acceptodds
Under review as a conference paper at ICLR 2027

PostCrop: Compose Before Crop for Context-Aware Fine-Grained Visual Perception

Abstract

Large vision-language models (LVLMs) have demonstrated strong fine-grained perception of relations among multiple entities by controlling visual context—expanding or reducing the surrounding context of multiple entities (e.g., zoom-out or locate-then-zoom-in). While existing training-free methods typically rely on either a coarse-to-fine search paradigm or a Crop Before Compose paradigm that crops entity regions before composing them with context, both face a context paradox: crops containing only the entity regions can obscure relations among entities, whereas crops keeping context introduce visual interference. To characterize this paradox, we fix the entity regions and vary the kept context. Across HR-Bench-4K and HR-Bench-8K, we find that the preferred amount of context varies across settings, motivating adaptive context selection. Further analysis shows that the early separation of entity regions in the Crop Before Compose paradigm makes this adaptation difficult. Leveraging these insights, we propose PostCrop, a training-free framework following the Compose Before Crop paradigm that considers relation-relevant context before the final crop is formed. The Evidence Localization and Composition module localizes the region that best matches each queried entity and composes the cue maps of multiple entities on the original visual grid to identify surrounding regions that support the queried relation. The Evidence Cropping and Answering module uses the composed map to determine how much surrounding context to retain and selects a compact crop containing the entity regions and the kept context, which is combined with the original image for answering. Extensive experiments on VSTAR, HR-Bench-4K, and HR-Bench-8K demonstrate that PostCrop consistently outperforms the compared baselines across all three evaluated LVLMs (e.g., 74.0% accuracy on the Fine-grained Cross-instance Perception (FCP) subtask of HR-Bench-4K with Qwen2.5-VL-7B), validating the effectiveness of the Compose Before Crop paradigm.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Vision-Language Models, Fine-Grained Visual Perception, Spatial Reasoning, Adaptive Visual Context Selection, Training-Free Inference

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_aZWoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a fine-grained perception method aimed at preserving relative spatial arrangements (such as left-to-right and top-to-bottom spatial relations) among multiple objects.

Critical strengths
  1. The motivation is reasonably sound, and the experimental results demonstrate the empirical effectiveness of the proposed approach.
  2. The manuscript is clearly written and easy to follow.
Critical weaknesses
  1. Flawed foundational premise compared to training-based methods: ZwZ represents a typical training-based paradigm that requires zero image-level manipulation/cropping during inference and naturally preserves the full contextual information of the scene. Consequently, it fundamentally sidesteps the core motivation claimed in this paper, namely, that excessive visual context inherently degrades fine-grained perception.
  2. Questionable necessity & trade-offs on general capabilities: The necessity of training-free methods remains unconvincing. Compared to training-based approaches, training-free interventions often lead to severe performance drops on general multimodal benchmarks (e.g., MMStar), a trade-off that is not properly evaluated here.
  3. Missing perspective against frontier models: How do state-of-the-art closed-source multimodal foundation models perform on these fine-grained perception benchmarks? Given their rapid progress, is such a training-free pipeline still relevant or practically justifiable?
  4. Missing latency/runtime benchmarks: A rigorous comparison of inference latency and computational overhead against training-based methods like ZwZ is noticeably absent.
Questions for authors
  1. What are the concrete advantages of the proposed method when compared directly with training-based approaches like ZwZ (in terms of accuracy, generalizability, and deployment cost)?
  2. Considering the significant per-sample inference latency typical of such methods, could the authors clarify the target real-world application scenarios where this approach would remain practical?
5 comments

We thank the reviewer for the thoughtful and constructive feedback. We would like to address each concern below. We have uploaded the rebuttal revision.

We compare PostCrop with ZwZ in terms of motivation, performance, and practical significance.

W1: How does the motivation of PostCrop compare with that of ZwZ?

  • First, the main motivation of the original manuscript is not simply that "the more context there is, the worse the model performs." Rather, a more complete view is that both too much and too little context can be harmful. The key is how to select appropriate context.

  • Second, our motivation remains valuable compared with the ZwZ paradigm. We acknowledge that our motivation is essentially about cropping, or, more specifically, better cropping. In fact, the ZwZ paradigm also relies on crop quality during training, because good crops can produce higher-quality synthetic data. PostCrop can contribute to this, and the two works are compatible.

To verify the complementarity between training-free visual localization methods and trained fine-grained perception methods, we use PostCrop-7B (Qwen2.5-VL-7B) to locate and crop key regions in the training images. We introduce the resulting high-quality crops into the Region-to-Image Distillation (R2I) training process of ZwZ-7B to provide more precise regional visual supervision, and train ZwZ-7B-postcrop using the same training configuration.


