acceptodds
Under review as a conference paper at ICLR 2027

Beyond Static Locality: Adaptive State Space Modeling for 3D Point Clouds

Abstract

State space models (SSMs) have recently been explored as an efficient framework for point cloud representation learning. However, existing point-cloud SSMs typically treat locality as a fixed architectural assumption, limiting their ability to adapt local modeling to diverse geometric structures and hierarchical feature representations. We propose AdaLoMamba3D, a locality-adaptive state-space framework that models locality as an adaptive property determined by learned representations. First, local interaction ranges are adapted according to representation hierarchy, enabling different representation levels to employ distinct neighborhood ranges. Center-neighbor relations are further normalized based on local feature statistics in a channel-wise manner, enabling different regions to obtain adaptive normalization references and reducing the bias introduced by uniform normalization. The resulting representations are propagated through bidirectional state-space modeling, where information exchange between complementary branches is conditioned on current token states. Together, these designs form a unified framework for adaptive locality. Experimental results show competitive and consistent performance across classification, few-shot recognition, and part segmentation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Point Cloud, State Space Models, Mamba, Adaptive Locality, 3D Vision

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_66XaOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors present AdaLoMamba3D, a locality-adaptive state-space framework for 3D point cloud representation learning. By introducing adaptive mechanisms into local feature aggregation and state-space interaction, the approach enables point cloud models to better accommodate diverse geometric structures and feature representations while maintaining the efficiency of the Mamba-based backbone.

Critical strengths

The paper inspires readers to propose locality-adaptive state-space frameworks for 3D point cloud learning.

Critical weaknesses
  1. In the introduction, the authors claim that AdaLoMamba3D possesses adaptive locality for representation hierarchy, local feature distributions, and token states, which is highly appealing. However, PLE only sets different fixed neighborhood sizes across different layers, identical to those in other methods that the authors criticize, making it difficult to support the authors’ claim of flexibility in capturing diverse geometric patterns.

  2. In SAF, the channel-wise modulation is predicted only for the reverse scan. What is the rationale behind this technical choice? Isn’t the forward scan on the same level as the reverse scan?

  3. The experiment lacks quantitative comparison with recent works. Furthermore, to demonstrate the effectiveness of AdaLoMamba3D, additional experiments on more challenging point cloud tasks are necessary, such as point cloud semantic segmentation on S3DIS.

  4. There have been many outstanding works on point cloud learning recently, such as [2] and [3], but the authors do not review these efforts. Therefore, as a submission to ICLR 2027, Section 2 is clearly inadequate. Furthermore, the last sentence in Section 2.2 is redundant with the previous sentence; the statement “Unlike prior methods, our approach focuses on the adaptivity of locality itself” is clear on its own.

  5. In the third paragraph of the introduction, the author should draw a definitive conclusion based on the above, i.e., that locality should be adapted according to geometry, representation hierarchy, and feature states; then, in the fourth paragraph, AdaLoMamba3D is introduced.

  6. In the second paragraph of the introduction, the authors present a misconception regarding Mamba3D [1]. Mamba3D does not include a bidirectional SSM for enhancing global modeling. The method focuses solely on locality perception, and its bidirectional SSM serves as a linear-complexity operator for channel modeling, as shown in Figure 2 of Mamba3D.

[1] X. Han, Y. Tang, Z. X. Wang, and X. Z. Li, ”Mamba3D: Enhancing Local Features for 3D Point Cloud Analysis via State Space Model,” in Proc. ACM International Conference on Multimedia (MM), Melbourne, VIC, Australia, Nov. 2024, pp. 4995--5004.

[2] D. K. Liang, T. R. Feng, X. Zhou, Y. M. Zhang, Z. K. Zou, and X. Bai, ”Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), vol. 47, no. 12, pp. 10949--10966, Dec. 2025.

[3] K. L. Qu, P. Gao, Q. Dai, and Y. H. Sun, ”Point-Focused Attention Meets Context-Scan State Space: Robust Biological Visual Perception for Point Cloud Representation,” in Proc. International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil, Apr. 2026, pp. 1--17.

Questions for authors

Please see the Critical Weaknesses section.

6 comments
AuthorsOpenReview ↗

Response to Weakness 1: Depth-Dependent Scheduling and Input-Dependent Adaptation

We appreciate the reviewer's insightful comments on the adaptive locality modeling in our framework. We acknowledge that the original manuscript did not sufficiently distinguish the adaptation mechanisms of the three components, particularly giving the impression that PLE dynamically selects neighborhood sizes based on input geometry. We apologize for this lack of clarity and provide the following clarification.

