acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Adversarial Patch Attacks on VLA: From Pixel Space to Physical Reality

12 views

Abstract

Vision-Language-Action (VLA) models are increasingly deployed in safety-critical manipulation tasks, making adversarial robustness a security-critical concern. Recent studies report that adversarial patches can reduce task success rate to near zero. But how much of this attack strength survives when a patch is fixed to a physical surface? We revisit this question by separating patch generation from patch evaluation. Pixel-space compositing places a patch directly in the input image, whereas a fixed physical texture must produce all of its appearances through camera projection. We cross these two settings in a study on OpenVLA, OpenVLA-OFT, and . Switching from pixel-space to geometry-consistent evaluation raises task success rate in all 24 combinations, by 3.0-69.2 percentage points. Re-optimizing under geometric constraints then recovers attack strength in 23 of these combinations. For example, OpenVLA's success rate on LIBERO-Spatial under UADA attack changes from 0.2% to 52.0% when evaluation changes, and to 17.4% after re-optimization. Printed patches also disrupt a physical robot task. Finally, we examine how the same constraints affect adversarial training, finding substantial gains on VLA models. Together, these results show why physical patch risk should be evaluated under an explicit deployment model, with attack generation and evaluation considered separately.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vision-language-action model, adversarial robustness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZzxiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper shows that adversarial patches added directly to images can overestimate their real-world impact on VLA models. Accounting for physical geometry weakens existing attacks, but re-optimizing patches under these constraints makes them stronger again. Real-robot experiments confirm that printed patches can still disrupt robot behavior, while geometry-consistent adversarial training can improve robustness.

Critical strengths
  1. A meaningful problem & better evaluation protocol. Separating patch generation from evaluation under physical constraints clarifies how deployment assumptions affect measured VLA vulnerability.

  2. Consistent empirical evidence. The deployment gap appears in all 24 tested combinations, while re-optimization strengthens attacks in 23, providing strong support for the central finding across different models and tasks.

  3. Practical relevance for robot robustness. Printed-patch attacks on a real robot and improved robustness through geometry-consistent adversarial training demonstrate concrete implications for both vulnerability assessment and defense.

Critical weaknesses
  1. Scope for deeper methodological insight. The paper addresses a meaningful and practically relevant problem, although its approach largely builds on existing attack objectives and geometric projection. Further analysis of the specific roles of temporal and multi-view consistency would help clarify the additional insights contributed to VLA robustness.

  2. Limited benchmark diversity. Although the experimental evaluation is extensive, covering 24 model–attack–suite combinations, the simulation experiments are confined to LIBERO. Evaluation on additional benchmarks would strengthen confidence in the generality of the findings across different tasks and environments.

Questions for authors
  1. LIBERO-Long robustness. Could the authors report clean performance before and after AT to help explain the low libero-long success rates in Table A7?

  2. LoRA-rank sensitivity. Are the large performance differences across ranks in Table A5 consistent across runs, and was Pixel AT tuned comparably?

  3. Physical evaluation. Could the authors explain the difference between Tables 4–5 (16.7% vs. 10.0%), report per-scene clean performance in Table 7, and clarify whether patches were re-optimized for each defense?

3 comments
AuthorsOpenReview ↗

Thank you for the valuable review

Thank you for the thoughtful review. We appreciate your recognition of the evaluation protocol, consistent empirical evidence, and practical implications. Below we clarify the existing evidence on temporal and multi-view consistency, rank sensitivity, and adaptive physical evaluation.

W1: Further analysis of temporal and multi-view consistency.

Thanks for your valuable suggestion. We have examined the role of multi-view consistency in Table 2. A physical patch must use one texture across camera views, so a texture optimized for one view may be less effective when observed from another. Under the same geometry-consistent evaluation, joint optimization reduces OpenVLA-OFT task success from 31.6% for either single-view texture to 25.2%. On π₀, success decreases from 85.6–86.4% to 72.0%. These results show the benefit of optimizing the shared texture through both camera projections.

