acceptodds
Under review as a conference paper at ICLR 2027

PerceptVLM: From Seeing Details to Connecting Evidence for Multimodal Perception

Abstract

Multimodal perception forms the foundation of multimodal intelligence. In particular, spatial perception and multimodal long-context understanding are two pivotal tasks that underpin embodied AI and multimodal agents. However, synthesizing targeted, high-quality training data for these tasks remains difficult and costly. Our key insight is that these tasks share an underlying capability, which we call connecting evidence: identifying relevant visual facts, binding them to the correct entities, recovering their relations, and composing them into an answer. Crucially, targeted training data are far easier to synthesize for this capability than for these tasks themselves. To strengthen this capability, we propose EvidenceConnect, an easy-to-implement data synthesis framework that uses zoomed views of related regions within a single image to generate questions requiring comparison, aggregation, or chaining of visual evidence. To ensure quality, a reference answer is retained only when independent VLMs agree on it, and region-ablation tests discard questions whose answer survives the removal of selected regions. We then train base VLMs on the resulting data with RLVR, obtaining PerceptVLM. Although the synthesized data are not tailored to any downstream benchmark, PerceptVLM improves fine-grained and spatial perception, with further gains on embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning (shown in Figure 1). These results establish EvidenceConnect as an effective framework for supervising evidence connection and suggest that strengthening this shared capability leads to more generalizable multimodal perception.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Large Language Models, Multimodal Perception, Synthetic Data

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_i7LsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EvidenceConnect, which uses cropped region views to generate visual question-answering data requiring evidence composition across multiple regions, and filters samples through cross-model answer verification and region-removal tests. The authors use approximately 20.4K examples to train three sizes of Qwen models with GRPO, producing PerceptVLM. Experiments cover 26 benchmarks across six task categories, with improvements over the corresponding base models at all three scales. Despite using only single-image training data, the models also improve on multi-image long-context understanding and agent tasks.

Critical strengths
  1. The data construction pipeline is fairly complete, and the authors carefully consider answer reliability, dependence on multi-region evidence, and referential ambiguity. Cross-model answer verification and region-removal tests can improve data quality, making the method practically useful.
  2. The evaluation is broad, with improvements at all three model scales. Training only on single-image data also improves multi-image long-context understanding and agent tasks, showing good cross-task transfer.
Critical weaknesses
  1. Although the authors explain that colored outlines reduce referential ambiguity, they do not report the proportions of marked and unmarked training examples or directly evaluate the effect of region markers on downstream performance. Moreover, comparisons with other single-image datasets provide indirect evidence for multi-region supervision, but differences in data quality and question distributions remain confounding factors; controlled ablations within the proposed data construction pipeline would better distinguish the contributions of region markers and multi-region evidence composition.
  2. The paper primarily compares against three general-purpose models, without representative strong multimodal or domain-specialized models, and external baseline results are missing on some benchmarks, leaving its claims of frontier-level competitiveness insufficiently supported.
Questions for authors
  1. What proportions of the training samples have region markers and have no markers? Downstream evaluations usually do not provide these markers. Could this difference affect generalization? Have the authors compared performance after training with and without region markers, keeping other training conditions the same?
  2. Comparisons with other datasets also involve differences in data quality and question types, making it difficult to determine how much of the improvement comes from multi-region evidence composition. Could the authors compare single-region and multi-region supervision within their own data construction pipeline, keeping the source images, verification procedures, data size, and training budget as consistent as possible?
  3. What was the basis for selecting these three external reference models? Could the authors provide results for stronger general-purpose multimodal models or relevant domain-specialized models under the same or comparable evaluation conditions to support the paper’s claims of frontier-level competitiveness?
8 comments
AuthorsOpenReview ↗

Response to Reviewer i7Ls (1)

Q1. Although the authors explain that colored outlines reduce referential ambiguity, they do not report the proportions of marked and unmarked training examples or directly evaluate the effect of region markers on downstream performance. Moreover, comparisons with other single-image datasets provide indirect evidence for multi-region supervision, but differences in data quality and question distributions remain confounding factors; controlled ablations within the proposed data construction pipeline would better distinguish the contributions of region markers and multi-region evidence composition.

A1. Thank you for the suggestions. We provide the requested statistics and two controlled ablation studies. Both studies use the same Qwen3.5-4B initialization and GRPO configuration, with 7K training examples for efficiency and two epochs of training.