Q1/W4: How does PostCrop compare with ZwZ in terms of performance?

In the first table below, we compare ZwZ-7B and ZwZ-7B-postcrop in terms of accuracy on VSTAR, computational cost, inference latency, and generalization ability. On V*, compared with ZwZ-7B, the model trained with PostCrop crops is expected to achieve higher fine-grained perception accuracy while maintaining roughly the same inference latency and memory usage as ZwZ-7B.

ModelV* Accuracy (191 questions)Average Evaluation LatencyPeak Memory
ZwZ-7B baseline70.16%0.315 sec/question15.64 GB
ZwZ-7B-postcrop70.68%0.337 sec/question15.66 GB

In the second table below, we compare ZwZ-7B and PostCrop-7B in terms of accuracy on VSTAR, computational cost, inference latency, and performance on MMStar, which reflects generalization ability. PostCrop brings higher computational cost and latency on VSTAR, and its generalization to general tasks on MMStar is somewhat weaker. However, PostCrop substantially improves fine-grained perception while largely preserving the model's basic capabilities.

ModelV* AccuracyAverage Evaluation LatencyPeak MemoryMMStar Accuracy
ZwZ-7B baseline70.16%0.315 sec/question15.64 GB61.80%
PostCrop-7B94.20%9.178 sec/question19.10 GB56.13%

Q2: What is the practical significance of PostCrop?

PostCrop aims to produce better crops and improve the perception of fine details. As shown in Figure 5, it controls the influence of context within a crop, particularly for multi-object perception involving depth relations and tables. This may make it suitable for industrial scenarios with high-resolution images and complex visual elements. Fine-grained perception should not rely on localization alone. Other visual elements also need to be considered because correct localization does not necessarily lead to a correct answer.


W2: How does PostCrop perform on MMStar?

PostCrop's performance on MMStar decreases slightly, but its fine-grained and coarse-grained perception abilities show the expected improvements. The specific reasons are provided in Appendix Table 13 of the manuscript.


W3: How does PostCrop compare with closed-source models?

For a comprehensive evaluation, we selected two representative closed-source models, GPT-4o and OpenAI o3, as shown in the table below.

MethodTraining freeVSTAR Attr.VSTAR SpatialVSTAR OverallHR-Bench-4K FSPHR-Bench-4K FCPHR-Bench-4K OverallHR-Bench-8K FSPHR-Bench-8K FCPHR-Bench-8K Overall
GPT-4o———66.070.048.059.062.049.055.5
OpenAI o3———95.7——————

Dear Reviewer aZWo,

Thank you for your valuable feedback. We have carefully addressed your concerns in our rebuttal.

As the discussion period is coming to an end, we would greatly appreciate it if you could review our responses and let us know if you have any remaining concerns.

Thank you for your time and consideration.

Best regards,

The authors of Paper Submission40345

Reviewer_aZWoOpenReview ↗

Feedback

  1. What evidence supports the claim that both too much and too little context can be harmful?

  2. The 70.16% result reported for the ZwZ-7B baseline is substantially below the performance reported by the ZwZ authors. As such, it does not seem to be a reliable representation of ZwZ. If this is the performance obtained under your setup, the original model may be a more appropriate baseline.

  3. The claim that you trained a model using the ZwZ recipe with crops generated by PostCrop raises several questions. What evidence shows that PostCrop produces better crops? A 0.5% gain seems too small to distinguish from run-to-run variation. More importantly, the time required to construct the training data should be considered: at scale, even a modest increase in per-example processing time could lead to substantial overhead. How many GPUs did you use, and how long did training take? I am also unsure whether the evaluation sufficiently addresses practical utility and the potential value of this work to the community.

  4. Based on the current results, training-free methods appear to be substantially less effective than trained methods, which raises questions about their practical viability.

  5. In light of the issues raised in the rebuttal, I am considering lowering my score. Some of the information provided could give readers an inaccurate impression of the state of the field, particularly those unfamiliar with related work. I find this concerning. At the same time, I appreciate that the authors have acknowledged some limitations of their experiments, especially regarding efficiency. This leaves me conflicted: I might view the work more favorably if its claims and contribution were framed more narrowly.

Reviewer_ujQnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PostCrop, a training-free method to improve fine-grained visual perception. It first constructs a composite image that includes all referred objects and keeps the spatial relations among them. Compared to existing crop-before-compose methods, this compose-before-crop method preserves the object relations in the input images as auxiliary information, which further improves the prediction accuracy. The pipeline includes searching for object evidence via a middle attention map, VLM verification, image composition, and final cropping and answering with composited layout information.

