We thank Reviewer FjXF for the valuable comments. Specific concerns are also addressed below.

Q1: Robustness Evaluation Coverage

Q1-1: Extreme Occlusion Evaluation

We thank the reviewer for highlighting this important aspect. We explicitly evaluate the robustness of our approach on both synthetic and real-world datasets under severe structural incompleteness and sensor degradation. Visualizations of extreme patterns are provided in Supplementary Sec. C (Line 911).

• On ShapeNet (synthetic), we tested with up to 75% of input points missing across 55 categories, as reported in Tables 2–3 of the main paper (CD-H metric).
• On KITTI (real-world), the input is even sparser (averaging 440 points, or 5.4% of the completion target, with some samples below 10 points), where our model still achieves strong performance.

Q1-2: Noise-Level Diversity

We inject varying noise intensities to improve robustness (see Supplementary Sec. A.2, Lines 816-829). Furthermore, KITTI scans naturally contain a wide range of noise intensities, density, and missing regions caused by sensor and environmental factors, making it an ideal benchmark for evaluating robustness.

Q1-3: Benchmark Coverage

To the best of our knowledge, our approach is among the most thoroughly evaluated, achieving SOTA on four major benchmarks (PCN, ShapeNet, MVP, KITTI), while most prior works report on only two or three. We welcome suggestions and are willing to conduct additional experiments as suggested.

Q2: Cross-dataset Experiment

In addition to KITTI, we evaluated additional cross-dataset experiments with the PCN-pretrained model on Completion3D dataset[1]:

Rebuttal Table A. Cross-dataset Results on Completion3D.

Q3: Explanation of Design Details

Q3-1: Choice of Farthest-Point Sampling (FPS)

We adopt FPS as the default masking strategy in this field, following a well-established standard [2-4], which has shown FPS produces more uniform and globally distributed missing regions than random or grid-based sampling. This characteristic is especially beneficial for our point cloud completion task, enabling the model to better infer unseen parts when facing missing regions.

Q3-2: Choice of Gaussian Noise

We choose Gaussian noise because it is the standard choice for simulating sensor artifacts in point cloud learning [5-6]. Measurement errors in real-world 3D sensors are well-approximated by a Gaussian distribution (per the central limit theorem).

Q3-3: Positional Encoder

The learnable positional encoder is implemented as an MLP that maps each patch center’s spatial coordinate into a high-dimensional embedding, a technique widely used in point cloud representation learning [3] [7].

The positional encoder, together with all other components of the Stochastic Masked Reconstruction module, is trained jointly in an end-to-end manner by minimizing the masked reconstruction loss (see Supplementary Sec. A1, Lines 780–815).

Q3-4: Key Hyperparameters

Please see Supplementary Sections A3–B (Lines 830–890) for details of key hyperparameters and their impact on stability and efficiency. We will move this key discussion into the main text in the final version.

[Dear Reviewer FjXF] We sincerely appreciate the reviewer’s attention to detail and high standards. We value your feedback and look forward to addressing any further questions or concerns.

[1] Topnet: Structural point cloud decoder.

[2] Pointnet++: Deep hierarchical feature learning on point sets in a metric space.

[3] Masked Autoencoders for Point Cloud Self-supervised Learning.

[4] Point transformer.

[5] PointCleanNet: Learning to Denoise and Remove Outliers from Dense Point Clouds.

[6] Point cloud denoising review: from classical to deep learning-based approaches.

[7] Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modeling.

Dear Reviewer FjXF, thank you very much for your constructive suggestions and for raising the rating. We truly value your recognition and support. We are grateful that your concerns have been addressed and would be pleased to discuss any further concerns you may have.

We thank Reviewer PV7J for the recognition and constructive comments.

Q1: Bi-Aux Units Visualization

We sincerely thank the reviewer for this suggestion. We agree that additional visualizations—such as attention maps, feature activations, and region-wise responses—will clarify how Bi-Aux Units focus on specific input regions and improve interpretability. We will include these in the revision and welcome further feedback.