(1) Region markers.

Among the 20,662 examples used to train the 4B model, 20,624 (99.82%) contain at least one region marker, while 38 (0.18%) use unmarked images. Counting each referenced region once per question, 71,485 of 71,615 evidence regions (99.82%) are marked. As explained in Section 3.3, these markers disambiguate the intended instances when several objects in the full image could match the same description. The model must still perceive and combine the visual evidence within the referenced regions. We have included these in the revised version, as reported in Appendix A.1.

To directly assess the effect of markers, we constructed paired marked and unmarked versions of the same 7K examples for training, preserving the source images, target evidence regions, and underlying tasks. For the unmarked version, we removed the outlines and replaced color-based references in the questions and answer options, where applicable, with natural-language descriptions, followed by answer verification. The unmarked variant achieves a mean score of 52.93, compared with 49.41 for the base model (+3.52 points). The marked variant achieves 54.08, improving over the base model by 4.67 points and over the unmarked variant by 1.15 points. These results show that the training data remain beneficial without markers, while retaining markers during training yields higher downstream performance even though no such markers are added at evaluation time. We attribute this gap to clearer visual references. Colored outlines directly identify the intended objects, whereas descriptions may be ambiguous among similar objects or reveal the queried attribute. Under RLVR, this ambiguity introduces reward noise by penalizing the model for selecting a plausible but unintended region. Markers reduce this confound, focusing training on locating relevant regions and combining their evidence. The gains persist without markers at evaluation, suggesting improved composition ability rather than reliance on outlines. This supports the transferability of marker-based supervision beyond the marked training inputs. We have updated our submission in Section 4.4, with details provided in Appendix A.1 (Table 4).

(2) Multi-region evidence composition.

We constructed single-region and multi-region supervision within our own pipeline, using the same source images, question-generation model, and answer-verification procedure. Both variants use a consistent region-reference format and the matched training settings described above.

In the comparison, single-region supervision achieves a mean score of 52.23, while multi-region supervision achieves 54.08, a gain of 1.86 points, with improvements on 9 of the 12 benchmarks. Because single-region and multi-region supervision necessarily involve different questions, this comparison evaluates the supervision procedure as a whole. With source images, answer verification, data size, and training settings controlled, the results provide direct evidence that requiring the model to compose evidence across regions improves downstream performance over single-region supervision. These have now been incorporated into Section 4.4, as summarized therein and detailed in Appendix A.1 (Table 4).

BenchmarkQwen3.5-4B baseWithout markers (7K)Single-region (7K)Ours (7K)
HR-4K82.3786.5084.1386.62
HR-8K78.3780.5080.0082.87
CountQA36.9139.2038.9445.09
O3-Bench53.9158.5558.8462.90
ZoomBench50.4154.7959.7655.98
MMStar74.1374.4773.3376.60
TreeBench50.3748.6450.8649.88
Chartography5.0013.008.009.00
MMSI-Bench30.7030.8031.1034.60
MindCube-Tiny47.6253.6250.9552.29
ERQA45.7549.7547.7547.50
ViewSpatial-Bench37.4345.3843.0745.68
Avg49.4152.9352.2354.08
AuthorsOpenReview ↗

Response to Reviewer i7Ls (2-1)

Q2. The paper primarily compares against three general-purpose models, without representative strong multimodal or domain-specialized models, and external baseline results are missing on some benchmarks, leaving its claims of frontier-level competitiveness insufficiently supported.

A2. Thank you for the advice. Evaluating external models across our suite of 25 benchmarks spanning six task categories incurs substantial API and inference costs, which constrained both the number of models and the completeness of their evaluation in the original submission. We therefore initially prioritized three general-purpose multimodal models to provide references across these diverse tasks. To address your concern, we have now expanded the model comparisons and completed the previously missing results for the original reference models. We have also revised the competitiveness claims to specify the models, tasks, and evaluation settings supported by these results.

(1) Broader external model comparisons.