1. Clarification of PLE and Framework-Level Adaptation

PLE employs a predefined depth-dependent neighborhood schedule [4, 8, 12]. Its primary objective is to align the spatial range of local feature interactions with the progressive development of feature representations across network depths. In earlier layers, smaller neighborhoods allow the model to focus on fine-grained geometric information near each patch center, whereas progressively larger neighborhoods provide broader spatial context for higher-level feature representations.

Furthermore, PLE maintains the same patch centers across network depths, ensuring consistent spatial references for local feature aggregation. When patch centers and distance metrics remain unchanged, progressively increasing KNN neighborhood sizes produces nested neighborhood sets, allowing the model to incorporate more distant neighbors while retaining the coverage of nearby ones. Thus, PLE introduces a depth-dependent structural inductive bias by aligning neighborhood expansion with representation hierarchy.

2. Comparison with Geometry-Based Adaptive Neighborhood Selection

To further evaluate progressive neighborhood scheduling, we additionally tested four adaptive neighborhood selection strategies based on local geometric statistics. Experiments were conducted on ScanObjectNN OBJ_ONLY using three independent random seeds (0, 1, and 42). The results are presented below:

Neighborhood Selection StrategyOA (%)
Dimensionality Agreement94.15±0.17
Eigenentropy93.69±0.26
Dimensionality Entropy94.11±0.35
Surface-Variation Peak93.92±0.20
PLE [4, 8, 12]94.61±0.26

PLE achieves the highest mean accuracy (94.61±0.26%) among the evaluated neighborhood selection strategies, outperforming the best-performing geometry-based adaptive strategy by 0.46 percentage points. These results provide further empirical support for the effectiveness of depth-dependent progressive scheduling within our current architecture.

3. Design Rationale

The key distinction lies in how neighborhood selection accounts for representation hierarchy. The evaluated geometry-based adaptive strategies primarily determine neighborhood sizes from local geometric statistics of the input point cloud, without explicitly considering the changing feature aggregation requirements across network depths. In contrast, PLE coordinates neighborhood expansion with progressive feature abstraction, allowing shallow layers to focus on nearby geometric structures while deeper layers integrate broader spatial context. Combined with the nested structure of KNN neighborhoods and consistent patch centers, this design aligns local information aggregation with representation depth.

Meanwhile, geometry-based adaptive neighborhood selection and depth-dependent scheduling are complementary rather than mutually exclusive. In future work, we intend to investigate dynamic neighborhood selection mechanisms that jointly consider local geometry, representation hierarchy, and task-specific requirements to further improve the modeling of diverse geometric patterns.

Following the reviewer's suggestion, we have refined the relevant descriptions in the Abstract, Introduction, and Method sections to clarify the respective roles of PLE, LSA, and SAF. We have also added three-seed comparisons between PLE and four geometry-based adaptive strategies in Table 9 of Section 4.3.

AuthorsOpenReview ↗

Response to Weakness 2: Rationale for SAF Gate Placement

We are grateful to the reviewer for the constructive comments regarding the gate placement in SAF. We further clarify the rationale for adopting a single-branch dynamic gating mechanism and specifically applying it to C-SSM, supported by additional ablation experiments.

1. Rationale for Single-Branch Dynamic Gating and Gate Placement

The objective of SAF is to dynamically adjust the relative contributions of the two state-space branches based on the current input features, which can be achieved by modulating the fusion weight of a single branch. Accordingly, we retain L+SSM as the direct fusion pathway and generate channel-wise gating weights from the current token representations to dynamically modulate the complementary information provided by C-SSM.

When the gating weights equal 1, the outputs of the two branches are directly summed. When the gating weights are below or above 1, the relative contribution of C-SSM with respect to L+SSM decreases or increases accordingly. Therefore, even without applying an additional gate to L+SSM, the model can achieve input-dependent modulation of the relative contributions of the two branches. Compared with independently gating both branches, this design avoids additional gating parameters while preserving the original additive fusion as a special case. This choice does not imply any difference in the importance of the two branches for feature modeling. Rather, it enables dynamic feature fusion through a simpler architecture by retaining the direct output of L+SSM and selectively integrating the complementary information from C-SSM.

2. Ablation Study of Dynamic Gate Placement

To further evaluate the effectiveness of this gate placement, we conducted additional ablation experiments with different gating configurations. On ScanObjectNN OBJ_ONLY, we evaluated three strategies: applying dynamic gating only to L+SSM, applying independent gates to both branches, and applying dynamic gating only to C-SSM (SAF).

All experiments used the same pretrained initialization and identical downstream experimental settings. We report the mean overall accuracy (OA) and standard deviation across three independent random seeds (0, 1, and 42):

Dynamic Gate PlacementOA (%)
L+SSM Only94.15±0.17
Both Branches (Independent Gates)94.49±0.29
C-SSM Only (SAF)94.61±0.26

The results show that SAF, which applies dynamic gating only to C-SSM, achieves the highest mean accuracy (94.61±0.26%) among all evaluated configurations. These findings further support our choice of gate placement, demonstrating that single-branch dynamic gating enables effective input-dependent feature fusion while maintaining a simple architecture.