For temporal consistency, the patch stays at a fixed world location, while its appearance changes with the camera during robot execution. This matters because the robot’s actions affect subsequent observations. We examine the resulting behavior in Section 4.4 through paired clean and attacked rollouts. Geometry-consistent EDPA increases the mean end-effector deviation from 5.74 cm to 6.56 cm and causes earlier divergence from clean execution. These results show that the attack affects the robot throughout execution, beyond the final task outcome. We will make these insights clearer in the discussion of Table 2 and Figure 3.

W2: Limited benchmark diversity.

Thanks for your valuable suggestion. We have added experiments with π0\pi_0 on five VLABench tasks in Appendix B (Table A1), using a VLABench-finetuned checkpoint and demonstrations from this benchmark to optimize new patches jointly across the five tasks. Each task is evaluated with 50 rollouts under clean, P/P, P/G, and G/G settings.

SettingCleanP/PP/GG/G
Average task success (%)49.238.042.832.4

On average, transferring pixel-generated patches from P/P to P/G increases task success from 38.0% to 42.8%, while re-optimization under G reduces it to 32.4%. These results support the overall deployment-gap and re-optimization trends beyond LIBERO. We have also added a reference to these experiments in Section 4.1.

Q1: Could the authors report clean performance before and after AT to help explain the low LIBERO-Long success rates in Table A7?

Thanks for your valuable suggestion. We evaluated the three OpenVLA-OFT models on LIBERO-Long without patches, using 500 episodes per model. The clean success rates are reported below.

ModelClean success rate (%)
Vanilla14.8
Pixel AT0.0
Geometry AT7.6

Q2: Are the large performance differences across ranks in Table A5 consistent across runs, and was Pixel AT tuned comparably?

Thanks for raising this question. We used rank 8 for both Pixel AT and Geometry AT from the outset, without a rank search for either method. The other ranks in Table A8 were evaluated afterward to examine sensitivity. Each configuration used one randomly selected training seed. We will clarify this experimental procedure.

Following your suggestion, we additionally trained Pixel AT with rank 16 for 100,000 steps and re-optimized a geometry-consistent EDPA patch against this model. We evaluated it over 500 episodes on LIBERO-Spatial, obtaining 81.0% clean success and 38.0% success under attack. The comparison with the existing results is shown below, higher success indicates stronger robustness.

LoRA rankPixel ATGeometry AT
855.00%72.40%
1638.00%46.80%

Geometry AT outperforms Pixel AT at both tested ranks, by 17.4 and 8.8 percentage points, respectively. This additional comparison supports the benefit of geometry-consistent training beyond rank 8.

Q3: Physical evaluation.

Thanks for raising these questions. The 16.7% in Table 4 and 10.0% in Table 5 correspond to different patch-placement settings: Table 4 uses random placement, whereas the default condition in Table 5 fixes the patch on the right side of the workspace. We will clarify this distinction in the captions.

We also report the clean success rates (%) for both scenes below and will include them in Table 7. Each unseen-scene result is based on 30 trials.

ModelOriginal sceneUnseen scene
Vanilla90.080.0
Pixel AT73.360.0
Geometry AT86.766.7

Finally, the patches were re-optimized separately against each evaluated model. The defense results in Tables 4 and 7 therefore use adaptive attacks, rather than patches generated against the vanilla model alone. We will state this explicitly in the experimental description.

Thank you for the clarifications and extra VLABench exp. Meaningful topic, I will maintain my positive recommendation.

AuthorsOpenReview ↗

Response to reviewer Zzxi's reply

Dear Reviewer Zzxi,

Thank you again for your timely feedback and your efforts in reviewing our rebuttal and additional experiments. We sincerely appreciate your recognition of the meaningful topic and your continued positive recommendation. If you have any additional questions, we are glad to further answer them.

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

The paper observes that existing patch attacks on VLA models are optimized and evaluated by pasting a patch at a fixed pixel location in every frame and view, which does not reflect a physical patch on the table. The authors introduce a geometry-consistent rendering operator (planar homography from calibrated cameras plus occupancy-based placement) and cross pixel-space and geometry-consistent settings at generation and evaluation time. Across OpenVLA, OpenVLA-OFT and \pi_0 on four LIBERO suites, pixel-generated patches lose much of their effect when deployed geometrically, while re-optimizing under geometric constraints recovers most of it. The paper also reports multi-view joint optimization, real-robot experiments with printed patches, and a LoRA-based geometry-aware adversarial training.