We initially selected general-purpose multimodal models because our evaluation spans six task categories. DeepSeek-V4.1-Flash and GLM-5.3-Flash provide references from two open-model families, while GPT-5.6-Luna provides a cost-effective proprietary reference that enabled broad evaluation coverage. This selection prioritized coverage and diversity of model families. To broaden the comparison, we have added MiMo-V2.6-Pro, Step5, and Claude-Haiku-5.5 (closed-source). In the expanded comparison, PerceptVLM remains competitive, achieving the highest scores among the reported configurations on 7 of the 15 benchmarks. PerceptVLM-27B achieves a mean score of 54.72, exceeding MiMo-V2.6-Pro (high: 47.30) and Step5 (low: 43.21; high: 41.96). Notably, PerceptVLM-27B exceeds the strongest external result in this table by 5.23 points on VStar, 4.47 points on MindCube-Tiny, and 4.13 points on ViewSpatial-Bench. These results have now been incorporated into the revised manuscript, summarized in Section 4.2 and detailed in Appendix A.5 (Table 8).

BenchmarkMiMo-V2.6-Pro highStep5 lowStep5 highHaiku-5.5 lowHaiku-5.5 highPerceptVLM-4BPerceptVLM-9BPerceptVLM-27B
Chartography5.009.0010.0035.0037.0014.0013.0012.00
VStar80.1082.2082.7286.3985.8689.0090.5891.62
BabyVision26.5529.6427.3234.5442.0118.5621.3928.35
O3-Bench72.4669.5771.3075.0781.4563.7767.8372.46
TreeBench57.7855.8060.2560.7462.4751.3651.6056.05
MMStar77.0778.4078.3375.9376.9376.2778.2779.07
ERQA51.7547.7544.5062.2559.2552.0055.5061.75
MMSI-Bench25.9016.8012.3041.6044.1032.3033.9037.10
MindCube-Tiny25.2424.5719.0550.2939.4353.0552.5754.76
ZoomBench52.5451.2449.8258.4659.5358.2259.5361.78
CountQA53.4036.1934.8867.3469.7047.1948.4354.32
Perception-Bench41.7040.8040.3354.8357.0337.8038.8944.93
SPAR-Bench47.4714.9611.6948.6047.2045.6947.2350.78
ViewSpatial-Bench38.5334.1730.6546.5046.3845.2645.2750.63
MIMIC53.9857.1356.3355.9356.5066.1064.9865.16
Avg47.3043.2141.9656.9057.6650.0451.2654.72
AuthorsOpenReview ↗

Response to Reviewer i7Ls (2-2)

(2) Comparison with a domain-specialized training approach.

We would like to clarify that our goal is to identify and strengthen a shared underlying capability, connecting visual evidence, that supports multiple task domains. Domain-specific data synthesis can be effective, but developing separate synthesis and verification procedures for each domain can require substantial additional effort. EvidenceConnect targets this shared capability through a unified, single-image data synthesis pipeline, aiming to improve performance across domains without separately curating training data for each one.

To assess this approach against domain-specialized supervision, Section 4.3 (Table 2) has already compared EvidenceConnect with ZwZ [1], which targets fine-grained visual perception through supervision grounded in individual regions. Using the same Qwen3.5-4B backbone and GRPO recipe, training on 20.7K EvidenceConnect examples achieves a mean score of 53.86 across 11 benchmarks, compared with 49.86 when training on 73K ZwZ examples, a 4.00-point improvement with fewer training examples. These results support our premise that explicitly training models to connect evidence across regions provides an effective route to broad performance gains, even when compared with supervision specifically designed for fine-grained perception.

(3) Completion of the original baseline results and scope of conclusions.

We have filled in the previously missing entries for the three original reference models in Section 4.2 (Table 1). These completed results, together with the expanded comparisons, provide a broader basis for assessing absolute performance. Separately, comparisons with the corresponding 4B, 9B, and 27B Qwen baselines establish the gains attributable to EvidenceConnect within each backbone.

[1] Wei, Lai, et al. "Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception." ICML 2026.

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

Spatial perception and multimodal long-context understanding are two critical capabilities for multimodal large language models (MLLMs). To address the challenge of curating training data for boosting performance on these two fronts, this work introduces EVIDENCECONNECT: an automated synthesis framework that produces visual questions requiring the model to connect visual evidence across distinct image regions. Retraining Qwen on the synthesized data yields substantial improvements on both tasks, with results even outperforming certain strong proprietary models such as GPT.

Critical strengths

A reasonable framework which is verified through experiments shows performance gains over the baseline models.

Critical weaknesses

My primary concern is the limited technical novelty of this work. The pipeline mainly consists of region selection, question generation, answer verification, and dataset curation. The core focus lies in data construction and empirical evaluation. Although the authors demonstrate effectiveness via extensive experiments and quantitative comparisons, the technical contributions remain somewhat incremental.