Following the reviewer's suggestion, we have further clarified the rationale for SAF's single-branch gating mechanism and gate placement in the revised manuscript and included the additional three-seed ablation results in Table 10 (Dynamic Gate Placement) of Section 4.3.

AuthorsOpenReview ↗

Response to Weakness 3: Recent Method Comparisons and S3DIS Semantic Segmentation

We are grateful to the reviewer for the constructive suggestions regarding the completeness of our experimental evaluation. We agree that comparisons with recent methods and validation on more challenging downstream tasks are important for comprehensively evaluating AdaLoMamba3D. In response, we have expanded the quantitative comparisons with recent studies and conducted additional semantic segmentation experiments on S3DIS.

1. Quantitative Comparisons with Recent Methods

We have expanded Tables 1–3 of the revised manuscript to include representative recent methods, such as RI-MAE (AAAI 2025), Point-MaDi (NeurIPS 2025), PointGST-M (TPAMI 2025), PointMSMamba (TVC 2025), and PointMamba++ (IET CV 2026). Together with existing comparisons involving Point-PQAE and StruMamba3D, these additions provide a more comprehensive evaluation of classification performance on ScanObjectNN and ModelNet40, as well as part segmentation performance on ShapeNetPart. ModelNet40 classification results are reported without voting.

The following table summarizes selected published results of recent methods on these benchmarks:

MethodPB_T50_RS (%)ModelNet40 (%)ShapeNetPart Ins. mIoU (%)
RI-MAE (AAAI 2025)85.80–84.3
Point-MaDi (NeurIPS 2025)89.5293.886.3
PointGST-M (TPAMI 2025)85.2993.585.8
PointMSMamba (TVC 2025)88.7293.686.4
PointMamba++ (IET CV 2026)87.9693.786.1
AdaLoMamba3D (Ours)94.5595.786.5

2. Additional S3DIS Semantic Segmentation Experiments

To further evaluate the effectiveness of AdaLoMamba3D on more challenging scene-level point cloud understanding tasks, we conducted additional semantic segmentation experiments on S3DIS Area 5. We evaluated AdaLoMamba3D and compared its performance with representative existing methods. The results are presented below:

MethodmAcc (%)mIoU (%)
PointNet++ (NeurIPS 2017)67.153.5
Point-BERT (CVPR 2022)70.360.8
Point-MAE (ECCV 2022)69.960.8
PointGPT-L (NeurIPS 2023)70.662.2
Point-PQAE (ICCV 2025)70.661.4
PCM (AAAI 2025)–63.4
Point-MaDi (NeurIPS 2025)71.061.2
PointGST (TPAMI 2025)68.458.6
RI-MAE (AAAI 2025)68.460.3
AdaLoMamba3D (Ours)71.063.4

AdaLoMamba3D achieves 71.0% mAcc and 63.4% mIoU on S3DIS Area 5, matching the highest reported values among the compared methods on both metrics. These results provide further evidence that our approach is effective not only for object-level classification and part segmentation but also for more challenging scene-level semantic segmentation, offering additional support for its point cloud representation learning capabilities.

Following the reviewer's suggestion, we have expanded the quantitative comparisons with recent methods in Tables 1–3 and added the S3DIS Area 5 semantic segmentation results in Table 4 of the revised manuscript, providing a more comprehensive evaluation across different downstream tasks.

Reviewer_5XDPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AdaLoMamba3D, a locality-adaptive state-space framework for 3D point cloud representation learning built upon Mamba3D. The authors argue that existing point-cloud Mamba architectures rely on fixed locality modeling mechanisms that may not adequately capture variations in geometric structures and feature representations.

The proposed framework introduces three components: (1) Progressive Locality Expansion (PLE), which increases neighborhood sizes across network stages; (2) Local Statistical Adaptation (LSA), which normalizes center-neighbor feature differences using neighborhood-specific, channel-wise statistics; and (3) State-Adaptive Mamba Fusion (SAF), which uses token-dependent gating to modulate the reverse Mamba branch.

The method is evaluated on object classification, few-shot classification, and part segmentation. The reported results demonstrate competitive performance, particularly on ScanObjectNN, while ablation studies examine the contributions of the proposed components.

Critical strengths
  • The proposed method achieves competitive performance across multiple point cloud benchmarks, including classification, few-shot recognition, and part segmentation.