Critical strengths
  1. The motivation is straightforward and the method is clear to understand.
  2. The results show that the proposed method achieves consistent gains across different backbones and benchmarks.
  3. The qualitative results further prove the effectiveness of the proposed method.
Critical weaknesses
  1. The motivation is good, but the novelty of the proposed method is very limited. [1] already proposes an attention-guided visual intervention method to help change the intermediate state. Actually, most recent papers have already utilized the attention map to localize the related objects and then applied further training-free calibration; thus, this paper's proposal is not new. Besides, [2] already combines semantic decomposition, attention-based localization, and layout-preserving view construction. The method in this paper overlaps a lot with [2], which makes the contribution incremental.
  2. In Appendix C, the baseline method uses an option letter to conduct scoring, while PostCrop generates a phrase and uses Yes/No to do the verification. Please make sure the generation and verification are identical.
  3. The ablation results show PostCrop improves limited performance. In Table 4, the adaptive method improves from 68 to 71, compared to Union, on 4K while keeping the same on 8K. In Table 5, different variants share the same performance on 4K and only 1 point on 8K. Does this mean the proposed method shows very little improvement?
  4. Table 1 reports Qwen2.5-VL FCP results of 74%/76%, whereas Tables 4–6 report 71%/78% for the complete PostCrop configuration. The ablations are described as using the complete FCP subsets. Please ensure the performance is consistent across different tables.
  5. Since the method in this paper requires an additional composition step compared to existing work, the latency of running such a pipeline is high. It would be better to have a cost analysis of the trade-off between them.

[1] Khayatkhoei, M., Chhikara, P., & Ilievski, F. (2025, May). Mllms know where to look: Training-free perception of small visual details with multimodal llms. In International Conference on Learning Representations (Vol. 2025, pp. 68194-68213).
[2] Liu, X., Hu, Y., Zou, Y., Wu, L., Xu, J., & Zheng, B. (2025). Hide: Rethinking the zoom-in method in high resolution mllms via hierarchical decoupling. arXiv preprint arXiv:2510.00054.

Questions for authors

Please see the weakness, especially about the novelty, limited improvement in the ablation study.

4 comments

We thank the reviewer for the thoughtful and constructive feedback. We would like to address each concern below. We have uploaded the rebuttal revision, revised the main text, and added further details to the appendix to address your concerns.

W1: What is the contribution of PostCrop beyond existing work on entity localization?

  • First, we acknowledge that the entity localization stage of PostCrop draws on existing studies, including semantic decomposition and attention. However, this has already been clearly explained in the original paper through the related work citations and the method description. We also emphasize in the paper that this part is not the focus of our research (see Section 3.2), and it is not included among our contributions (see the Introduction).

  • Second, our actual core contribution is the study of visual context. After entity localization, in addition to preserving the relative positions of entities, we investigate whether there is other useful visual information for perception in the context and, if so, how to capture this information.

  • Third, as we also summarize in the Conclusion, fine-grained perception cannot rely on localization alone ([1], [2]). It also needs to consider the use and composition of visual information. This expresses that our core work is visual recomposition and points to a possible direction for future development.


W2: How is answer readout handled to avoid bias from answer options?

To avoid bias from answer options, we improved the readout method compared with scoring based on option letters. To further ensure a fair comparison, we present the phrases generated by PostCrop and compare them using Yes/No verification, as shown in Table 1 of the manuscript.


W3: How much improvement does adaptive cropping bring over Union, and how does it perform on another benchmark?

  • First, we acknowledge that the performance improvements in some ablation experiments are limited. In particular, on HR-Bench-8K, B* and Union achieve the same accuracy. This is consistent with the diagnostic results in Section 2.1: on the current 8K test set, retaining the full Union already provides sufficient relational information. Further compression may instead cause evidence loss. Therefore, there is limited room for adaptive cropping to improve over Union.

  • Second, to further verify the effectiveness and generalization of the method, we added ablation experiments on TreeBench (Wang et al., ICLR 2026). Using the same localization and answering process, we compare the performance of LPD, Union, and B*, as shown below. These results show that the benefit of adaptive context selection is closely related to the visual information required by the task, and can bring more noticeable improvements in some complex, high-resolution scenes.

TreeBench
LPD34.57%
Union34.32%
PostCrop (B)*35.31%

W4: Which Qwen2.5-VL FCP results are correct?

After checking and rerunning the experiments, we will correct the Qwen2.5-VL FCP results reported in Table 1 from 74%/76% to 71%/78%, respectively, and make the corresponding corrections elsewhere in the manuscript.


W5: What is the stage-wise runtime of the PostCrop modules?