Questions for authors

The authors only expriment with the Qwen series models.

I would like to see the retraining results on other advanced open-source models.

2 comments
AuthorsOpenReview ↗

Response to Reviewer EUTm (1)

Q1. My primary concern is the limited technical novelty of this work. The pipeline mainly consists of region selection, question generation, answer verification, and dataset curation. The core focus lies in data construction and empirical evaluation. Although the authors demonstrate effectiveness via extensive experiments and quantitative comparisons, the technical contributions remain somewhat incremental.

A1. Thank you for acknowledging the effectiveness of our approach. We would like to clarify the central insight and methodological contribution, which are not fully captured by describing the framework through the generic stages of region selection, question generation, verification, and curation.

Core insight: Our central insight concerns how to construct transferable supervision for tasks whose training data are difficult to synthesize directly. As discussed in Section 1, spatial perception and multimodal long-context understanding underpin embodied systems and multimodal agents, respectively, yet obtaining reliable supervision for these settings is challenging. We hypothesize that identifying, binding, and composing distributed visual evidence constitute a shared training target. EvidenceConnect operationalizes this hypothesis through single-image, multi-region compositional questions, providing a tractable setting for constructing and auditing supervision. The research question is whether training in this simpler setting transfers to tasks involving multiple images, long contexts, and interactions.

Methodological contribution: The methodological contribution is the construction of multi-region compositional supervision with explicit checks of evidence dependence. Merely mentioning multiple regions in a question does not ensure that answering it requires combining their evidence. In Section 3.1, EvidenceConnect constructs coherent region groups and assigns regions explicit roles in aggregation, closed-set comparison, or dependency chains. Question generation also specifies each region’s contribution and the expected effect of removing it. Section 3.2 then addresses two distinct requirements: cross-model answer anchoring assesses answer reliability, while adaptive region-removal tests assess whether the intended evidence is actually needed. Agreement on an answer alone cannot detect questions that can be solved through a subset of the intended evidence. These design choices are therefore directly motivated by the challenge of producing meaningful compositional supervision.

Distinctions from prior work: Besides, Section 2 discusses the distinctions from closely related work. ZwZ [1] studies region-to-image supervision for fine-grained perception, whereas EvidenceConnect targets composition across a coherent group of regions. CRIT [2] focuses on cross-modal multi-hop reasoning, and HopChain [3] constructs instance-grounded dependent reasoning chains. EvidenceConnect supports aggregation, closed-set comparison, and dependency chains within a single-image setting, coupled with explicit region-level dependence checks. These distinctions concern the structure and validation of the supervision, beyond the broad pipeline stages listed in the review.

Empirical evidence: The experiments establish the practical value of this formulation. With the same Qwen3.5-4B backbone and GRPO recipe, Table 2 in Section 4.3 reports an average score of 53.86 for EvidenceConnect versus 49.86 for ZwZ, using 20.7K versus 73K training examples. Furthermore, Section 4.2 (Table 1) shows that training exclusively on single images improves multi-image long-context understanding and multimodal agent tasks. These findings support our hypothesis that within-image evidence composition provides useful supervision beyond the training setting.

We would appreciate clarification on which specific aspects of this formulation or verification design are considered incremental, and which closely related methods would help make the comparison more precise.

[1] Wei, Lai, et al. "Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception." ICML 2026.

[2] Sung, Junyoung, et al. "CRIT: Graph-Based Automatic Data Synthesis to Enhance Cross-Modal Multi-Hop Reasoning." arXiv preprint arXiv:2604.01634 (2026).

[3] Wang, Shenzhi, et al. "Hopchain: Multi-hop data synthesis for generalizable vision-language reasoning." arXiv preprint arXiv:2603.17024 (2026).

AuthorsOpenReview ↗

Response to Reviewer EUTm (2)

Q2. The authors only expriment with the Qwen series models. I would like to see the retraining results on other advanced open-source models.

A2. Thank you for this suggestion. To evaluate whether the benefits of EvidenceConnect extend beyond the Qwen family, we additionally trained GLM-4.1V-9B-Thinking on EvidenceConnect. We selected this model because it provides a capable backbone from a different model family while supporting native-resolution visual encoding, making it well suited to the fine-grained, multi-region perception targeted by EvidenceConnect. We compared the model after training on our proposed dataset with the original model below.

