acceptodds
Under review as a conference paper at ICLR 2027

ConScore-VG: Consistency-Score-Guided Test-Time Reinforcement Learning for Visual Grounding

Abstract

Visual grounding requires models to localize image regions according to language instructions and is a fundamental capability of vision-language models. Recent approaches based on supervised fine-tuning or reinforcement learning have achieved strong performance, but adapting them to new scenarios without additional annotations remains challenging. Test-time reinforcement learning (TTRL) offers a label-free alternative, yet extending it to visual grounding is challenging because continuous coordinate outputs are not naturally compatible with standard consensus-based pseudo-labeling, and spatial consensus among predictions does not necessarily imply correct localization. To address these challenges, we propose ConScore-VG, a consistency-guided TTRL framework that combines spatial and semantic consistency scores for reliable pseudo-label selection. The Spatial Consistency Score measures how well each candidate localization agrees with the overall spatial consensus, while the Semantic Consistency Score evaluates region-instruction alignment through visual information gain. Extensive experiments on referring expression comprehension (REC) and GUI grounding benchmarks demonstrate that ConScore-VG consistently improves grounding performance across both settings. Notably, ConScore-VG yields consistent gains on REC with only 256 unlabeled samples and improves over the baseline by 6.68% on ScreenSpot-v2. These results demonstrate the potential of TTRL for visual grounding and establish ConScore-VG as a data-efficient approach for adapting vision-language models to downstream grounding tasks without labels.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Grounding, Test-Time Reinforcement Learning, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JQf8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a pseudo-label construction method for bounding-box prediction tasks. Using these pseudo-labels, ConScore-VG performs test-time reinforcement learning with GRPO on only 256 samples to improve the model’s visual grounding performance.

Critical strengths
  1. The paper addresses the challenge of constructing pseudo-labels for continuous bounding-box outputs. Its spatial consistency score provides a simple and effective way to extend consensus-based selection to visual grounding, with Figure 3 showing that it favors candidates with higher GT IoU.
Critical weaknesses
  1. The experiments are conducted exclusively on Qwen2.5-VL-3B. However, more recent models such as Qwen3 and Qwen3.5 are available. Given that larger and more recent models generally exhibit superior grounding capabilities, it remains unclear whether the proposed method would provide consistent benefits to these more advanced models.

  2. In the main experiments presented in Table 1, the improvements on the test sets across the three benchmarks are only 0.3, 0.7, and 0.5, respectively. Furthermore, I think that applying test-time reinforcement learning specifically on grounding data could potentially affect the model’s general image understanding or other capabilities.

Questions for authors
  1. Have you evaluated the proposed method on more advanced models with stronger grounding capabilities?

  2. Have you evaluated whether grounding-specific test-time reinforcement learning affects the model’s general image understanding or other capabilities? Given the modest grounding gains in Table 1, evidence of whether adaptation preserves these capabilities would help assess the overall benefit.

  3. In Table 4, adding the semantic consistency score to spatial-only yields small improvements. Since the combined score may favor spatially clustered boxes, which could cover similar semantic content, it remains unclear how often semantic scoring improves pseudo-label selection. Could you provide additional experiments, such as ablations on the other benchmarks or a quantitative analysis similar to Figure 3 comparing spatial-only and combined selection?

  4. The improvements shown in Table 1 are small, and since the experiments rely on only 256 randomly selected samples, it remains unclear whether these results are sensitive to the specific sample selection.

2 comments

Dear Reviewer JQf8,

Thank you for your valuable feedback and thoughtful comments on our manuscript. We provide our responses below and have incorporated the corresponding changes into the revised manuscript.

W1+Q1(Experiments with Newer Models)

A: We initially selected Qwen2.5-VL-3B-Instruct as our base model to ensure a fair comparison with the baseline. Following your suggestion, we extend our evaluation to Qwen3-VL-4B-Instruct, a newer backbone with stronger grounding capabilities. We evaluate ConScore-VG on GroundingSuite for general visual grounding and on the ScreenSpot series benchmarks for GUI grounding. The table below summarizes the results.

MethodGroundingSuiteSSv2.avgSSv1.avgSSPro.avg
Qwen3-VL-4B-Instruct71.785.8582.7843.07
ConScore-VG73.088.5285.3847.82

The results show that ConScore-VG consistently improves grounding performance on the more advanced model with stronger grounding capabilities. We discuss these findings on page 10 of the revised manuscript, where Table 6 presents the quantitative comparison.