To analyze the overhead of the PostCrop modules, we provide additional runtimes for each stage by question type on VSTAR using Qwen2.5-VL-7B. The average total time per question for Attribute and Spatial questions is 4.4782 seconds and 16.2932 seconds, respectively. For Spatial questions, the time is mainly spent on Cue Attention and Entity Localization, Identity Verification, and Relation Evidence Composition. The results for each step are provided in Appendix Tables 14–15.

Dear Reviewer ujQn,

Thank you for your valuable feedback. We have carefully addressed your concerns in our rebuttal.

As the discussion period is coming to an end, we would greatly appreciate it if you could review our responses and let us know if you have any remaining concerns.

Thank you for your time and consideration.

Best regards,

The authors of Paper Submission40345

Dear Reviewer ujQn,

Thank you for your valuable feedback. We have carefully addressed your concerns in our rebuttal.

As the discussion period is coming to an end, we would greatly appreciate it if you could review our responses and let us know if you have any remaining concerns.

Thank you for your time and consideration.

Best regards,

The authors of Paper Submission40345

Dear Reviewer ujQn,

Thank you for your valuable feedback. We have carefully responded to your concerns in our rebuttal.

As the discussion period is coming to an end, we would greatly appreciate it if you could review our responses and let us know if you have any remaining concerns.

Thank you for your time and consideration.

Best regards,

The authors of Paper Submission40345

Reviewer_LTS3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PostCrop tackles multi-entity relation questions on high-resolution images, where training-free zoom-in methods either crop too tightly and lose the context between entities or keep the full bounding box and admit distractors. Instead of cropping each entity first, it composes a joint evidence map for the entities on the original token grid via bidirectional attention transfer, then picks the crop between “entities only” and “full box” that maximizes evidence-weighted coverage. It beats HiDe substantially on relation subsets (HR-Bench FCP 61→74 on Qwen2.5-VL-7B) but not on single-entity questions, at 2.2× the cost.

Critical strengths

1: Clean diagnostic design. Entity boxes are held fixed and only the amount of retained context is varied, separately at 4K and 8K. The conclusions (context helps, the right amount varies by setting, and selecting context from single-entity response maps is no better than random) are causal, not benchmark deltas. The Fig. 3 result that Source, Partner and Sum all equal Random is especially valuable because it rules out an intuitive approach.

2: It identifies a structural blind spot, not a tuning gain. “Crop first and you can never see what lies between the entities” is a falsifiable claim rather than “our trick works better”. It moves the contribution from engineering to analysis.

3: Large, targeted, model-consistent gains. HR-Bench FCP rises from 55–61 to 74–76, and the same pattern holds across three backbones and two Qwen2.5-VL sizes: relation questions improve a lot, single-entity questions stay flat or dip slightly. Gains that appear only where the method is supposed to act are more convincing than uniform small improvements.

Critical weaknesses

1: Cost and practicality. PostCrop needs many forward passes per question, giving 39 min on VSTAR vs. 17.7 min for HiDe (Table 7); the ViCrop timing is approximate with no stated source. Please report forward passes and token counts per question, and compare against a compute-matched HiDe variant (e.g., HiDe + option verification).

2: Inconsistencies across tables for the same configuration Table 1 reports PostCrop (Qwen2.5-VL-7B) HR-Bench FCP = 74.0 (4K) and 76.0 (8K). Table 4, which is the full method ("PostCrop (B*)", same backbone, "complete FCP subset"), reports 71.0 (4K) and 78.0 (8K). Tables 5 and 6 also report 71.0 / 78.0 for the full configuration. FSP matches (93.0) in both tables, so this is not a different run of the whole benchmark. Which number is the method?

Questions for authors

1: What are the sources and backbones of the ZoomRefine / CVSearch / DeepScan / DyFo / DeepEyes numbers in Fig. 1(c)? Can ZoomEye or DyFo be run on Qwen2.5-VL-7B with the same readout?

2: Given Section 2.1 and Table 10 show that 1.1× expansion beyond Union helps in some settings, how do you reconcile this with Proposition 3?

3: How sensitive is accuracy to the choice of middle layer (e.g., layers 11–19 on Qwen2.5-VL-7B)? Why is the InternVL tile budget 36 for PostCrop but 128 for HiDe?

4 comments

We thank the reviewer for the thoughtful and constructive feedback. We would like to address each concern below.

Q1: What are the sources and backbones of the methods in Fig. 1(c), and can ZoomEye and DyFo be included using the same answer readout?

The results for ZoomRefine, CVSearch, DeepScan, DyFo, and DeepEyes in Fig. 1(c) are taken from the corresponding papers, with Qwen2.5-VL-7B as the backbone. We also include ZoomEye and DyFo for comparison using the same answer readout, as shown below.