  • The method achieves 94.55% accuracy on ScanObjectNN PB_T50_RS with pre-training, outperforming several existing point-cloud Mamba methods.

  • The paper provides component-wise ablation studies for PLE, LSA, and SAF, demonstrating the individual contributions of each component.

  • Combining all three components improves accuracy from 93.19% to 94.55% on PB_T50_RS, supporting their complementary effectiveness.

  • Additional ablation studies investigate different neighborhood schedules and bidirectional fusion strategies, providing further evidence for the effectiveness of the proposed design.

Critical weaknesses
  • Limited methodological novelty: The proposed components mainly involve relatively straightforward modifications, including progressive neighborhood expansion, local statistical normalization, and sigmoid-based feature gating.

  • Insufficient justification of adaptive locality: Although PLE progressively increases neighborhood sizes across network stages, the values (4, 8, and 12) are predefined rather than dynamically determined by input geometry or learned representations. This raises questions about the extent of adaptivity introduced by PLE.

  • Limited distinction from existing techniques: The paper does not sufficiently clarify the novelty of PLE compared with hierarchical neighborhood modeling approaches or LSA compared with Mamba3D's LNP module.

  • Inconsistent baseline results: The baseline accuracy in Table 4 (93.19%) differs from the Mamba3D result in Table 1 (92.05%) on PB_T50_RS, without a clear explanation of the differences in experimental settings.

  • Underused page limit: Although the paper uses only 8 of the 9 permitted pages, the remaining space could have been used to better justify the proposed design choices and provide deeper analysis of the results.

Questions for authors
  • 1. Adaptive locality: Although Table 5 demonstrates the benefit of progressively increasing neighborhood sizes (4, 8, and 12) over fixed schedules, could the authors clarify the methodological novelty of PLE compared with existing hierarchical neighborhood modeling approaches, such as PointNet++? What are its conceptual and experimental advantages over these approaches?

  • 2. LSA ablation: When LSA is disabled in Table 4, is the original Local Norm Pooling (LNP) module from Mamba3D retained, or is local normalization removed entirely? Could the authors clarify this to better establish the specific contribution of LSA over LNP?

  • 3. Baseline consistency: The baseline accuracy in Table 4 (93.19%) differs from the reported Mamba3D accuracy in Table 1 (92.05%) on PB_T50_RS. Could the authors explain this discrepancy and clarify whether both results were obtained under identical experimental settings?

  • 4. SAF ablation: In Table 6, the authors evaluate only a fixed gating coefficient of 0.5. Could the authors provide additional ablation studies using different fixed coefficients (e.g., 0.25 and 0.75) to better demonstrate the effectiveness of SAF over fixed fusion strategies?

  • 5. Related work and comparisons: Could the authors discuss the closely related works Enhancing Point Cloud Analysis with Multi-Scale State Space Modeling and Edge Graph Augmentation (2025) and PointMamba++: Rethinking Ordering and Convolution Strategy of State Space Model for Point Cloud Analysis (2026), clarify the methodological differences from the proposed approach, and, if possible, include experimental comparisons with these methods?

7 comments
AuthorsOpenReview ↗

Response to Weakness 1: Methodological Novelty and Component Design

We thank the reviewer for the careful assessment of our methodological design. Although progressive neighborhood expansion, local normalization, and gating are simple operations, our designs address the irregular geometry of point clouds, depth-dependent spatial context requirements, heterogeneous local statistics, and varying token representations. PLE, LSA, and SAF target neighborhood organization, feature calibration, and state-space branch fusion, respectively.

1. PLE: Neighborhood Organization Aligned with Hierarchical Feature Representations

As representations evolve from fine-grained local geometry toward higher-level context, different depths require different neighborhood ranges. PLE therefore uses a progressive schedule [4, 8, 12] with fixed patch centers, allowing shallow layers to focus on nearby geometry and deeper layers to aggregate broader spatial context.

When patch centers and distance metrics remain unchanged, progressively increasing KNN neighborhood sizes produces nested neighborhood sets, where larger neighborhoods contain the nearest neighbors included in smaller ones. This incorporates more distant neighbors while preserving token spatial correspondence without extra downsampling.

2. LSA: Feature Scale Calibration Based on Local Statistical Variations

Local geometry and feature response magnitudes vary across regions and channels. To reduce their influence on aggregation, LSA computes local RMS statistics from center–neighbor feature differences and performs normalization independently for each neighborhood and feature channel. This adapts feature scales to current local statistics.

Since the normalization scale is determined by feature differences within each neighborhood and channel, the mean squared normalized feature difference is bounded above by one. Furthermore, when the local feature energy is sufficiently large relative to the numerical stability term, the normalized features remain approximately invariant to uniform scaling of feature magnitudes. Consequently, LSA can help reduce the influence of scale variations across neighborhoods and channels on subsequent Softmax-based weighted aggregation, allowing local feature aggregation to better account for relative feature variations.