The results reveal that the trained model improves on all six benchmarks, with gains ranging from 2.00 to 3.62 percentage points. Its unweighted average score increases from 48.54 to 51.41, a gain of 2.87 percentage points. These results provide additional evidence that the benefits of EvidenceConnect extend to a second model family. We have updated our submission in Section 4.5 (Figure 3) and Appendix A.3 (Table 6).

BenchmarkGLM-4.1V-9BGLM-4.1V-9B trained on our datasetGain
HR-4K72.8874.88+2.00
HR-8K64.7568.37+3.62
CountQA28.9932.40+3.41
O3-Bench37.1039.71+2.61
ZoomBench42.8445.92+3.08
TreeBench44.6947.16+2.47
Avg48.5451.41+2.87
Reviewer_u4kkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a high-quality multimodal data synthesis pipeline EvidenceConnect. This framework first discovers relatable regions from high-resolution complex images by frontier MLLMs, then constructs privileged evidence layouts to synthesize aggregation, comparison/matching, or dependency-chain questions. EvidenceConnect further introduces rigorous question quality assurance procedure, including cross-model answer anchoring and adaptive drop-one testing to ensuring the correctness and difficulty of the generated question. Extensive experiments on several benchmarks and synthetic datasets demonstrate the effectiveness of the proposed data generation pipeline.

Critical strengths
  1. The data generation pipeline is well-motivated and reasonable. The implementation of EvidenceConnect is clearly written in main text and Appendix.

  2. The paper employs rigorous quality assurance on synthetic data. Cross-model Answer Anchoring and Adaptive Drop-one Testing help ensure the correctness of the QA pairs and the difficulty of generated data.

  3. The experiments in Table 2 strongly demonstrates substantial performance gains compared with other synthetic datasets.

Critical weaknesses
  1. The comparisons with Advanced Models in Table 1 are of limited value, because the 4B base model already consistently outperforms GPT and DeepSeek, while the 27B base model is already comparable to GLM. The improvements achieved by PerceptVLM demonstrate consistent gains over the base models, but do not establish that its superiority over GPT, GLM, and DeepSeek is attributable to the EvidenceConnect data. In my assessment, the experiment in Table 2 is the most important one for demonstrating the effectiveness of EvidenceConnect and therefore deserves a larger evaluation budget.

  2. The scope of applicability of the data generation pipeline is unclear. The experiments demonstrate that it can produce high-quality data from complex natural images such as those in SA-1B, but whether the same pipeline can also produce high-quality data from images in other domains remains unknown.

Questions for authors

See Weakness 2.

4 comments
AuthorsOpenReview ↗

Response to Reviewer u4kk (1)

Q1. The comparisons with Advanced Models in Table 1 are of limited value, because the 4B base model already consistently outperforms GPT and DeepSeek, while the 27B base model is already comparable to GLM. The improvements achieved by PerceptVLM demonstrate consistent gains over the base models, but do not establish that its superiority over GPT, GLM, and DeepSeek is attributable to the EvidenceConnect data. In my assessment, the experiment in Table 2 is the most important one for demonstrating the effectiveness of EvidenceConnect and therefore deserves a larger evaluation budget.

A1. Thank you for this observation. We agree that the external models in Table 1 primarily provide references for absolute performance, and that advantages already present in the Qwen backbones should not be attributed to EvidenceConnect. Comparisons under the same backbone and training recipe provide more direct evidence of the effectiveness of our training data.

At the same time, the strength of these backbones makes the improvements over them particularly informative. As you note, the Qwen base models already perform strongly against the external references on the evaluated tasks. Nevertheless, as reported in Section 4.2 (Table 1), training with EvidenceConnect further improves their average scores by 5.39, 4.96, and 4.99 points at 4B, 9B, and 27B, respectively, with gains on all 25 benchmarks at each scale. These results demonstrate that EvidenceConnect provides additional benefits even when the starting models already possess strong multimodal capabilities. Moreover, these improvements extend from single-image training to multi-image and interactive evaluations, supporting the breadth of transfer across the six task categories.