MethodTraining freeVSTAR Attr.VSTAR SpatialVSTAR Overall
Qwen2.5-VL-7B—83.580.382.2
ViCrop✓86.176.382.2
HiDe✓92.289.591.1
ZoomEye✓84.384.284.3
DyFo✓88.785.587.4
PostCrop (ours)✓93.994.794.2

W1: What is the inference cost of PostCrop compared with ViCrop and HiDe?

The table below shows the inference cost comparison from our reruns, including the original ViCrop, the original HiDe, a HiDe variant, and PostCrop.

MethodPeak memory (GiB)Time (min)
ViCrop20.18.0
HiDe19.316.1
HiDe variant (adapted for PostCrop)19.327.2
PostCrop19.129.2

Q2: How can the results beyond Union be reconciled with Proposition 3?

Regarding the control experiments beyond Union, where the score may increase while accuracy does not, we will further explain the role of Proposition 3 and introduce another benchmark as additional evidence for our conclusion.

  • First, we need to clarify that Section 2.1 and Table 10 examine different things. Section 2.1 compares the final crop regions with different extents, in order to diagnose how context affects perception. Table 10, on the other hand, compares the final performance when adaptively selecting B* after expanding the upper bound of the search range.

  • Second, we emphasize that Proposition 3 states that Union is a sufficient search boundary under the conditions of complete cue coverage, accurate localization, and relation locality. When these conditions are not satisfied, using a larger upper bound may lead to gains that are not sufficiently reliable. We conducted a significance analysis of the results in Table 10, where the accuracy with a 1.1× search upper bound is 74.29%, compared with 72.86% for Union. In the paired comparison for each question, two cases change from incorrect to correct, while one changes from correct to incorrect. The statistical test gives p = 1.000. Therefore, this increase in accuracy is not statistically significant.

  • Third, we use TreeBench (Wang et al., ICLR 2026) as additional evidence. TreeBench provides ground truth that can be used to determine whether localization is accurate. We therefore selected 86 correctly localized questions that fit the PostCrop setting as a validation set, in order to compare the effects of different upper bounds when the conditions are satisfied. Union and 1.3× Union both achieve 20.93%, while 1.1× Union achieves 18.60% and 1.5× Union achieves 17.44%. The results do not show that expanding the search range leads to a monotonic increase in accuracy. The detailed experimental settings and results are provided in Appendix Table 11, immediately following Table 10.

Q3: How sensitive is accuracy to the choice of middle layer, and why are different tile budgets used for PostCrop and HiDe?

The table below compares layers 11–19 of Qwen2.5-VL-7B on VSTAR. Accuracy is relatively insensitive to the choice of middle layer.

MethodLayer 11Layer 12Layer 13Layer 14Layer 15Layer 16Layer 17Layer 18Layer 19
PostCrop (Qwen2.5-VL-7B)89.01%93.72%92.67%89.53%94.24%93.19%91.10%90.58%90.58%

The tile budget for InternVL is 36 in PostCrop and 128 in HiDe, following HiDe's original setting of 128. To make the comparison consistent, we set the tile budget to 36 for both methods. The results are shown below.

MethodTraining freeVSTAR Attr.VSTAR SpatialVSTAR OverallHR-Bench-4K FSPHR-Bench-4K FCPHR-Bench-4K OverallHR-Bench-8K FSPHR-Bench-8K FCPHR-Bench-8K Overall
InternVL2.5-4B (36 tiles)—76.575.075.979.055.067.065.053.059.0
HiDe (ICML'26)✓87.877.683.884.053.068.583.051.067.0
PostCrop (ours)✓86.180.383.881.060.070.584.057.070.5

W2: Which Qwen2.5-VL FCP results are correct?

After checking and rerunning the experiments, we will correct the Qwen2.5-VL FCP results reported in Table 1 from 74%/76% to 71%/78%, respectively, and make the corresponding corrections elsewhere in the manuscript.

I thank the authors for the response. Some concerns are resolved, but several remain:

W1: The new times (29.2 vs. 16.1 min) also differ from Table 7 (39 vs. 17.7 min). Please explain the difference.

W2: Thank you for the correction (74/76 → 71/78). Please also update the abstract. With 71.0 on 4K, the gain of B* over Union is 3 points on 4K and zero on 8K. Is that correct?

Q3: At a matched 36 tiles, PostCrop ties HiDe on VSTAR (83.8 vs. 83.8), and the HR-Bench margins shrink to 2–3.5 points.

The added ZoomEye/DyFo comparison is helpful. I can keep my score for now. I also share Reviewer FsM1’s concerns about baseline provenance and benchmark-specific routing.