3. SAF: Input-Dependent Fusion of Complementary State-Space Branches

L+SSM performs state-space modeling along the original token sequence, whereas C-SSM provides an alternative feature modeling pathway through channel reversal. The two branches offer complementary representations, but their relative contributions may vary across different input tokens. Consequently, a fixed fusion weight may not adequately accommodate such variations.

To address this issue, SAF preserves the direct output of L+SSM while generating channel-wise gating weights from the current token representations to dynamically modulate the contribution of C-SSM. The gating function employs a sigmoid scaled by a factor of two, constraining the modulation coefficients to (0, 2) and allowing bounded adjustments relative to the original equal-weight fusion. In addition, the gating network is initialized with zero weights and biases, ensuring that the initial gating coefficients equal one. This preserves the original additive fusion behavior at initialization while allowing input-dependent modulation weights to be learned during training. In this way, SAF enables dynamic integration of complementary branch information without modifying the underlying state-space modeling pathways.

4. Multi-Seed Component Ablation Experiments

To further evaluate the effectiveness of these designs, we conducted additional multi-seed component ablation experiments on ScanObjectNN OBJ_ONLY. All configurations used the same pretrained initialization and identical downstream training and evaluation settings. We report the mean overall accuracy (OA) and standard deviation across three independent random seeds (0, 1, and 42):

ConfigurationOA (%)
Baseline (w/o PLE, LSA, SAF)93.31±0.38
w/o PLE94.29±0.19
w/o LSA94.21±0.36
w/o SAF94.06±0.22
Full Model (AdaLoMamba3D)94.61±0.26

The complete model achieves 94.61±0.26% overall accuracy. Removing PLE, LSA, and SAF individually decreases mean accuracy by 0.32, 0.40, and 0.55 percentage points, respectively. These results provide additional empirical support for the contributions and complementary effects of the three components in point cloud feature modeling.

In response to the reviewer's comments, we have further clarified the design motivations and technical mechanisms of each component in the Method section and incorporated the additional multi-seed ablation results into Table 5 (Architecture Ablation) in Section 4.3 of the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness 2: Depth-Dependent Neighborhood Scheduling in PLE

We are grateful to the reviewer for the constructive feedback on the adaptivity of PLE. We acknowledge that the original manuscript did not sufficiently distinguish predefined depth-dependent scheduling from input-dependent adaptation, and we apologize for this lack of clarity.

1. Clarification of PLE Adaptivity

PLE employs a predefined progressive neighborhood schedule [4, 8, 12] to adjust local aggregation ranges according to network depth. Our framework combines this depth-dependent structural prior with the input-dependent adaptation mechanisms of LSA and SAF.

2. Comparison with Geometry-Based Adaptive Neighborhood Selection

To further evaluate the effectiveness of progressive neighborhood scheduling, we additionally implemented and tested four adaptive neighborhood selection strategies based on local geometric statistics. Experiments were conducted on ScanObjectNN OBJ_ONLY using three random seeds (0, 1, and 42). The results are presented below:

Neighborhood Selection StrategyOA (%)
Dimensionality Agreement94.15±0.17
Eigenentropy93.69±0.26
Dimensionality Entropy94.11±0.35
Surface-Variation Peak93.92±0.20
PLE (Progressive [4, 8, 12])94.61±0.26

PLE achieves an overall accuracy (OA) of 94.61±0.26%, with a mean accuracy 0.46 percentage points higher than that of the best-performing geometry-based adaptive strategy. These results provide further empirical support for the effectiveness of depth-dependent progressive neighborhood scheduling within our current architecture.

3. Design Rationale

A key distinction lies in how neighborhood scales are aligned with hierarchical feature representations. As network depth increases, feature representations progressively evolve from fine-grained local geometry toward higher-level contextual information, leading to different requirements for neighborhood coverage across representation levels. PLE aligns neighborhood expansion with this feature abstraction process, allowing earlier layers to focus on nearby geometric information while deeper layers progressively incorporate broader spatial context.

Furthermore, when patch centers and spatial distance metrics remain fixed, progressively increasing KNN neighborhood sizes produces nested neighborhood sets: larger neighborhoods contain the nearest neighbors included in smaller ones. This allows spatial coverage to expand progressively across network depths while retaining nearby neighbors. In contrast, the evaluated geometry-based adaptive strategies primarily determine neighborhood sizes from local geometric statistics of the input point cloud, without explicitly accounting for the changing requirements of feature representations across network depths. Thus, PLE explicitly aligns neighborhood expansion with representation hierarchy, a design distinction supported by its higher mean accuracy in these experiments.