W2 (Performance on RefCOCOs)

A: The magnitude of ConScore-VG’s gains on RefCOCOs is acceptable in light of results reported by the representative visual grounding method [1]. Beyond RefCOCOs, ConScore-VG achieves substantial gains of 4.0% on GroundingSuite for general visual grounding (Table 2, Page 9) and 6.68% on ScreenSpot-v2 for GUI grounding (Table 3, Page 9). Together, these results demonstrate consistent improvements across grounding scenarios, supporting ConScore-VG’s reliability.

W2+Q2 (Preservation of General Image Understanding Capabilities)

A: ConScore-VG uses rewards derived from the model’s own predictions to perform only a few policy updates on a small unlabeled adaptation set. We therefore expect it to preserve performance on general image understanding tasks. To examine this further, we evaluate the model adapted on GroundingSuite across three representative general VQA benchmarks, including MMBench, RealWorldQA and MMStar (Appendix E.2 and Table 8, Page 23). These results indicate that our proposed ConScore-VG enhances visual grounding performance without degrading the model's general vision-language capabilities, consistent with our earlier analysis.

MethodMMBench-ENMMBench-CNRealWorldQAMMStar
Qwen2.5-VL-3B-Instruct76.576.965.152.8
ConScore-VG76.576.966.552.6

Q3 Ablation of Semantic and Spatial Consistency Scores on the other benchmarks.

A: Following your suggestion, we examine how often adding semantic consistency scoring improves pseudo-label selection on two GUI grounding benchmarks, ScreenSpot-v2 and ScreenSpot-Pro. We report the results in Table 5 (Page 10) of the revised manuscript. The full ConScore-VG combines both scores and achieves the best performance across all benchmarks. On ScreenSpot-Pro, a challenging full-screen GUI benchmark with more visual distractors, it outperforms the spatial-only variant by 3.10%. This gain suggests that semantic consistency helps select candidate predictions better aligned with the referring instruction, providing more reliable pseudo-labels for policy optimization.

MethodSemanticSpatialRefCOCORefCOCO+RefCOCOgSSv2.avgSSPro.avg
Base89.282.985.180.1120.18
Semantic only✓89.483.785.481.7624.48
Spatial only✓89.783.485.986.4028.08
ConScore-VG✓✓90.083.986.186.7931.18

Q4 Ablation Study on Different Adaptation Subsets

A: We investigate the impact of adaptation sample selection on ConScore-VG by randomly drawing three distinct subsets from GroundingSuite, each containing 256 unlabeled test samples. The experimental results are as follows. The results show that performance gains vary across randomly sampled adaptation sets, potentially reflecting differences in sample difficulty. In particular, sets containing too many examples that the base model cannot solve may limit the gains from policy optimization. Despite this variability, ConScore-VG consistently improves performance across the sampled sets, supporting its robustness to variations in adaptation data. We have added this discussion on the Appendix E.8 (Page 25) of the revised manuscript and Table 10 (Page 26) presents the quantitative results.

SubsetStuffPartMultiSingleAll
Base94.031.962.266.969.0
Subset196.130.878.367.573.0↑+4.0
Subset296.441.574.371.675.3↑+6.3
Subset396.027.972.867.171.4↑+2.4

We again thank the reviewer and are happy to clarify further if any questions remain.


References

[1] Liu, Yuqi, et al. Visionreasoner: Unified reasoning-integrated visual perception via reinforcement learning. ICLR, 2026.

Dear Reviewer JQf8,

Thank you for your time and effort in reviewing our paper and for your valuable suggestions. We have carefully responded to your comments in our rebuttal and revised manuscript. As the discussion period is coming to an end, we would like to check whether our responses and the revision have addressed your concerns. If you are satisfied, we kindly request you to consider updating your score to reflect the improvements. We remain committed to addressing any remaining points you may have.

Best regards,

The Authors

Reviewer_a1gfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

see weakness.

Critical strengths

see weakness.

Critical weaknesses

The paper studies test time reinforcement learning for visual grounding with unlabeled samples, but the literature review, technical novelty, empirical depth, and presentation are insufficient for ICLR.