We thank the reviewer for the thoughtful and constructive feedback.

W1: Thank you for pointing out this inconsistency. The times in Table 7 were from an earlier run and were not updated with the latest measurements. We have corrected Table 7 and the Efficiency Analysis accordingly. To minimize external interference, we conducted an additional runtime comparison on the full 191-question VSTAR evaluation, excluding model loading and warm-up. The measured times are 29.2 minutes for PostCrop and 16.1 minutes for HiDe, respectively. We apologize for this oversight.

MethodPeak memory (GiB)Time (min)
ViCrop20.1±0.18.0±0.1
HiDe19.3±0.116.1±0.1
HiDe variant (adapted for PostCrop)19.3±0.227.2±0.2
PostCrop19.1±0.229.2±0.1

W2: Thank you for pointing this out. We have updated the abstract accordingly. Yes, B* shows no improvement over Union on HR-Bench at 8K. However, this does not necessarily imply that context selection is ineffective. Through a case-by-case analysis of the incorrect predictions from both methods, we found that HiDe has more cases where the relevant region is correctly localized but the final answer remains incorrect.


Q3: Yes. Although PostCrop and HiDe achieve the same overall VSTAR score with 36 tiles, PostCrop performs better on the Spatial subset, suggesting that context selection remains beneficial.

Reviewer_FsM1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

PostCrop is a training-free framework for high-resolution LVLM fine-grained perception. It extracts object and attribute cue maps, verifies entity regions, composes multi-entity evidence on the original visual grid using bidirectional attention transfer, and searches expansion/compression candidates to select a context view B*. The model answers from the original image together with the selected view. Experiments on VSTAR, HR-Bench-4K/8K, and MME-RealWorld-Lite with Qwen2.5-VL, Qwen3-VL, and InternVL report the largest gains on cross-instance relation questions, supported by diagnostics, ablations, context-source controls, and a timing analysis.

Critical strengths
  1. The paper identifies a useful and practically relevant context trade-off: entity-only views can lose relational evidence, while broad crops can introduce interference. The Compose-Before-Crop formulation and the bidirectional evidence construction are a coherent response to this problem.

  2. The revised evaluation gives a clearer basis for the comparison. Appendix C.1 specifies a common answer prompt, option order, and scoring procedure for the open-source methods in Table 1, and the supplementary per-question records reproduce the reported overall scores. The clearest gain is on HR-Bench-8K FCP with Qwen2.5-VL-7B, where PostCrop reaches 78% compared with 63% for HiDe. The revised Table 2 also reports results on the 1,819 MME-RealWorld-Lite questions outside calibration. The Reasoning advantage over HiDe persists (50.15% vs. 45.69%), while overall accuracy is slightly lower (53.71% vs. 53.88%). These results support my revised recommendation of weak acceptance for the focused contribution to fine-grained relation perception.

Critical weaknesses
  1. The gains from adaptive crop selection itself are modest. With the other components held fixed, Table 4 shows a three-point FCP gain over a fixed Union crop on HR-Bench-4K and a tie at 78% on 8K. The larger Table 1 gains over HiDe therefore describe the full pipeline. The TreeBench comparison is also conditional on a selected set of 86 questions with role-consistent localization overlap, which limits what it establishes about search boundaries in general.

  2. Cue extraction and inference cost limit the scope of the result. The rebuttal clarifies that fixed HR-Bench question IDs were used to reuse earlier cue records and that the full pipeline can extract cues from the question and choices. The reported cue-extraction failures still affect performance on unfamiliar examples. On VSTAR, revised Table 7 reports 29.2 minutes for PostCrop versus 16.1 minutes for HiDe. The additional ZwZ comparison draws on results from different evaluation setups, so it offers limited evidence about how the methods compare in practice.

Questions for authors
  1. For the adaptive crop versus fixed Union comparison in Table 4, could you report how many questions improved and how many regressed under adaptive cropping at each resolution? This would show whether the 8K tie reflects similar answers or a trade-off across examples.

  2. For the ZwZ fine-tuning experiment discussed in the rebuttal, which split was used for the final evaluation? Please clarify whether it excluded the HR-Bench-8K examples and source images used to select training crops.

4 comments

We thank the reviewer for the thoughtful and constructive feedback. We would like to address each concern below. We have uploaded the rebuttal revision, revised the main text, and added further details to the appendix to address your concerns.

Q1/W1: Which results were re-evaluated under the same prompts and rules for matching answers?

We re-evaluated Base, ViCrop, HiDe, and PostCrop using the same prompts, option order, and rules for matching answers. See Table 1 in the manuscript.