Critical strengths
  1. A timely question. Prior VLA patch attacks assume pixel-level control of a fixed image region in every frame (and, for EDPA, a separate texture per camera). The paper makes this implicit threat model explicit and asks how much attack strength survives for a single texture fixed in the workspace.
  2. A clean experimental design. The 2×2 generation/evaluation design (Table 1) separates the transfer loss of an existing patch (P/P → P/G) from the gain of re-optimizing for deployment (P/G → G/G), with consistent trends across 24 closed-loop model–attack–suite combinations.
  3. A balanced conclusion. Pixel-space numbers should not be read as physical risk, but weak transfer does not imply robustness: re-optimization recovers much of the attack (e.g., 0.2% → 52.0% → 17.4% on OpenVLA/UADA/Spatial). The multi-view and printed-patch experiments add further support.
Critical weaknesses
  1. Overstated physical attack. The proposed method only adds planar homography and placement, then pastes the patch onto the frame. Lighting, print color, camera response and occlusion by the arm or objects are not modeled (App. C only has synthetic occlusion, and 20% of it already cuts the EDPA gap from 7.0 to 1.6). So G is a better geometric overlay, not physical deployment.
  2. Sim and real results mismatch. In sim \pi_0 is barely hurt in any setting (72.0–79.0% on Spatial, up to 96.4% elsewhere; Table A4). On the real robot, the printed original EDPA patch drops \pi_0 from 90.0% to 23.3%, close to the geometry-consistent one (16.7%, two trials apart), and stays robust in Table 5. Different task and data, sure, but the pixel-vs-geometry gap that motivates the paper does not show up. Also, Table 3 compares the 30×30 patch with P/G; the fair comparison is G/G (Table A2), where the geometry-consistent attack is stronger (Spatial 17.4 vs. 22.8, Goal 2.0 vs. 33.0, Long 9.2 vs. 18.6). The results support "pixel patches transfer poorly to planar overlays in sim", not "pixel-space evaluation overstates physical risk".
  3. Limited novelty. Optimizing under the deployment transformation is the core idea of EOT and physical patches (Athalye et al., 2018; Brown et al., 2017; Eykholt et al., 2018), and homography rendering is standard there. Bringing it to VLA attacks is useful, but this is mainly an evaluation-protocol paper and should be framed that way.
Questions for authors
  1. Could the P/G and G/G results be recomputed for at least one model with occlusion-aware compositing and photometric variation (lighting, print color/material), to test whether the conclusions hold beyond geometric alignment?
  2. In simulation π0 is only mildly affected under all settings, yet printed patches reduce its real-robot success to 16.7–23.3%, with little difference between original and geometry-consistent EDPA. How do you reconcile this with the claim that pixel-generated patches transfer poorly to physical deployment?
  3. In the area-matched comparison (Table 3), why is the 30×30-pixel attack compared against P/G rather than G/G? Against G/G, the geometry-consistent attack is stronger on three of four suites; how does this square with the claim that pixel-space evaluation overstates the threat?
4 comments
AuthorsOpenReview ↗

Thank you for the valuable review (1/2)

Thank you for the careful review. We appreciate your recognition of the controlled generation/evaluation design and your suggestions for clarifying its implications. Below we distinguish what the current experiments establish from the broader physical effects that remain outside our simulation model.

W1: Physical factors beyond geometric alignment.

We use G to isolate how deployment geometry affects VLA patch attacks. This controlled setting allows us to examine the mismatch between direct pixel modifications and a patch fixed in the workspace. Our real-robot experiments further evaluate printed patches under actual physical conditions (Section 4.6). We have also added experiments with dynamic occlusion and photometric variations in Appendix F (Table A7). Please see our response to Q1 for the details.

W2: Simulation–real-world discrepancy and the area-matched comparison.

Please see our response to Q2 for the comparison between simulation and real-robot results, and our response to Q3 for the area-matched comparison and how it relates to our findings on the deployment mismatch.

W3: Limited methodological novelty relative to EOT and physical patches.

We respectfully disagree with this assessment. Our main contribution is that we identify an important mismatch overlooked in existing VLA patch studies. These studies evaluate attacks by directly modifying image pixels, while a physical patch must produce its appearances across views and timesteps from one fixed texture and placement. The original attack settings already include EOT, but their pixel-space evaluation still leaves this mismatch unresolved. Our work brings this issue to light and systematically examines its consequences.