  1. The literature review is inadequate and shows limited awareness of unsupervised visual grounding and pseudo label based localization work.
    The paper discusses TTRL and GUI RCPO but does not systematically cover prior work on pseudo label selection, self training, noisy label learning, and weakly supervised or unlabeled visual grounding. This omission makes the claimed gap and the novelty of consistency guided pseudo labels difficult to assess.

  2. The technical idea is common and the contribution is incremental.
    Combining spatial consensus among sampled boxes with a semantic consistency score is a natural extension of self consistency, pseudo label selection, and perturbation based scoring, and the paper does not provide a strong argument for why this combination is fundamentally new. The semantic score based on region perturbation and information gain is also close to existing attribution and contrastive scoring ideas, so the method reads as an engineering combination rather than a distinct technical advance.

  3. The empirical workload is not solid and feels rushed.
    The main adaptation uses only 256 unlabeled samples, three epochs, and a single Qwen2.5 VL 3B backbone, with limited seeds and no significance testing. Several comparisons rely on reproduced baselines or borrowed GUI rewards, and the ablations do not sufficiently separate pseudo label selection from policy optimization or control for extra sampling and scoring cost.

  4. The experimental protocol and analysis do not fully support the claims.
    ConScore without updates and ConScore GRPO are compared, but the gains may come from additional rollouts and selection rather than from test time reinforcement learning itself. The paper also lacks detailed analysis of pseudo label noise, failure cases, and sensitivity to the perturbation design, the quantile threshold, and the scoring prompt, which are central to the proposed semantic score.

  5. The presentation is difficult to follow and key details are poorly organized.
    The manuscript contains long line number blocks, inconsistent equation numbering, dense and unclear tables, and important protocol details buried in the appendix. The writing also has a template generated flavor, with repeated claims and vague phrasing, which makes the technical contribution harder to verify.

Questions for authors

see weakness.

4 comments

Dear Reviewer a1gf,

Thank you for your careful review and constructive feedback. We have addressed your comments point by point below and incorporated the corresponding changes into the revised manuscript.

W1 (Related Work on Unsupervised Visual Grounding and Pseudo-Label-Based Localization).

A: Specifically, prior efforts that reduce reliance on bounding-box annotations primarily fall into three training-time paradigms:

  1. Semi-supervised detection and grounding methods [1] use a teacher model trained with labeled data to generate pseudo-boxes for unlabeled images.
  2. Unsupervised visual grounding methods [2] rely on external pretrained detectors to extract candidate regions.
  3. Weakly supervised visual grounding methods [3] learn from image–query pairs without bounding-box annotations by matching queries to region proposals generated offline by external detectors.

ConScore-VG differs fundamentally from them in three key aspects:

  1. Test-time vs. Training-time: ConScore-VG performs test-time reinforcement learning (TTRL) on a small set of unlabeled test samples.
  2. Self-contained vs. Detector-dependent: ConScore-VG requires no external detectors and derives rewards directly from its own rollouts.
  3. Pretrained VLMs vs. Specialist Models: ConScore-VG adapts pretrained VLMs by leveraging their intrinsic multimodal priors.

Table 1 (Page 8) shows that ConScore-VG substantially outperforms conventional pseudo-labeling baselines. We have expanded the corresponding discussion in Appendix A.1 (Page 19) of the revised manuscript.

MethodRefCOCO valRefCOCO testARefCOCO+ valRefCOCO+ testARefCOCOg valRefCOCOg testSum
Pseudo-Q [4]56.058.338.945.146.347.4292.0
Omni-Q [2]67.471.646.251.763.662.4362.9
CPL [5]70.774.651.858.360.260.1375.7
SVCR [3]74.576.956.464.2---
ConScore-VG90.092.483.989.486.186.0527.8

W2 (Novelty of Combining Spatial and Semantic Consistency)

A: We respectfully clarify that our proposed ConScore-VG is a principled synergistic design specifically tailored to Test-Time Reinforcement Learning (TTRL) in continuous coordinate spaces:

  1. Adapting Consensus to Continuous Coordinates : Most TTRL methods rely on majority voting over discrete answers, which does not directly extend to continuous bounding-box predictions. Our Spatial Consistency Score bridges this gap by measuring group-wise pairwise IoU consensus among sampled rollouts. Compared with [6], our formulation preserves object extent and yields scores strongly and positively correlated with localization quality (Figure 3, Page 5).
  2. Breaking Confirmation Bias via Task Duality: Recent studies [7, 8] show that relying solely on prediction consensus can lead to confirmation bias. To mitigate this bias, ConScore-VG uses task duality (Section 3.2.2, Page 5) to verify candidate boxes without relying on external models.
  3. Technical Novelty of the Semantic Consistency Score: Our semantic consistency score is novel in how it verifies candidate regions through visual referring, the inverse of visual grounding, not merely in its use of contrastive scoring. Unlike [9], ConScore-VG perturbs only the candidate region and preserves the surrounding context.