Q2: Are cue extraction and question-category routing automated at inference time?

Cue extraction and question-category routing are automated at inference time. The supplementary code includes a fixed mapping from HR-Bench question IDs to categories, as well as a deterministic parser for VSTAR. Cue extraction is implemented through prompts. The main function of the prompt is to extract visual cues from the questions and answer options in benchmark VQA tasks. Appendix G ("Prompt Templates") provides the unified prompt for cue extraction and its automatic branches.


W2: Are the HR-Bench mapping and VSTAR parser part of the cue-extraction prompt?

The mapping from HR-Bench question IDs to categories and the VSTAR parser are used for downstream rule-based routing. They are not part of the prompt for cue extraction itself. Meanwhile, Appendix G ("Observed Cue Extraction Failure Cases") also explains that errors in cue extraction or routing may still affect the results. We consider these to be failure cases.


Q3: How were the hyperparameters selected, and how should the results beyond Union be interpreted?

We used 50 questions from the Other2Other subset and 50 questions from the Ego2Pedestrian subset in the official Relation_Interaction category of MME-RealWorld-Lite as calibration subsets for hyperparameter analysis. We fixed the intermediate layer at L15, following HiDe. We did not search over layers in this round. We conducted 28 comparisons, changing one hyperparameter at a time. The results from both subsets support the current parameter choices for PostCrop with Qwen2.5-VL-7B. The parameter settings are provided in Appendix H ("MME-RealWorld-Lite Hyperparameter Screening") and Table 16.

Regarding the issue in the control experiments beyond Union, where the score may increase while accuracy does not, we will further explain the role of Proposition 3 and introduce another benchmark as additional evidence for our conclusion.

  • First, we emphasize that Proposition 3 states that Union is a sufficient search boundary under the conditions of complete cue coverage, accurate localization, and relation locality. When these conditions are not satisfied, using a larger upper bound may lead to gains that are not sufficiently reliable. We conducted a significance analysis of the results in Table 10, where the accuracy with a 1.1× search upper bound is 74.29%, compared with 72.86% for Union. In the paired comparison for each question, two cases change from incorrect to correct, while one changes from correct to incorrect. The statistical test gives p = 1.000. Therefore, this increase in accuracy is not statistically significant.

  • Second, we use TreeBench (Wang et al., ICLR 2026) as additional evidence. TreeBench provides ground truth that can be used to determine whether localization is accurate. We therefore selected 86 correctly localized questions that fit the PostCrop setting as a validation set, in order to compare the effects of different upper bounds when the conditions are satisfied. Union and 1.3× Union both achieve 20.93%, while 1.1× Union achieves 18.60% and 1.5× Union achieves 17.44%. The results do not show that expanding the search range leads to a monotonic increase in accuracy. The detailed experimental settings and results are provided in Appendix Table 11, immediately following Table 10.


Q4: Does the Structure Adapter affect the reported results, and what is the improvement on Reasoning tasks?

The Structure Adapter is legacy code retained from an early stage of method development. It is not part of the final PostCrop method, and the reported results on HR-Bench do not depend on this branch. We will provide cleaner source code.

We acknowledge that the overall improvement of PostCrop on MME-RealWorld-Lite is small (+0.21 percentage points). However, the results by task show that PostCrop achieves a significant improvement on Reasoning tasks (+4.80 percentage points, p = 0.0074).

We sincerely apologize for our delayed response. We initially overlooked that your comments were updates to the original questions. Thank you for your careful review and constructive feedback. We address your four concerns below.

Q1: Which Table 1 results were rerun, and how was a consistent evaluation protocol ensured?

Thank you for pointing out the inconsistency. We confirm that all results in the current Table 1 were evaluated under a unified protocol, with identical prompts, answer-option ordering, and answer-matching rules.

We sincerely apologize for the conflicting description in Appendix C.1. The statements attributing some results to the earlier HiDe and ActiveScope tables were carried over from an initial manuscript version, and we inadvertently failed to update Appendix C.1 during the previous rebuttal revision.

We have now corrected Appendix C.1 to accurately describe the evaluation underlying Table 1. In addition, we have included per-question evaluation results in the supplementary materials, allowing the reviewer to verify that the compared methods followed a consistent evaluation protocol and to inspect the reported comparisons.


Q2: Do the 100 MME-RealWorld-Lite calibration questions overlap with the 1,919-question evaluation, and what happens when they are excluded?

Thank you for raising this concern. We confirm that the 100 hyperparameter-calibration questions (50 Other2Other and 50 Ego2Pedestrian) are a subset of the 1,919 MME-RealWorld-Lite evaluation questions.