The results show that this mismatch substantially changes our understanding of VLA vulnerability. The same textures become weaker under geometry-consistent evaluation in all 24 tested combinations, while re-optimization strengthens attacks in 23 combinations. Thus, poor transfer of an existing patch does not establish robustness against physical patches. We further show that the deployment assumptions affect multi-view patch optimization and adversarial training. We believe these findings offer new insights into VLA security and provide practical guidance for both attack evaluation and defense.

Q1: Evaluation with occlusion-aware compositing and photometric variations.

Thanks for your valuable suggestion. We have conducted additional experiments with occlusion-aware compositing and photometric variations, as reported in Appendix F (Table A7). Specifically, we extend G to G+, which uses per-frame scene depth to handle occlusion by the robot and objects, and incorporates brightness, gamma, color-channel gain, and patch-color variations. We evaluate EDPA on OpenVLA across all 10 LIBERO-Spatial tasks, with 50 episodes per task for each condition. The initial states and photometric parameters are matched across conditions, and both attacks use the same patch placements.

Here, P/G+ evaluates the original pixel-generated patch under G+, while G+/G+ optimizes and evaluates the patch under G+. Clean uses the same photometric variations. Lower task success indicates a stronger attack.

SettingTask success rate (%)
Clean82.6
P/G+66.4
G+/G+19.8

Under the same G+ evaluation, the patch optimized under G+ achieves a task success rate of 19.8%, compared with 66.4% for the pixel-generated patch, a difference of 46.6 percentage points. These results show that the benefit of accounting for deployment constraints persists when dynamic occlusion and photometric variations are included.

AuthorsOpenReview ↗

Thank you for the valuable review (2/2)

Q2: How do the simulation and real-robot results support the claimed deployment gap?

Thanks for raising this question. We would like to clarify that both printed patches were optimized using data collected from the real-robot setup (Section 4.6), rather than transferred from LIBERO. The simulation results cover ten tasks per suite, whereas the physical experiments evaluate attacks optimized for the block-stacking setup. Thus, the two experiments involve different attack optimization problems, and the absolute success rates should be interpreted in their respective settings.

We also would like to point out that the physical comparison extends beyond the default condition in Table 4. In Table 5, geometry-consistent EDPA reduces task success to 10.0% versus 30.0% for original EDPA under front placement, and to 3.3% versus 23.3% under the left-rear viewpoint. It achieves equal or lower success rates in all tested conditions. These results show a practical benefit from accounting for deployment geometry, particularly in the reported placement and viewpoint variations, even though both methods can produce effective printed patches.

Q3: Why is the 30×30-pixel attack compared against P/G rather than G/G?

Thanks for your valuable suggestion. Table 3 uses P/G because we wanted to examine whether a smaller pixel-space patch remains stronger than directly deploying the original pixel-generated texture. We agree that G/G should also be included to compare attacks optimized for their respective deployment settings. The G/G results are already reported in Table A4, and we provide the complete comparison below. Lower task success indicates a stronger attack.

SettingSpatialObjectGoalLong
Pixel-space, 30×30 pixels22.828.833.018.6
P/G, 12×12 cm52.047.053.227.6
G/G, 12×12 cm17.444.22.09.2

G/G indeed produces stronger attacks on three suites. This result demonstrates the value of addressing the deployment mismatch: on Goal, for example, geometry-consistent optimization reduces success from 53.2% to 2.0%. Our finding about overestimation concerns the same generated texture, as established by the P/P–P/G comparisons in Tables A4–A6. It does not imply that an attack optimized for physical deployment must remain weaker than a pixel-space attack. We will include G/G in Table 3 and clarify this distinction in Section 4.5.

Reviewer_hKsMOpenReview ↗

Reply to authors

Thank you for the clarification and additional experiments. The rebuttal addresses most of my concerns. I will raise my score to a positive rating.

AuthorsOpenReview ↗

Response to reviewer hKsM's reply

Dear Reviewer hKsM,

We sincerely appreciate your decision to raise the rating and your time and effort in reviewing our rebuttal and additional experiments. Your suggestions have been valuable in helping us improve the manuscript. If you have any additional questions, we are glad to further answer them.