Meanwhile, geometry-based adaptive neighborhood selection and depth-dependent scheduling are not mutually exclusive. In future work, we intend to explore dynamic neighborhood selection mechanisms that jointly consider local geometric structures, representation hierarchy, and task-specific requirements to better accommodate the diverse geometric patterns of point clouds.

Following the reviewer's suggestion, we have refined the relevant descriptions in the Abstract, Introduction, and Method sections and added comparisons between PLE and four geometry-based adaptive strategies in Table 9 of Section 4.3.

AuthorsOpenReview ↗

Response to Weakness 3: Methodological Distinctions of PLE and LSA

We thank the reviewer for the constructive suggestion to better distinguish our approach from existing methods. We acknowledge that the original manuscript did not sufficiently clarify the differences between PLE and existing hierarchical neighborhood modeling approaches, or between LSA and the original Local Norm Pooling (LNP), which may have made our specific contributions less clear.

1. Distinction Between PLE and Existing Hierarchical Neighborhood Modeling

Hierarchical methods such as PointNet++ typically construct point cloud representations at different spatial resolutions through progressive sampling, neighborhood grouping, and feature aggregation. In these approaches, the expansion of the receptive field is generally accompanied by a reduction in the number of sampled points and changes in spatial resolution. In contrast, PLE progressively adjusts local neighborhood sizes according to representation depth while preserving patch centers and spatial correspondences between tokens, without requiring additional downsampling or reconstruction of patch centers.

The key distinction lies in decoupling neighborhood size adjustment from changes in token spatial resolution while aligning neighborhood scheduling with the feature representation requirements at different network depths. On the one hand, maintaining fixed patch centers preserves consistent spatial correspondences across network layers without reorganizing tokens through additional sampling operations. On the other hand, progressive neighborhood scheduling enables a deliberate allocation of modeling capacity between fine-grained local geometry at shallower layers and broader spatial context at deeper layers. Therefore, PLE does not merely increase neighborhood sizes; instead, it introduces a depth-dependent neighborhood scheduling mechanism that preserves token structure while aligning local aggregation ranges with representation depth.

To further evaluate the effectiveness of progressive neighborhood scheduling, we compared progressive, reversed, and fixed neighborhood schedules on ScanObjectNN OBJ_ONLY using three random seeds (0, 1, and 42). The results are presented below:

Neighborhood ScheduleOA (%)
[4, 4, 4]94.29±0.19
[8, 8, 8]94.43±0.48
[12, 12, 12]94.03±0.60
[12, 8, 4]94.35±0.35
PLE94.61±0.26

The progressive schedule achieves 94.61±0.26% overall accuracy (OA), with higher mean accuracy than all three evaluated fixed neighborhood configurations and the reversed schedule. These results provide further empirical support for the effectiveness of depth-dependent neighborhood scheduling within our current architecture.

2. Technical Improvements of LSA over LNP

The original LNP uses a shared scalar standard deviation for feature normalization, which may not adequately account for variations in feature magnitudes across different local neighborhoods and channels. To address this limitation, LSA introduces two specific improvements in statistical normalization. First, it computes local RMS statistics from center–neighbor feature differences to characterize relative feature variations within each neighborhood. Second, it estimates a separate normalization scale for each feature channel within each neighborhood, enabling neighborhood-specific, channel-wise statistical calibration. Mathematically, the proposed normalization ensures that the mean squared normalized feature difference within each neighborhood and channel does not exceed one and approaches one when the local feature energy is substantially greater than the numerical stability term. This property helps reduce scale differences across neighborhoods and channels, potentially mitigating their influence on subsequent Softmax-based weighted aggregation.

To further evaluate the improvements introduced by LSA over the original LNP, we conducted a two-factor ablation study comparing normalization and aggregation strategies under identical experimental conditions. The results are summarized below:

Normalization / AggregationMean OA (%)Softmax OA (%)
Shared Scalar Std93.69±0.2694.21±0.36 (Original LNP)
Local Channel-wise RMS94.15±0.4594.61±0.26 (LSA)

Compared with the original shared scalar standard deviation normalization, the proposed local channel-wise RMS normalization improves mean accuracy by 0.46 and 0.40 percentage points under uniform averaging and Softmax-based weighted aggregation, respectively. These results show that neighborhood-specific, channel-wise statistical calibration achieves higher mean accuracy under both aggregation strategies, providing further empirical support for the technical improvements introduced by LSA.

We have clarified these methodological distinctions in the Method section and added the corresponding multi-seed comparisons in Tables 7 and 8 of Section 4.3.

