acceptodds
Under review as a conference paper at ICLR 2027

Mantis: Mamba-native Tuning is Efficient for 3D Point Cloud Foundation Models

Abstract

Pre-trained 3D point cloud foundation models (PFMs) have demonstrated strong transferability across diverse downstream tasks. However, full fine-tuning these models is computationally expensive and storage-intensive. Parameter-efficient fine-tuning (PEFT) offers a promising alternative, but existing PEFT approaches are primarily designed for Transformer-based backbones and rely on token-level prompting or feature transformation. Mamba-based backbones introduce a granularity mismatch between token-level adaptation and state-level sequence dynamics. Consequently, straightforward transfer of existing PEFT approaches to frozen Mamba backbones leads to substantial accuracy degradation and unstable optimization. To address this issue, we propose Mantis, the first Mamba-native PEFT framework for 3D PFMs. Specifically, a State-Aware Adapter (SAA) is introduced to inject lightweight task-conditioned control signals into selective state-space updates, enabling state-level adaptation while keeping the pre-trained backbone frozen. Moreover, different valid point cloud serializations are regularized by Dual-Serialization Consistency Distillation (DSCD), thereby reducing serialization-induced instability. Extensive experiments across multiple benchmarks demonstrate that Mantis achieves competitive performance while training less than 5% of the backbone parameters. Our code is available in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point coud, Parameter-efficient fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3onaOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose Mantis, the first Mamba-native PEFT framework for 3D PFMs. The authors identify the mismatch between existing Transformer-native PEFT methods and Mamba-based backbones, and address it through state-level adaptation and cross-serialization regularization. Extensive experiments demonstrate that Mantis achieves strong downstream performance while training less than 5% of the backbone parameters.

Critical strengths
  1. The motivation is clear, and the experiments are extensive.
  2. The paper inspires readers to propose parameter-efficient fine-tuning methods for Mamba-based works.
Critical weaknesses
  1. Figure 2 is key to understanding Mantis's pipeline. However, the figure lacks necessary links between data streams, causing confusion and making it difficult to understand the overall pipeline of the proposed method after reading Section 3.1. Furthermore, Eq. (3) represents a residual connection, but in Figure 2, it is incorrectly shown as linked to S6 rather than ⊕. Additionally, it is recommended that the methodology illustrate, whenever possible, how each operator affects dimension, such as φ(⋅)\varphi \left( \cdot \right), to facilitate readers' understanding.

  2. In the initial pipeline, Mantis is consistent with most previous related works. In terms of order indicators, differing from existing approaches that employ feature-wise affine transformations, the authors model the serialization order as a conditioning signal that modulates the representation in a residual manner. However, there is a lack of insight into this technical change.

  3. SAA essentially modulates the frozen S6 parameters by encoding input signals for parameter-efficient fine-tuning. Why is it considered to integrate task-conditioned control signals into the Mamba state-space updates? It appears that DSCD is the one that integrates task-conditioned control signals.

  4. The authors claim that the Hilbert and Trans-Hilbert curves capture complementary structural cues under different traversal orders. In fact, the Hilbert and Trans-Hilbert curves result from traversing the Hilbert curve in the xyz and yxz orders, respectively. Why are they complementary? Can xyz and zxy be complementary?

  5. Since S6 possesses the property of unidirectional modeling, the Hilbert and Trans-Hilbert traversals are concatenated. When S6 performs state updates on the concatenated point sequence, are the authors considering that a single point in two different spatial relationships could cause structural confusion? Does this confusion affect the state-aware adaptation of SAA?

  6. In Mamba3D, S6's bidirectional modeling operates at the channel level. How can Mantis be used to achieve efficient parameter tuning for Mamba3D?

  7. The authors mention several recent outstanding works in the paper but do not compare them in Tables 1 & 2, such as PointGST[1].

[1] 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.

Questions for authors

Please see the Critical Weaknesses section.

13 comments

We appreciate your recognition of our clear motivation and extensive experimental validation. We hope the responses below will clarify your concerns.

Q1: Could the authors clarify the data-flow connections, residual operation, and tensor dimensions in Figure 2 and the methodology?