Reviewer_ecrTOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines how physical deployment constraints change adversarial patch attacks on VLA models. It separately varies patch generation and evaluation between pixel-space compositing and geometry-consistent projection. Across three models and four LIBERO suites, fixed planar deployment weakens transferred patches, while re-optimization usually strengthens them. The paper also evaluates closed-loop trajectories, printed patches, and adversarial training.

Critical strengths
  1. The paper addresses an important and practically relevant problem.

  2. The experiments are well controlled.

Critical weaknesses
  1. Qualitative comparisons are limited. More rollout images of clean execution, original EDPA, and geometry-consistent EDPA across several tasks would make the differences clearer.

  2. Patch visibility is unclear. How are occlusions by the robot and moving objects handled? The authors need to provide implementation details.

  3. The Appendix reports overall computation times, but does not analyse the additional cost of the proposed modules.

Questions for authors

The main finding is well supported, and the evaluation framework is useful. Please address the two weaknesses above.

4 comments
AuthorsOpenReview ↗

Thank you for the valuable review

Thank you for the constructive review and for recognizing the practical relevance and controlled experiments. We clarify the implementation of patch placement and visibility below.

W1: Limited qualitative comparisons.

Thanks for your valuable suggestion. We have added paired rollout images of π0\pi_0 on LIBERO-Spatial in Appendix G (Figure A2), comparing clean execution, original EDPA under P/P and P/G, and geometry-consistent EDPA under G/G. All four rollouts start from the same initial state, and we show both front and wrist observations at the same timesteps. The images show how pixel-space patches remain fixed in each image, whereas a workspace-fixed patch changes its projected appearance and visibility as the camera moves. In this example, P/P causes task failure, while P/G succeeds, making the deployment mismatch directly visible.

W2: How are occlusions by the robot and moving objects handled?

Thanks for raising this question. We use the spatial information of objects provided by the simulator to construct an occupancy grid and select an unoccupied tabletop region for patch placement (Appendix A.2). The patch remains fixed in world coordinates throughout each rollout and is projected using the current camera parameters. We would like to clarify that the occupancy grid ensures valid initial placement, while the original projection-based compositor does not include per-frame depth masking for occlusion by the robot and moving objects.

To address this issue, we have conducted additional experiments using an extended compositor, G+, which uses per-frame scene depth to handle occlusion by the robot and objects. We also incorporate photometric variations, including brightness, gamma, RGB gains, and patch contrast. Specifically, we evaluate EDPA on OpenVLA across all 10 LIBERO-Spatial tasks, with 50 episodes per task under each condition. The results are shown below. Lower task success indicates a stronger attack.

SettingTask success rate (%)
Clean82.6
P/G+66.4
G+/G+19.8

Here, P/G+ evaluates the original pixel-generated patch under G+, while G+/G+ optimizes and evaluates the patch under G+. Under the same evaluation conditions, geometry-consistent optimization reduces task success from 66.4% to 19.8%, a difference of 46.6 percentage points. These results show that the benefit of geometry-consistent optimization remains when dynamic occlusion and photometric variations are included. We have added the experimental details and results in Appendix F (Table A7).

W3: Additional computational cost of the proposed modules.

Thanks for your valuable suggestion. We will report the runtime of each proposed module in Appendix C. Across three LIBERO-Spatial scenes, occupancy-grid construction takes 30.20–30.56 ms on CPU, and placement selection, including grid construction, takes 30.95–31.51 ms. Projection and compositing take 2.65–2.66 ms for one camera view and 4.83–4.85 ms for two views on an RTX 5090 GPU. During evaluation, placement is selected once at the beginning of each rollout, while projection and compositing are repeated at each timestep to maintain a patch fixed in world coordinates. These modules introduce only a small additional computational cost during evaluation.

AuthorsOpenReview ↗

Hoping for further feedback

Dear Reviewer ecrT,

Thank you again for the time and effort you put into this manuscript. We eagerly await your further feedback on our rebuttal. Have your concerns been addressed? If you still have questions about our paper, we are happy to answer them and further improve our manuscript.