Table 2 in Section 4.3 provides the complementary, more direct comparison of training-data effectiveness. Under a common Qwen3.5-4B backbone and GRPO recipe, EvidenceConnect improves the average score across 11 benchmarks from 47.16 to 53.86 (+6.70 points), compared with 49.86 (+2.70 points) for ZwZ, the strongest alternative in this comparison. EvidenceConnect outperforms ZwZ on 10 of the 11 benchmarks, using 20.7K versus 73K training examples. We would further emphasize what Table 2 reveals about existing training data. Four of the five alternative datasets reduce the 11-benchmark average below the Qwen3.5-4B base model (47.16): MMHops reaches 42.73 (-4.43), TVG 45.48 (-1.68), F2RVLM 46.42 (-0.74), and CRIT 46.61 (-0.55), despite comparable or larger data volumes (19.4K-37K examples). Only ZwZ improves over the base model (49.86, +2.70). For five of the six existing training sets we evaluated, further RL training therefore does not improve the base model's perception at all. Thus, we believe the current evaluation already establishes two distinct findings: consistent improvements over already strong backbones across scales and tasks, and greater effectiveness than the alternative training datasets evaluated under a common backbone.

In addition, we agree that extending the dataset comparisons in Table 2 would further strengthen the evaluation. We therefore conducted an additional experiment using the DeepEyes [1] training dataset. We selected DeepEyes because it combines fine-grained visual perception, chart understanding, and multimodal reasoning data, making it a relevant and diverse training-data baseline for our evaluation. We evaluated the resulting model on seven perception benchmarks, with the results reported below.

Training on DeepEyes yields small improvements on HR-4K and O3, but decreases performance on the other five benchmarks. The average score changes from 62.08 to 61.63 (−0.44 points), with the largest decreases on CountQA (−2.49 points) and ZoomBench (−1.65 points). Thus, training on DeepEyes does not yield an overall improvement under this training and evaluation setting.

This additional experiment complements the dataset comparisons in Table 2: further training on diverse visual perception and reasoning data does not automatically translate into improved downstream perception. Together with the gains obtained using EvidenceConnect in the controlled comparisons, these findings underscore the importance of the training-data construction procedure. We have updated our submission in Section 4.3, with detailed results provided in Appendix A.2 (Table 5).

ModelsVStarTreeBenchHR-4KO3HR-8KZoomBenchCountQAAvg
Base82.2050.3782.3753.9178.3750.4136.9162.08
Trained on DeepEyes81.6849.3884.0055.0778.1248.7634.4261.63

[1] Zheng, Ziwei, et al. "DeepEyes: Incentivizing ‘Thinking with Images’ via Reinforcement Learning." International Conference on Learning Representations. Vol. 2026. 2026.

AuthorsOpenReview ↗

Response to Reviewer u4kk (2)

Q2. The scope of applicability of the data generation pipeline is unclear. The experiments demonstrate that it can produce high-quality data from complex natural images such as those in SA-1B, but whether the same pipeline can also produce high-quality data from images in other domains remains unknown.

A2. Thank you for highlighting the need to evaluate the pipeline beyond natural images. In our study, we selected SA-1B for its diverse natural scenes and high-resolution images, which provide rich visual evidence for constructing multi-region questions. Our original experiments demonstrate that training on data synthesized from natural images transfers to multiple downstream domains. To further examine the applicability of the data generation pipeline itself, we conducted an additional experiment on two different image domains.

Specifically, we applied EvidenceConnect to infographics from InfographicVQA [1] and mobile UI screenshots from ScreenQA [2], synthesizing 3,000 training examples in total. Starting from Qwen3.5-4B, we trained on these examples alone for three epochs. Compared with Qwen3.5-4B, the resulting model improves HRBench-4K from 82.37 to 86.00 and Insight-O3 from 53.91 to 57.39, with an average improvement of 1.87 over 7 benchmarks. These results provide additional evidence that EvidenceConnect can generate useful training data from infographic and GUI images, extending its demonstrated applicability beyond natural photographs. We have included these experiments in Section 4.5 of the revised manuscript, with details provided in Appendix A.4 (Table 7).

BenchmarkQwen3.5-4BAfter training
HRBench-4K82.3786.00
HRBench-8K78.3782.50
CountQA36.9138.02
Insight-O353.9157.39
VStar82.2084.82
ZoomBench50.4150.77
TreeBench50.3748.15
Avg62.0863.95

[1] Mathew, Minesh, et al. "Infographicvqa." 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2022.

[2] Hsiao, Yu-Chung, et al. "Screenqa: Large-scale question-answer pairs over mobile app screenshots." Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Thank you to the authors for their detailed responses, which have adequately addressed my concerns. I will maintain my original score.