W3 (Experimental Workload)

W3/Q1: Concerns About the Experimental Workload

A: Our existing experiments follow established evaluation practice [6, 10] and have provided sufficient empirical support for the original manuscript’s main claims. For further evidence, we also evaluate ConScore-VG on GroundingSuite [11], a more challenging benchmark. Using only 256 unlabeled samples , ConScore-VG substantially improves visual grounding, enabling a 3B model to achieve performance comparable to the closed-source Gemini-1.5-Pro (Table 2, Page 9).

MethodGemini-1.5-ProGPT-4oInternVL3-8BDeepSeek-VL2Ferret-13BQwen2.5-VL-3BConScore-VG
GroundingSuite74.317.352.360.855.169.073.0

W3/Q2: Adaptation Sample Size.

A: A key advantage of ConScore-VG is improving performance with few unlabeled samples. We examine the effect of adaptation sample size with optimization steps held fixed in the original manuscript (Figure 8, Page 26) and present additional GroundingSuite experiments with a fixed number of training epochs, as shown below (Appendix E.9 and Table 14, Page 26). These results show that moderate adaptation sets yield substantial gains, consistent with recent work [7], while larger sets offer limited additional benefits and substantially increase adaptation time and computational cost.

SizeStuffPartMultiSingleAll
Base94.031.962.266.969.0
N=25696.130.878.367.573.0↑+4.0
N=51296.956.389.377.182.5↑+13.5
N=102496.110.178.265.169.5↑+0.5

W3/Q3: Number of Adaptation Epochs.

A: In the original manuscript, we examined the effect of the number of adaptation epochs (Appendix E.7 and Table 12, Page 25). Additional epochs yield only marginal gains, while prolonged adaptation can slightly reduce performance and increase runtime.

W3/Q4: Single-Backbone Evaluation.

A: Following [6], we used Qwen2.5-VL-3B-Instruct in our main experiments to enable direct comparison using the same backbone within our compute budget. At your suggestion, we also evaluated ConScore-VG with Qwen3-VL-4B-Instruct. ConScore-VG consistently improves grounding performance across both backbones (Table 6, Page 10).

MethodGroundingSuiteSSv2.avgSSv1.avgSSPro.avg
Qwen3-VL-4B-Instruct71.785.8582.7843.07
ConScore-VG73.088.5285.3847.82

W3/Q5: Random Seeds and Statistical Significance.

A: Across multiple random seeds on GroundingSuite, ConScore-VG consistently improves grounding performance over the base model, suggesting that its gains are not driven by a particular seed (Appendix E.6, Page 24; Table 11, Page 25).

MethodStuffPartMultiSingleAll
Base94.031.962.266.969.0
Seed=4296.130.878.367.573.0↑+4.0
Seed=24295.931.774.466.571.8↑+2.8
Seed=44295.927.075.366.771.6↑+2.6
Mean ± Std96.0±0.129.8±2.576.0±2.066.9±0.572.1±0.8↑+3.1

W3/Q6: GUI Reward Reuse, Baseline Reproduction, and Disentangling the Effects of Pseudo-Label Selection and Policy Optimization

A:

  1. GUI Reward Reuse: ConScore-VG selects reliable pseudo-boxes using consistency scores without introducing new reward functions. This enables seamless integration with standard label-based rewards, a key advantage of the framework.
  2. Baseline Reproduction: Following [6, 10], we reproduce the base models’ results to ensure fair comparison and isolate ConScore-VG’s contribution. Reproducing prior work [6] on RefCOCOs is also essential to demonstrate that its design is ill-suited to general visual grounding (Table 4 and Figure 5, Page 9).
  3. Disentangling the Effects of Pseudo-Label Selection and Policy Optimization. The results in the original manuscript already disentangle the respective contributions of pseudo-label selection and policy optimization. Table 1 (Page 8) shows that pseudo-label selection alone improves grounding without parameter updates. Policy optimization updates the model using selected pseudo-labels from a small adaptation set, yielding gains under greedy decoding. Table 5 (Page 10) further shows that pseudo-label quality affects the gains from policy optimization.