A1: We thank the reviewer for these constructive suggestions. We acknowledge that the original Figure 2 did not sufficiently illustrate certain data connections and residual operations. We have revised the figure and provide the following clarifications.

(1) Data flow. We have added explicit arrows to Figure 2 to clarify the connections between modules and the overall information flow, making the pipeline easier to follow.

(2) Residual connection. The original illustration was intended to represent the residual formulation in Eq. (3), but the connection to the addition operator was not explicitly depicted, potentially causing ambiguity. Following the reviewer's suggestion, we have revised Figure 2 to explicitly show the residual path connecting to the addition operator (⊕), ensuring consistency between the figure and Eq. (3).

(3) Tensor dimensions. We have supplemented Sec. 3.1 with clearer tensor shapes and dimensional transformations for key operations. Furthermore, Sec. 3.2 already provides the dimensional definitions of the projection matrices and control signals in SAA. We hope these clarifications will facilitate readers' understanding of the methodology.

Q2: Could the authors explain the motivation and insights behind residual order conditioning instead of feature-wise affine transformations?

A2: We sincerely thank the reviewer for this insightful comment. We would like to clarify that our design is motivated by the different roles of serialization information in PointMamba and Mantis.

In PointMamba[1] (NeurIPS 2024) , order-specific affine transformations encode different serialization orders by directly scaling and shifting token features, allowing the Mamba backbone to distinguish different traversal patterns. This is an effective design for order identification. However, Mantis focuses on state-level adaptation with frozen pre-trained backbones, where serialization information is needed not only to distinguish traversal orders but also to guide the adaptation of state-space dynamics.

To this end, our residual formulation retains the original token features while introducing an input-dependent, order-conditioned correction. The resulting order-aware representation is then used as a conditioning signal for the SAA controller, which dynamically modulates the selective SSM coefficients. Therefore, the key motivation is not that residual transformations are inherently superior to affine transformations, but that our formulation is designed to support serialization-aware state adaptation rather than solely feature-level order identification.

We appreciate the reviewer for providing this opportunity to clarify the motivation behind our design choice.

Q3: How does SAA generate task-conditioned control signals??

A3: We sincerely thank the reviewer for this insightful comment. We apologize that the distinction between SAA and DSCD may not have been sufficiently clear. We would like to clarify that SAA and DSCD serve fundamentally different purposes.

Specifically, SAA generates input-dependent control signals, rather than relying on an explicit task embedding. Its controller takes the current SSM input, intermediate block features, and order-aware representation as inputs to generate control signals that modulate the selective SSM coefficients (BB, CC, and Δ\Delta). Importantly, the controller parameters are optimized under the downstream task objective, enabling the generated control signals to adapt to the requirements of different downstream tasks. Consequently, task-conditioned refers to the task-adapted controller parameters, rather than an explicit task embedding.

In contrast, DSCD serves as a training-time regularization mechanism that encourages cross-serialization consistency in feature representations and predictions. Although its loss can indirectly influence the optimization of SAA through joint training, DSCD does not generate or directly inject control signals into the SSM updates. SAA is responsible for state-level control, whereas DSCD regularizes cross-order consistency.

Q4: Why are Hilbert and Trans-Hilbert considered complementary?

A4: We thank the reviewer for raising this important question. We would like to clarify both the motivation for our serialization choice and the role of different traversal orders in Mantis.

Regarding the serialization choice, our design follows established practices in previous point cloud architectures. Specifically, PointMamba[1] (NeurIPS 2024) adopts Hilbert and Trans-Hilbert as two scanning strategies to provide different spatial locality perspectives. Similarly, Point Transformer V3[2] (CVPR 2024) explores multiple space-filling curve patterns and their axis-reordered variants, showing that changing the traversal-axis priority can provide alternative perspectives on spatial relationships. Following these works, we adopt Hilbert and Trans-Hilbert as an established and computationally efficient dual-serialization configuration, rather than claiming that the xyz/yxz combination is theoretically optimal among all possible axis permutations.