Q2: Computational Cost

In Rebuttal Table A, we report computational cost in terms of model parameters, FLOPs, and ms/sample. As shown in the table, our model is lightweight compared to existing ones, while still achieving the best performance. TTA adds a relatively small amount of overhead to the inference time since most gains are obtained within the first few adaptation steps. The inference time is reported under three adaptation steps. Please note that ms/sample is reported using one Nvidia V100.

Rebuttal Table A. Computational cost.

Q3: Evaluation under Extreme Patterns

We thank the reviewer for this important question. We explicitly evaluate the robustness of our approach on both synthetic and real-world datasets under severe structural incompleteness and sensor degradation. Visualizations of extreme patterns are provided in Supplementary Sec. C (Line 911).

• On ShapeNet (synthetic), we tested with up to 75% of input points missing across 55 categories, as reported in Tables 2–3 of the main paper (CD-H metric).
• On KITTI (real-world), the input is even sparser (averaging 440 points, or 5.4% of the completion target, with some samples below 10 points), where our model still achieves strong performance.

Importantly, TTA with meta-auxiliary learning provides a general framework: the auxiliary task is not restricted to enumerating all possible failure modes. Instead, it can be tailored to different objectives as needed, fundamentally overcoming the limitations of static, pre-defined strategies. This enables dynamic, per-sample adaptation to previously unseen patterns, allowing our model to robustly recover fine details where traditional static training or enumerative augmentation would fail.

Q4: Hyperparameter Choices

We appreciate the reviewer’s concern regarding hyperparameter sensitivity. Please see Supplementary Sec. A4 (Lines 867–881), where we found that the number of gradient steps is the most impactful factor and have provided a thorough analysis. Furthermore, our adaptive $\lambda$-calibration with MAML is designed to alleviate the need for expensive manual hyperparameter tuning by learning suitable hyperparameters during meta-training. As such, the learning rate simply follows the initial task-specific setting.

[Dear Reviewer PV7J] Thank you again for your rigorous and constructive review, which has helped us further improve the quality of our work.

Dear Reviewer PV7J, thank you very much for your positive response. We appreciate your feedback and are grateful that your concerns and questions have been addressed. We are always ready to discuss any further concerns you may have.

We sincerely thank Reviewer SjCL for constructive comments.

Question

Q1: Limitations and Failure Cases

While TTA improves completion in many challenging cases as compared with existing approaches, it still has inherent limitations under certain conditions. Specifically, our sample-specific TTA is unable to “hallucinate” plausible fine-grained geometry beyond what is already encoded in the model’s priors. Therefore, when objects have extremely complicated geometric structures and the input point cloud provides minimal information, the adaptation process may still overfit to the limited input, resulting in blurry completions, especially for categories like birdhouses.

For future work, we intend to explore TTA for post-training quantized (PTQ) generative models [1], aiming to enhance completion accuracy in challenging scenarios while maintaining the diversity of generated results. We will clarify this limitation and discuss these potential directions in the revised manuscript.

Q2: Computational Cost

In Rebuttal Table A, we report computational cost in terms of model parameters, FLOPs, and ms/sample. As shown in the table, our model is lightweight compared to existing ones, while still achieving the best performance. TTA adds a relatively small amount of overhead to the inference time since most gains are obtained within the first few adaptation steps. The inference time is reported under three adaptation steps. Please note that ms/sample is reported using one Nvidia V100.

Rebuttal Table A. Computational cost.

Weakness

W1: Cherry-picking Concern on Figure 1

Figure 1 shows a typical example from the PCN test set. The improvement can also be observed in the results shown in Fig. 4 and Fig. 5 of the main paper. They illustrate the primary benefit of our sample-specific adaptation. We believe the underlying reason for getting the improvement is that the inner-loop optimization in TTA with MAML updates model parameters along the gradient of the self-supervised loss, gradually steering the model output toward globally and semantically plausible completions (e.g., an airplane with a reasonable tail) for each test sample. The outer-loop meta-optimization, driven by the completion task, ensures that the auxiliary branches provide effective adaptation signals. Please refer to Alg. 2 in Supplementary A.3.2 (Lines 846–866).