W3/Q7: Computational Overhead.

A: We compare ConScore-VG with label-based GRPO, measuring per-step training time and peak memory usage per device. ConScore-VG improves grounding through test-time adaptation on unlabeled samples while incurring only modest computational overhead (Table 9, Page 24).

MethodTime Per Step (seconds)Peak Memory Usage (GiB)
Label-based GRPO17.5442.70
ConScore-VG21.7446.56

W4 (More Ablations)

W4/Q1: The gains may stem from additional rollouts and candidate selection.

A: Table 1 (Page 8) shows that ConScore without updates improves grounding through additional rollouts and selection but substantially increases inference time. ConScore-GRPO incorporates these gains into model parameters through test-time reinforcement learning on only 256 unlabeled examples, retaining improved performance under greedy decoding.

W4/Q2: Failure Cases

A: We analyze failure cases of ConScore-VG, as illustrated in Figure 13 (page 29). These cases fall into two categories: mismatches between annotations and referring instructions, and genuine localization errors. The former affect evaluation but not policy optimization, as ConScore-VG relies on its own predictions for supervision. The latter mainly arise with very small targets or complex instructions. Addressing them is orthogonal to our current research focus and offers a direction for future work.

W4/Q3: Ablation Study on Perturbation Strategies.

A: Following your suggestion, we compare three perturbation strategies: mean padding, zero padding, and Gaussian blur. As shown below, the mean padding described in Appendix B.1 (Page 20) performs best overall, possibly because it suppresses discriminative cues while better preserving natural image statistics. In contrast, zero padding introduces a high-contrast black patch that may act as an unintended visual cue, whereas Gaussian blur may leave some discriminative details intact.

StrategyRefCOCORefCOCO+RefCOCOgGroundingSuite
Base89.282.985.169.0
Mean Padding90.083.986.173.0
Zero Padding90.083.785.571.0
Gaussian Blur89.583.985.771.2

W4/Q4: Ablation Study on The Quantile Parameter ρ\rho.

A: ConScore-VG consistently outperforms the base model across ρ∈{0,0.25,0.5,0.75}\rho \in \{0, 0.25, 0.5, 0.75\}, indicating robustness to this hyperparameter (Figure 7, Page 24). Among these settings, ρ=0.25\rho = 0.25 yields the highest accuracy on all three datasets. This may reflect a trade-off, as ρ=0\rho = 0 retains more unreliable token signals, whereas overly large values may down-weight informative tokens.

Quantile ParameterRefCOCORefCOCO+RefCOCOg
Base89.282.985.1
ρ=0\rho=089.683.585.8
ρ=0.25\rho=0.2590.083.986.1
ρ=0.5\rho=0.589.783.885.6
ρ=0.75\rho=0.7589.983.785.7

W4/Q5: Analysis of Pseudo-Label Noise.

A: We assess the impact of pseudo-label noise by varying the proportion of adaptation samples with incorrect initial predictions. As shown in the table below, lower noise rates yield larger performance gains, consistent with the expectation that more reliable pseudo-labels provide better supervision.

Noise Ratio(%)0255075100Base
GroundingSuite73.071.068.965.159.269.0

W4/Q6: Sensitivity to the Scoring Prompt

A: Our experimental setup follows representative work on visual grounding with large vision–language models [10,12] and on test-time reinforcement learning [6,13,14], in which prompt wording is not a primary focus of ablation studies. Since tuning the scoring prompt is not a core innovation of our method, we do not include a separate analysis of sensitivity to its wording. To facilitate reproducibility, we provide the exact scoring prompt used in our experiments in Appendix F (Page 30) of the revised manuscript.

W5 (Manuscript Writing and Organization)

We have reviewed the manuscript and found no such issues, as explained below.