Regarding complementarity, different traversal orders produce different point arrangements and sequential contexts. Since Mamba performs order-sensitive state propagation, these different arrangements can induce distinct contextual representations. We agree that other permutations, such as xyz and zxy, could also provide complementary spatial perspectives in principle.

Importantly, the novelty of Mantis does not lie in introducing these serialization curves, but in how we exploit their complementary representations. Unlike simply using multiple serialization orders to diversify spatial contexts, our proposed DSCD explicitly enforces cross-order consistency at both the feature and prediction levels. This encourages the model to learn consistent semantic representations and task predictions despite variations in serialization order. Furthermore, Table 6(a) provides empirical support for our choice, showing that Hilbert and Trans-Hilbert achieve the best performance among the evaluated serialization combinations. We do not claim universal optimality over all possible axis permutations.

We sincerely thank the reviewer for this suggestion, which helps us better distinguish the role of established serialization strategies from the proposed DSCD contribution.

Q5: Does concatenating the two traversal sequences cause structural confusion or affect SAA?

A5: We sincerely thank the reviewer for raising this important concern. In our implementation, the Hilbert and Trans-Hilbert sequences are concatenated and processed by the S6 layers as a single continuous sequence without resetting the hidden state, following the established practice in PointMamba[1] (NeurIPS 2024). This design is intentional: rather than treating the two traversal orders independently, we allow the state accumulated from the first sequence to provide contextual information for processing the second sequence. The two sequences contain the same geometric points but have distinct order-aware representations and sequential neighborhoods. Therefore, their different spatial relationships reflect alternative traversal contexts rather than conflicting geometric observations.

For SAA, different sequential neighborhoods do not inherently undermine state-aware adaptation. Its controller dynamically generates modulation signals based on the current SSM input, intermediate features, and order-aware representation, allowing state updates to adapt to the actual context. Furthermore, DSCD enforces feature-level and prediction-level consistency across the two traversals, regularizing representation and prediction discrepancies induced by different serialization orders.

Q6: How can Mantis achieve parameter-efficient tuning for Mamba3D?

A6: We sincerely thank the reviewer for raising this question. We note that Mamba3D[3] employs two SSM branches: L+SSM performs forward modeling along the point-token dimension, while C-SSM operates on channel-flipped features to capture feature-level dependencies. The proposed SAA operates at the selective SSM parameter level rather than being tied to a specific sequence dimension. Consequently, Mantis can be extended to Mamba3D by incorporating SAA into both SSM branches. Each SAA generates control signals based on its respective branch inputs to modulate the corresponding selective SSM coefficients. This retains the original Bi-SSM architecture while freezing its backbone parameters and training only lightweight SAA modules and task-specific heads, thereby enabling parameter-efficient adaptation.

Q7: Could the authors include PointGST in Tables 1 and 2 for comparison?

A7: We sincerely thank the reviewer for this valuable suggestion. As discussed in Appendix C.3, we initially selected PMA as a representative PEFT baseline to keep the experimental comparisons focused. We had also included PointGST[4] (TPAMI 2025) in Table 3(c) to provide an additional comparison with existing PEFT methods. However, we acknowledge that this comparison was limited in scope and did not sufficiently cover the main classification and segmentation benchmarks. Following the reviewer's suggestion, we have further incorporated PointGST into Tables 1 and 2 of the revised manuscript, providing more comprehensive comparisons across different backbones and downstream tasks.

Object classification on ScanObjectNN and ModelNet40.