This represents a paradigm shift from traditional static, one-size-fits-all inference. As a result, improvements are consistent rather than cherry-picked. This is also validated in Tables 1–5 (across all benchmark datasets) of the original submission. We will include more compelling examples in the final version to further demonstrate its generalization ability.

W2: Specifics of the Claim (Conventional Generic Completion vs. Our Sample-specific Completion with TTA)

We appreciate the reviewer’s insightful suggestion and agree that it is important to clarify this distinction. Basically, in point cloud completion, there are two main types of models: deterministic models and probabilistic models (e.g., diffusion models). In this work, we study deterministic models that usually yield averaged (generic) predictions in existing works due to mean regression. That motivates us to develop a sample-specific adaptation strategy to predict the geometry details for each test sample.

W3: Concern about Application and Significance

As shown in Rebuttal Table A, our method consistently achieves improved accuracy while maintaining a reasonable computational cost. Moreover, for real-world deployment, where test and training data often differ significantly due to changes in sensors, objects, and other factors, domain shift can lead to performance degradation when deploying a fixed, pre-trained model. On the contrary, our TTA approach enables continuous personalization to new environments and samples by dynamically adapting to each test instance. This claim is also validated in our cross-dataset evaluation as shown in Table 5 of the main paper and Fig. 7 in the Supplementary (Line 911)

[1] Post-training quantization on diffusion models.

[Dear Reviewer SjCL] We value your feedback and look forward to discussing any further concerns you may have.

Dear Reviewer SjCL, thank you very much for your constructive comments and positive rating. We appreciate your acknowledgment that most of your concerns have been addressed. If you have any further questions or suggestions, we would be grateful for the opportunity to discuss them.

We sincerely thank Reviewer sY8X for the valuable suggestions.

Q1: Computational Cost

In Rebuttal Table A, we report computational cost in terms of model parameters, FLOPs, and ms/sample. As shown in the table, our model is lightweight compared to existing ones, while still achieving the best performance. TTA adds a relatively small amount of overhead to the inference time since most gains are obtained within the first few adaptation steps. The inference time is reported under three adaptation steps. Please note that ms/sample is reported using one Nvidia V100.

Rebuttal Table A. Computational cost.

Q2: Evaluating TTA on General Backbones

We agree that such experiments are essential for demonstrating the broader applicability of our method. As shown in Rebuttal Table B, we have validated the plug-and-play effectiveness of our TTA on PoinTr and SnowflakeNet, to demonstrate practical generalization.

Rebuttal Table B. TTA on general backbones.

Q3: Figure Clarity and Workflow

We will revise Figures 2 and 3 accordingly in the final version. Please also refer to Algorithm 2 in Supplementary A.3.2 for the detailed Meta-Auxiliary Training process. We will release the source code upon acceptance.

Q4: Expand Related Work

We agree that adding recent TTA methods for 3D point clouds, such as 3DD-TTA [1] and CloudFixer [2], will strengthen our related work section.

We will include these approaches, which primarily focus on input-level adaptation (e.g., modifying the test data distribution via diffusion or denoising). In contrast, our model adopts a model-level adaptation strategy, dynamically updating model parameters for each test sample. We will clarify these distinctions and expand our discussion in the revised manuscript. We welcome any further recommendations from the reviewer.

[1] Test-time adaptation of 3d point clouds via denoising diffusion models.

[2] Cloudfixer: Test-time adaptation for 3d point clouds via diffusion-guided geometric transformation.

[Dear Reviewer sY8X] We value your feedback and look forward to discussing any further concerns you may have.

Dear Reviewer sY8X, we hope that our experiments and analyses have addressed your concerns. Thank you very much for your constructive suggestions. If you have any additional concerns, we would sincerely appreciate the opportunity to discuss them.