  1. Line Numbers and Equation Numbering. Line numbers in the PyTorch-style pseudocode (Appendix B.2, Page 21) are included for readability, while all displayed equations are numbered sequentially throughout the manuscript without omissions or duplicates.
  2. Organization of Protocol Details. We present all essential experimental protocol details in Section 4.1 (Experimental Setup, Page 7) of the revised manuscript to keep the main text easy to follow, while detailing supplementary settings in the Appendix to accommodate the page limit.
  3. Table Layout and Writing Style. Regarding the table layout, our tables present key experimental comparisons within the main manuscript’s space constraints, with descriptive captions linking each table to the corresponding analysis. As for the repeated claims, we revisit core findings where relevant to maintain continuity and keep individual sections coherent and self-contained, consistent with common academic writing practices, including those followed in ICLR papers.

We again thank the reviewer and are happy to address any remaining concerns.


References

[1] Xu, Mengde, et al. End-to-end semi-supervised object detection with soft teacher. ICCV, 2021.

[2] Wang, Sai, Yutian Lin, and Yu Wu. Omni-Q: Omni-directional scene understanding for unsupervised visual grounding. CVPR 2024.

[3] Xu, Lian, et al. Learning structured visual compositional representations for weakly supervised referring expression comprehension. ECCV, 2026.

[4] Jiang, Haojun, et al. Pseudo-q: Generating pseudo language queries for visual grounding. CVPR 2022.

[5] Liu, Yang, et al. Confidence-aware pseudo-label learning for weakly supervised visual grounding. ICCV, 2023.

[6] Du, Yong, et al. Test-time reinforcement learning for gui grounding via region consistency. AAAI, 2026.

[7] He, Bingxiang, et al. How Far Can Unsupervised RLVR Scale LLM Training?. ICLR, 2026.

[8] Lin, Hongxiang, et al. Detecting and Mitigating the Correct-Answer Extinction Window in Test-Time Reinforcement Learning with Majority Voting. arXiv preprint arXiv:2605.19444 (2026).

[9] Wang, Zhenhailong, et al. Perception-aware policy optimization for multimodal reasoning. ICLR, 2026.

[10] Liu, Yuqi, et al. Visionreasoner: Unified reasoning-integrated visual perception via reinforcement learning. ICLR, 2026.

[11] Hu, Rui, et al. Groundingsuite: Measuring complex multi-granular pixel grounding. ICCV, 2025.

[12] Shen, Haozhan, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615 (2025).

[13] Zuo, Yuxin, et al. Ttrl: Test-time reinforcement learning. NeurIPS, 2026.

[14] Singh, Akshit, et al. Ttrv: Test-time reinforcement learning for vision language models. CVPR, 2026.

Dear Reviewer a1gf,

We appreciate your detailed comments and have made significant efforts to address them in our rebuttal and revised manuscript. The discussion period will end in about 24 hours. If you have any remaining questions about our responses or the revision, please let us know, and we will reply as soon as possible before the deadline. If your concerns have been addressed, we hope you will take this into account in your final assessment.

Best regards,

The Authors

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

The paper addresses a core challenge: how to estimate a surrogate localization target as pseudo labels used to calculate reward of each sampled trajectory in GRPO algorithm.

Critical strengths
  1. The use of spatial consistency score to select pseudo labels is well motivated, authors demonstrate that stronger consistency with the policy’s spatial consensus is associated with better localization quality.
  2. To address the concern that multiple sampled predictions may exhibit strong spatial agreement around a region that is semantically inconsistent with the language instruction, causing consensus-based selection to reinforce an incorrect localization, authors inverse visual grounding into region description generation and use the token-level log-probability differences of original instruction computed before and after perturbing the candidate region as the semantic consistency score. The analysis in Fig. 4 is convincing.
Critical weaknesses

N/A

Questions for authors

N/A

1 comment

Dear Reviewer WpHs,

Thank you for taking the time to review our manuscript and for your positive assessment of ConScore-VG. We particularly appreciate your recognition of the rationale for using spatial consistency and of the role of semantic consistency in addressing the limitations of relying on spatial agreement alone.

ConScore-VG is a test-time reinforcement learning method that uses complementary spatial and semantic consistency scores to select pseudo-labels for reward computation. Across both general visual grounding and GUI grounding, ConScore-VG achieves consistent performance gains with only a small number of unlabeled examples and a few adaptation steps, demonstrating its sample efficiency.

Thank you again for your thoughtful feedback. We would be happy to clarify any remaining questions.

Best,
Authors of Submission 23138

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.