- 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
- 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
-
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.
-
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
-
Have you evaluated the proposed method on more advanced models with stronger grounding capabilities?
-
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.
-
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?
-
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 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
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.
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.
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.
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.
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.