Reviewer_nY4BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

AdaLoMamba3D extends Mamba3D with progressive neighborhood expansion (PLE), local statistical adaptation (LSA), and token-conditioned bidirectional fusion (SAF). The method reports improvements on point-cloud classification, few-shot recognition, and part segmentation, with component ablations on ScanObjectNN.

Critical strengths
  1. The modifications introduce modest reported overhead: 17.1M parameters and 4.0G FLOPs, compared with 16.9M and 3.9G for Mamba3D.
  2. The method reports consistent improvements over Mamba3D across classification, few-shot recognition, and part segmentation. Table 4 evaluates all combinations of the three components, providing useful evidence of their complementary contributions.
Critical weaknesses
  1. The baseline configuration is insufficiently clarified, affecting attribution of the gains. The component-free baseline in Table 4 achieves 93.19% on PB_T50_RS, whereas Mamba3D achieves 92.05% in Table 1. This may reflect a stronger reproduction or different training settings, but the distinction is not explained. Consequently, the reported 2.50-point improvement over Mamba3D cannot be attributed entirely to the proposed components; the controlled ablation shows a 1.36-point gain.

  2. The claimed adaptivity and specific design contributions are not sufficiently established. PLE uses a predetermined [4,8,12] neighborhood schedule rather than learning or predicting neighborhood sizes from representations. LSA combines neighborhood-wise channel normalization with softmax-weighted aggregation, but their contributions are not isolated. SAF is evaluated only against summation and a fixed reverse-branch coefficient of 0.5, leaving unclear whether token-dependent gating is necessary compared with simpler learnable fusion. These limitations weaken the broader claim of advancing adaptive locality beyond the particular architectural combination.

Questions for authors
  1. What explains the 93.19% versus 92.05% baseline difference? Please specify any differences in pre-training, augmentation, optimization, and evaluation, and identify the matched baseline.
  2. Can you separately ablate LSA’s local normalization and weighted aggregation, and compare SAF with learned input-independent scalar/channel weights or a parameter-matched fusion alternative?
  3. Does the progressive schedule outperform a reversed schedule under a comparable neighborhood budget? Are the smaller ablation gains consistent across independent runs?
5 comments
AuthorsOpenReview ↗

Response to Weakness 1: Baseline Configuration and Performance Attribution

We thank the reviewer for the careful attention to the baseline configurations and performance attribution. The primary difference between the 92.05% accuracy in Table 1 and the 93.19% accuracy in the original Table 4 (renumbered as Table 5 in the revised manuscript following the addition of S3DIS semantic segmentation results) lies in the pretraining architecture and the source of pretrained weights.

Specifically, the 92.05% accuracy is the ScanObjectNN PB_T50_RS classification result reported in the original Mamba3D paper, obtained using the original Mamba3D architecture for both pretraining and downstream fine-tuning. In contrast, the 93.19% accuracy was obtained from our component ablation experiments under shared pretraining initialization. We first pretrained the complete AdaLoMamba3D architecture and subsequently loaded the resulting pretrained weights for downstream fine-tuning, with PLE, LSA, and SAF disabled. Therefore, the 93.19% result represents an ablation baseline for evaluating the contributions of the proposed components, rather than an independent reproduction of the original Mamba3D architecture throughout pretraining and fine-tuning.

Regarding downstream training and evaluation, our data augmentation, optimizer, learning rate schedule, training duration, and evaluation protocol are consistent with those of the original Mamba3D. Further experimental details are provided in our response to Question 1.

We appreciate the reviewer for highlighting the need for clarification. We have revised the manuscript to specify the source of pretrained weights and the ablation baseline configuration, making the experimental conditions and purposes of the two comparisons clearer.

AuthorsOpenReview ↗

Response to Weakness 2: Adaptivity and Component Contributions

We appreciate the reviewer's attention to the adaptivity of our framework and the contributions of its individual components. In response, we have clarified the scope of adaptivity in PLE and conducted additional controlled experiments to evaluate the individual contributions of LSA and SAF.

1. PLE: Depth-Dependent Neighborhood Scheduling

PLE employs a predefined depth-dependent neighborhood schedule [4, 8, 12] to adjust local aggregation ranges according to representation depth. To further evaluate this design, we compared PLE with four adaptive neighborhood selection strategies based on local geometric statistics on ScanObjectNN OBJ_ONLY using three random seeds (0, 1, and 42):

Neighborhood Selection StrategyOA (%)
Dimensionality Agreement94.15±0.17
Eigenentropy93.69±0.26
Dimensionality Entropy94.11±0.35
Surface-Variation Peak93.92±0.20
PLE (Ours)94.61±0.26

PLE achieves a mean accuracy 0.46 percentage points higher than that of the best-performing geometry-based adaptive strategy. These geometry-based methods primarily determine neighborhood sizes from local geometric statistics, whereas PLE explicitly considers the aggregation requirements at different network depths, emphasizing fine-grained geometric information at shallower layers and broader spatial context at deeper layers. Furthermore, PLE maintains fixed patch centers and exploits the nested structure of progressively enlarged KNN neighborhoods, gradually incorporating more distant neighbors while preserving the coverage of nearby points. Therefore, PLE aligns neighborhood scheduling with representation depth. In future work, we intend to explore dynamic neighborhood selection mechanisms that jointly consider local geometry and representation hierarchy.

