.57, exceeding BAGEL and BAGEL Self CoT by 5.87 and 2.32 points, respectively.

These results provide additional evidence of VSRL's effectiveness on an image-generation benchmark with substantial semantic reasoning demands, and also suggest a useful direction for future work combining visual structure reasoning with semantic reasoning.

We will add the complete T2I-CoReBench results to the section on supplementary benchmark evaluations in Appendix A and provide the corresponding evaluation settings in Appendix F, “Evaluation Protocols and Result Sources,” of the revised manuscript.

Thank you again for your valuable suggestions, which have helped us improve both the experimental evaluation and the explanation of our method.

Thank you for clarifying the RL objective and adding the consistency analysis, paired example, and T2I-CoReBench results. These additions partially address my concerns. However, the reasoning mean increases only from 45.42 to 45.56 compared with BAGEL Self CoT. I would therefore appreciate category-wise comparisons on GIR-Bench-T2I. Its numerical and spatial reasoning tasks require inferring the correct object counts and layouts from given constraints, providing a focused test of whether VSRL improves the reasoning needed to determine the intended scene structure. This is directly relevant to the paper’s reasoning claims. Comparisons with BAGEL Self CoT under consistent settings, limited to the T2I subset, would be sufficient.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.