Dear Reviewer u4kk,

Thank you for your positive feedback and for acknowledging that our rebuttal has adequately addressed your concerns. We sincerely appreciate your time and effort in reviewing our work and engaging in the discussion.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all three reviewers for their careful reviews, and are encouraged that they found the pipeline "well-motivated and reasonable" with "rigorous quality assurance on synthetic data" (Reviewer u4kk), "fairly complete" in treating "answer reliability, dependence on multi-region evidence, and referential ambiguity" (Reviewer i7Ls), and the evaluation "broad" with "good cross-task transfer" (Reviewer i7Ls).

Point-by-point replies are posted under each review. Because several requests recurred across reviews, we summarize the five new experiments run during the discussion period. All new content is highlighted in blue in the revised PDF.

1. Controlled ablations within our own pipeline (Reviewer i7Ls, Q1/Q3/Q4)

As noted in Q1, cross-dataset comparisons leave data quality and question distribution as confounders. We therefore ran two ablations inside our pipeline, holding fixed the source images, question-generation model, verification procedure, initialization (Qwen3.5-4B), data size (7K), GRPO configuration, and epochs (two).

  • Region markers. 99.82% of training examples carry at least one colored outline, and the same proportion holds per referenced region. Training without markers reaches 52.93 and with markers 54.08, against 49.41 for the base model. The supervision thus remains useful without outlines, and marker-based training transfers to evaluation inputs, which never contain outlines.
  • Single- vs. multi-region supervision. Single-region reaches 52.23 and multi-region 54.08 (+1.86), with higher scores on 9 of 12 benchmarks. This isolates evidence composition, the design choice defining EvidenceConnect, from the data source.

Section 4.4; Appendix A.1 (Table 4).

2. Generalization beyond the Qwen family (Reviewer EUTm, Q2)

GLM-4.1V-9B-Thinking trained on our data improves from 48.54 to 51.41 (+2.87), with gains on all six benchmarks (2.00-3.62). With the Qwen3-VL-2B/4B results, this extends the benefit to a second model family.

Section 4.5 (Figure 3); Appendix A.3 (Table 6).

3. Applicability beyond natural images (Reviewer u4kk, Q2)

Applying EvidenceConnect to InfographicVQA infographics and ScreenQA mobile UI screenshots, we synthesized 3K examples and trained Qwen3.5-4B on them alone for three epochs. The seven-benchmark mean rises from 62.08 to 63.95 (+1.87), indicating the pipeline is not specific to natural photographs.

Section 4.5; Appendix A.4 (Table 7).

4. An additional training-data baseline (Reviewer u4kk, Q1)

Training the same backbone on 20K DeepEyes examples with the same recipe moves the mean from 62.08 to 61.63, whereas our data reaches 68.83 in the identical setting, 7.19 points higher and better on all seven benchmarks. Diverse perception and reasoning data therefore do not by themselves yield downstream gains, underscoring the role of the construction procedure.

Section 4.3; Appendix A.2 (Table 5).

5. Broader external comparison, with narrowed claims (Reviewer i7Ls, Q2/Q5; Reviewer u4kk, Q1)

We added MiMo-V2.6-Pro, Step5, and Claude-Haiku-5.5, and completed the missing entries for the original references. Across its three sizes PerceptVLM attains the best score on 7 of 15 benchmarks; PerceptVLM-27B reaches 54.72 mean, above MiMo-V2.6-Pro (47.30) and Step5 (43.21/41.96), and exceeds the strongest external result by 5.23 on VStar, 4.47 on MindCube-Tiny, and 4.13 on ViewSpatial-Bench. In this comparison, only the closed-source Claude-Haiku-5.5 attains a higher mean, leading our largest, 27B-scale model by 2.18 and 2.94 points in its low and high settings.

Section 4.2; Appendix A.5 (Table 8).

On the technical contribution (Reviewer EUTm, Q1)

Our motivation is that identifying, binding, and composing distributed visual evidence form a shared training target for tasks whose own supervision is hard to synthesize, and that this target is trainable in the tractable single-image, multi-region setting. The methodological content lies in making composition necessary and verifiable: explicit region roles in aggregation, closed-set comparison, and dependency chains, plus region-removal tests of whether the intended evidence is actually required, which answer agreement alone cannot establish. Ablation 1 above gives direct evidence: holding the pipeline fixed and removing only the composition requirement costs 1.86 points. We would welcome the reviewer's view on which aspects remain incremental, and which related methods would sharpen the comparison.

We hope these additions address the shared concerns, and are glad to run further experiments during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.