2. LSA: Two-Factor Ablation of Normalization and Aggregation

To clarify the contribution of LSA over the original LNP, we compared the shared scalar standard deviation normalization used in LNP with the proposed local channel-wise RMS normalization, under both mean and Softmax-based weighted aggregation:

Normalization / AggregationMean OA (%)Softmax OA (%)
Shared Scalar Std93.69±0.2694.21±0.36 (LNP)
Local Channel-wise RMS94.15±0.4594.61±0.26 (LSA)

The results show that local channel-wise RMS normalization improves mean accuracy by 0.46 and 0.40 percentage points under mean and Softmax-based aggregation, respectively. The combination of RMS normalization and Softmax-based weighted aggregation achieves the highest mean accuracy (94.61±0.26%), further supporting the effectiveness of LSA's local statistical calibration.

3. SAF: Ablation of Fusion Strategies and Gate Placements

We extended the evaluation to include different fixed fusion coefficients, learnable but input-independent scalar and channel-wise weights, and alternative dynamic gate placements. All experiments used the same pretrained initialization and downstream training and evaluation protocols:

Fusion StrategyOA (%)
Base94.06±0.22
Fixed Coefficient (0.25)93.59±0.19
Fixed Coefficient (0.50)93.83±0.15
Fixed Coefficient (0.75)94.01±0.40
Learnable Scalar Weight93.98±0.75
Learnable Channel-wise Weights94.32±0.30
Dynamic Gate (Forward Only)94.15±0.17
Dynamic Gates (Both Branches)94.49±0.29
SAF94.61±0.26

SAF achieves the highest mean accuracy (94.61±0.26%) among all evaluated fusion strategies. Unlike input-independent weighting strategies that learn fusion weights shared across inputs, SAF dynamically modulates the contribution of the channel-reversed SSM branch based on the current token representations, enabling input-dependent integration of complementary information. These results further support the effectiveness of SAF's dynamic fusion mechanism and gate placement.

Following the reviewer's suggestions, we have clarified the scope of adaptivity in PLE and incorporated these additional experiments into Tables 6, 8, 9, and 10 in Section 4.3 of the revised manuscript. Further experimental details and analyses are provided in our responses to Questions 2 and 3.

AuthorsOpenReview ↗

Response to Question 1: Baseline Difference and Experimental Settings

We are grateful to the reviewer for the careful attention to the fairness and consistency of our baseline experiments. We provide the following clarification regarding the difference in baseline accuracy between Table 1 and the original Table 4.

1. Pretraining Configuration and Baseline Difference

The two results differ in the source of pretrained weights and the model architecture used during pretraining. The 92.05% accuracy in Table 1 is the ScanObjectNN PB_T50_RS classification result reported in the original Mamba3D paper, obtained using the original Mamba3D architecture for both pretraining and downstream fine-tuning. In contrast, the 93.19% accuracy in the original Table 4 (renumbered as Table 5 in the revised manuscript following the addition of S3DIS experiments) was obtained from our component ablation experiments.

Specifically, we first pretrained the complete AdaLoMamba3D architecture and then loaded the resulting pretrained weights for downstream fine-tuning, with PLE, LSA, and SAF disabled, yielding a baseline accuracy of 93.19%. To ensure comparability across the ablation experiments, all configurations in Table 5 use the same pretrained initialization, with only the activation of individual components varied during downstream fine-tuning. Therefore, the 93.19% result represents an ablation baseline under shared pretraining initialization, rather than an independent reproduction of the original Mamba3D architecture throughout pretraining and fine-tuning.

2. Downstream Training and Evaluation Settings

Apart from the aforementioned differences in pretraining architecture and pretrained weight sources, our data augmentation, optimizer, learning rate schedule, training duration, and evaluation protocol are consistent with the corresponding settings of the original Mamba3D. Specifically, we use the AdamW optimizer with an initial learning rate of 0.0005 and a weight decay of 0.05, together with a cosine learning rate schedule over 300 epochs and a 10-epoch warm-up period. The batch size is 32, and each input point cloud contains 2,048 points.

We have revised the manuscript to specify the sources of pretrained weights, the component ablation settings, and the distinction between the two baseline results, thereby improving the clarity of our experimental comparisons.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.