Our original plan was to use two randomly sampled 50-question subsets for cross-validation of hyperparameter selection, with the aim of reducing sensitivity to a particular calibration sample. We recognize, however, that overlap between calibration and evaluation warrants a separate check.

To address this directly, we additionally evaluated performance after excluding Other2Other and Ego2Pedestrian, leaving 1,819 questions. The results are shown below.

ModelPer.Rea.Overall
qwen-2.5vl51.4142.6248.27
HiDe58.4345.6953.88
PostCrop55.6950.1553.71

These results show that PostCrop's gains over the Base model persist after excluding the calibration questions: its overall accuracy is 53.71%, compared with 48.27% for Base (+5.44 percentage points). Compared with HiDe, PostCrop maintains a +4.46 percentage-point advantage on Reasoning (50.15% vs. 45.69%), while its overall accuracy is 0.17 percentage points lower (53.71% vs. 53.88%). Thus, we do not claim an overall advantage over HiDe on this exclusion subset.


Q3: Why was fixed question-ID routing used, and how does PostCrop handle cue extraction on unseen question types or datasets?

Thank you for the opportunity to clarify the end-to-end procedure. The fixed question-ID routing used in HR-Bench served to reuse visual-cue records extracted during an initial prompt-based pass, facilitating subsequent experiments on the core PostCrop modules. Fixed question IDs are not a required component of the complete inference pipeline.

In the complete end-to-end method, the evaluated backbone model extracts visual cues from the raw question text and answer options through prompting. Our communication-overhead evaluation also uses this complete procedure, including the cue-extraction step.

For unseen question families or new datasets, we use the same general prompt-based framework to extract cues, without requiring a predefined question-ID mapping. We acknowledge that routing or cue extraction may fail when the backbone model cannot adequately interpret the question semantics or relevant visual relations. This dependence on the backbone model's cue-understanding capability is a limitation of PostCrop.

The observed failure rates are:

  • VSTAR: 1.57%
  • HR-Bench: 6.50%
  • TreeBench: 7.40%

Finally, the reported evaluation uses the raw question text and answer options to obtain visual cues through prompting; no additional benchmark-specific metadata is provided as input to cue extraction.

Q4: How was the TreeBench subset selected, what does it establish about search boundaries, and how was the +4.80 pp MME Reasoning improvement determined?

Thank you for these questions. First, we correct a numerical error: the selected TreeBench subset contains 86 questions, not 90.

We applied a two-stage selection procedure. First, we selected multi-object TreeBench questions with two relevant ground-truth (GT) bounding boxes, providing the multi-object relational setting relevant to PostCrop.

Second, we retained a question only when the two final boxes generated by PostCrop's first module satisfy a one-to-one, role-consistent positive-overlap criterion: the box localized from the source cue must overlap the source-role GT box, and the box localized from the target cue must overlap the target-role GT box, with IoU > 0 for both pairs (i.e., a strictly positive intersection area).

This cue-to-GT role-consistency requirement excludes role-swapped or mismatched localizations. For example, q189 was excluded because the box localized from the source cue corresponded to the other target's GT box.

The second criterion is important because PostCrop follows the Compose-Before-Crop principle. Directly using GT boxes from the first-stage selection would bypass model-generated localization and evidence composition, and would therefore not faithfully evaluate this principle. Section 2.2 provides supporting analysis for the Compose-Before-Crop design.

The selected 86 questions satisfy the first two conditions of Proposition 3: Complete Cue Coverage and Complete Localization. These questions involve multiple objects, and the two boxes produced by the first module satisfy the one-to-one, role-consistent positive-overlap criterion with their corresponding GT boxes. The third condition, Relation Locality, follows a commonsense assumption: for the queried relation, the relevant entities and the visual context within their Union provide sufficient information to determine the relation. We therefore use these 86 questions to examine the effectiveness of the Union search boundary under the three conditions of Proposition 3.

Regarding the reported +4.80 percentage-point improvement on MME Reasoning, this refers to the full 1,919-question evaluation in Table 2, where PostCrop achieves 47.47% and the HiDe baseline achieves 42.67% on Reasoning (47.47 − 42.67 = +4.80 pp).

On the 1,819-question evaluation excluding calibration items described in Q2, the corresponding difference is 50.15 − 45.69 = +4.46 pp. These figures correspond to different evaluation subsets and should not be conflated.

Dear Reviewer FsM1,

Thank you for your valuable feedback. We have carefully addressed your concerns in our rebuttal.

As the discussion period is coming to an end, we would greatly appreciate it if you could review our responses and let us know if you have any remaining concerns.

Thank you for your time and consideration.

Best regards,

The authors of Paper Submission40345

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.