Best, Authors

The added rollout comparisons and runtime measurements address part of my concerns. However, the rebuttal confirms that the original evaluator did not handle occlusion by the robot and moving objects. For a study centered on physical deployment, this is a limitation of the evaluation itself. The G+ results support the benefit of re-optimization for EDPA on OpenVLA across LIBERO-Spatial, but they do not show how the original G-generated patches or defended models perform under the revised visibility model. The impact of the omitted occlusion handling on the broader multi-view and defense results therefore remains unresolved. Figure A2 also clarifies that moving from P/P to P/G replaces the wrist-specific texture with the front-view texture. Without a P/P comparison that uses the same texture in both views, it is unclear how much of the multi-view gap comes from geometric projection. These concerns reduce my confidence in the physical interpretation of the main results, and I revise my rating from 3 to 2.

Thank you for the follow-up and for recognizing the added rollout comparisons and runtime measurements. With approximately three hours remaining before the rebuttal deadline, we are unable to complete the additional evaluations you suggest. We would like to clarify how the existing experiments and the original evaluation design support our conclusions.

1. Occlusion handling and the additional G+ evaluation.

The original G evaluator models a fixed planar patch under camera projection and placement constraints, but does not include dynamic occlusion. To examine whether the benefit of geometry-consistent optimization persists beyond this setting, we added G+, which incorporates per-frame depth-based occlusion by the robot and objects, together with photometric variations.

Under the same G+ evaluation, re-optimization reduces task success from 66.4% to 19.8% for EDPA on OpenVLA/LIBERO-Spatial (Appendix F, Table A7). This provides evidence that the re-optimization benefit persists when dynamic occlusion and photometric variations are included. We agree that evaluating the original G-generated patches and defended models under G+ would additionally quantify the sensitivity of the original results to the visibility model; this comparison is not available in the current results.

2. Physical evidence for attacks and defenses.

Our conclusions about physical attacks and defenses are also supported by direct experiments with printed patches on a real robot, where occlusion, illumination, and camera response arise from the actual physical setup.

Table 4 shows that geometry-consistent EDPA reduces task success from 90.0% to 16.7%. Table 5 further evaluates changes in illumination, patch placement, and camera viewpoint, with geometry-consistent EDPA achieving equal or lower observed task success than original EDPA in every tested condition.

For defense, Table 7 reports 80.0% success for Geometry AT versus 43.3% for Pixel AT in the original scene, and 50.0% versus 40.0% in the unseen scene with additional objects. The patches were re-optimized separately against each defended model, so these are adaptive physical attacks. These experiments directly demonstrate attack effectiveness and defense benefits in the tested physical task and scenes, providing evidence beyond the original simulation evaluator.

3. Shared textures and the purpose of the multi-view comparison.

We would like to clarify that P/P follows the pixel-space attack protocols evaluated in our study [1,2]. In particular, EDPA [2] explicitly applies separate adversarial patches to the primary and wrist cameras in its multi-camera evaluation (Section 4.3). We retain this setting to compare against the original attack protocol.

Our study asks how the effectiveness measured under that protocol changes when the attacker is constrained to one fixed physical patch. Sharing a texture across views is therefore part of the deployment constraint being investigated. Under P/G, we use the front-view texture in both views to enforce this requirement, together with geometric projection.

The multi-view P/P–P/G comparison consequently measures the overall deployment mismatch, including both texture sharing and projection. We agree that a shared-texture P/P control would be useful for separating their individual contributions. We will make the scope of this comparison explicit: the reported gap concerns the transition from the original pixel-space protocol to a fixed planar-patch setting, rather than the effect of projection alone.

[1] Taowen Wang, et al. Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics. ICCV, 2025.

[2] Haochuan Xu, et al. Model-agnostic Adversarial Attack and Defense for Vision-Language-Action Models. arXiv:2510.13237, 2025.

Comments on the paper

1 comment

Dear reviewers,

We thank you again for the valuable and constructive comments. We have provided detailed responses to your comments and conducted additional experiments following your suggestions. We are looking forward to hearing from you about any further feedback.

If you find our responses satisfactory, we hope you might view this as a sufficient reason to reconsider the rating further.

If you still have questions about our paper, we are willing to answer them and improve our manuscript.

Best,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.