MethodsReference#TP (M)FLOPs (G)OBJ_BGOBJ_ONLYPB_T50_RSModelNet40
Point-MAEECCV2222.1 (100%)4.890.0288.2985.1893.8
+ PointGSTTPAMI250.6 (2.7%)4.891.74 (+1.72)90.19 (+1.90)85.29 (+0.11)93.5 (-0.3)
ReConICML2343.6 (100%)5.395.1893.6390.6394.1
+ PointGSTTPAMI250.6 (1.4%)4.894.51 (-0.67)92.72 (-0.91)89.37 (-1.26)93.5 (-0.6)
PointMambaNeurIPS2412.3 (100%)3.194.3292.6089.3193.6
+ PointGSTTPAMI250.6 (4.9%)4.593.21 (-1.11)91.77 (-0.83)86.93 (-2.38)93.5 (-0.1)
+ Mantis (Ours)This Paper0.6 (4.9%)4.094.65 (+0.33)93.14 (+0.54)89.96 (+0.65)93.7 (+0.1)
Mamba3DACM MM2416.9 (100%)3.993.1292.0892.0594.7
+ PointGSTTPAMI250.6 (3.6%)4.692.08 (-1.04)90.42 (-1.66)87.38 (-4.67)94.2 (-0.5)
+ Mantis (Ours)This Paper0.8 (4.7%)4.093.29 (+0.17)92.77 (+0.69)93.48 (+1.43)94.7 (+0.0)
ZigzagPointMambaNeurIPS2512.3 (100%)3.194.1592.1088.6593.2
+ PointGSTTPAMI250.6 (4.9%)5.193.53 (-0.62)90.72 (-1.38)85.04 (-3.61)93.3 (+0.1)
+ Mantis (Ours)This Paper0.6 (4.9%)4.594.02 (-0.13)93.27 (+1.17)90.11 (+1.46)93.7 (+0.5)

Part segmentation on ShapeNetPart.

MethodsParam. (M)Inst.mIoUCls.mIoU
PointMamba12.285.382.6
+ PointGST0.685.0 (-0.3)82.5 (-0.1)
+ Mantis (Ours)0.685.1 (-0.2)82.9 (+0.3)
Mamba3D15.585.683.6
+ PointGST0.684.1 (-1.5)82.8 (-0.8)
+ Mantis (Ours)0.786.1 (+0.5)84.1 (+0.5)
ZigzagPointMamba12.285.884.2
+ PointGST0.684.4 (-1.4)82.1 (-2.1)
+ Mantis (Ours)0.685.9 (+0.1)84.5 (+0.3)

Indoor semantic segmentation on S3DIS Area 5 and ScanNet V2.

MethodsParam. (M)S3DIS mAccS3DIS mIoUScanNet VmIoUScanNet PmIoU
Point-MAE21.869.960.851.250.8
+ PointGST1.768.458.644.141.9
PointMamba12.270.160.950.350.1
+ PointGST0.665.155.843.543.1
+ Mantis (Ours)0.669.359.849.849.5
Mamba3D15.468.258.451.350.7
+ PointGST0.665.855.045.644.9
+ Mantis (Ours)0.768.058.150.450.0

Notably, the additional results reveal that PointGST does not consistently recover the full fine-tuning performance on Mamba-based backbones. This may be related to the architectural differences between Transformer-based feature interactions and Mamba's selective state-space modeling, as PointGST was originally designed to perform spectral-domain adaptation within Transformer layers rather than SSM state updates. These results further support the effectiveness of Mantis's direct modulation of selective SSM updates.

References

[1] D. Liang, X. Zhou, W. Xu, X. Zhu, Z. Zou, X. Ye, X. Tan, and X. Bai, “PointMamba: A Simple State Space Model for Point Cloud Analysis,” Advances in Neural Information Processing Systems (NeurIPS), 2024.

[2] X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y. Qiao, W. Ouyang, T. He, and H. Zhao, “Point Transformer V3: Simpler, Faster, Stronger,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[3] X. Han, Y. Tang, Z. Wang, and X. Li, “Mamba3D: Enhancing Local Features for 3D Point Cloud Analysis via State Space Model,” Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM), 2024.

[4] D. Liang, T. Feng, X. Zhou, Y. Zhang, Z. Zou, and X. Bai, “Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025.

Reviewer_918XOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Mantis, a parameter-efficient fine-tuning (PEFT) framework designed for pretrained Mamba-based 3D point cloud models. The authors argue that existing PEFT methods, primarily developed for Transformer architectures, do not adequately account for Mamba's selective state-space dynamics and sensitivity to point-cloud serialization. To address these limitations, Mantis introduces two components: (1) a State-Aware Adapter (SAA) that uses input-dependent sparse control signals and low-rank modulation to adapt the selective coefficients (B_t), (C_t), and (Delta_t) while freezing the pretrained backbone; and (2) Dual-Serialization Consistency Distillation (DSCD), which encourages representation-level and prediction-level consistency across different serialization orders of the same point cloud.

The framework is evaluated on three pretrained backbones—PointMamba, Mamba3D, and ZigzagPointMamba—across object classification, few-shot classification, part segmentation, and indoor semantic segmentation. The reported results demonstrate competitive or improved performance compared with full fine-tuning and existing PEFT approaches while updating fewer than 5% of the backbone parameters in the principal configurations. Additional ablations examine the contributions of SAA and DSCD, controller design, adapted selective coefficients, and serialization strategies.

Critical strengths
  • The paper introduces a Mamba-specific parameter-efficient fine-tuning framework that directly adapts selective state-space coefficients through contextual conditioning, sparse control signals, and low-rank modulation while keeping the pretrained backbone frozen.

  • Mantis achieves competitive or superior performance compared with full fine-tuning across multiple Mamba-based backbones, including PointMamba, Mamba3D, and ZigzagPointMamba.

  • On ScanObjectNN PB_T50_RS, Mantis improves Mamba3D accuracy from 92.05% to 93.48% while reducing the number of trainable parameters from 16.9M to approximately 0.84M (Table 1).

  • The method is evaluated across diverse downstream tasks, including object classification, few-shot classification, part segmentation, and indoor semantic segmentation, demonstrating its applicability across different 3D point cloud tasks.

  • The paper provides comprehensive ablation studies investigating the contributions of SAA, feature-level and prediction-level consistency (Table 3a), the impact of inserting SAA into different encoder layers (Table 3b), comparisons with existing PEFT methods (Table 3c), different scanning curves (Table 5a), bottleneck dimensions (Table 5b), controller designs (Table 5c), fusion strategies (Table 6a), and modulation of individual state-space coefficients (Table 6b).

Critical weaknesses
  • Insufficient clarity regarding scanning strategies and the fairness of comparisons: Section 3.1 introduces dual serialization using Hilbert and Transposed Hilbert curves. However, PointMamba, Mamba3D, and ZigzagPointMamba employ different scanning mechanisms in their original architectures, and the paper does not explicitly clarify whether Mantis preserves each backbone's native scanning strategy or introduces a common serialization approach. Furthermore, results compare Mantis with existing PEFT methods, including PMA, PoinTriE, and LoRA, without clearly specifying whether identical scanning strategies, input preprocessing, and training configurations are used. Since Appendix Table 5a demonstrates that scanning strategies substantially affect performance, this ambiguity raises concerns about the fairness of the comparisons and makes it difficult to isolate the improvements attributable to the proposed adaptation mechanism.

  • Missing PoinTriE comparisons on segmentation tasks: Table 1 includes PoinTriE comparisons on Mamba-based backbones, and Table 2b reports PoinTriE results on Point-MAE for semantic segmentation. However, PoinTriE is omitted from the Mamba-based segmentation comparisons in Table 2a (ShapeNetPart) and Table 2b (S3DIS and ScanNet V2). If technically feasible, including PoinTriE on the same Mamba backbones would provide a more comprehensive comparison with existing PEFT methods.

Questions for authors

1. Clarification of scanning strategies and fairness of comparisons: Could the authors clarify the exact scanning strategies used when applying Mantis to PointMamba, Mamba3D, and ZigzagPointMamba? Does Mantis preserve each backbone's original scanning mechanism, or does it introduce Hilbert and Transposed Hilbert serialization across all three architectures? Furthermore, are the competing PEFT methods, including PMA, PoinTriE, and LoRA, evaluated under identical scanning strategies, input preprocessing, pretrained checkpoints, and training configurations? Although Appendix Table 7 provides controlled comparisons with LoRA targeting the native state-space coefficients under dual serialization and DSCD, the exact scanning configurations remain unclear. Given the sensitivity to scanning strategies demonstrated in Appendix Table 5a, clarification and comparisons under matched scanning conditions would help establish whether the reported improvements can be attributed to the proposed adaptation mechanism.

2. Missing comparisons with existing PEFT methods: Could the authors clarify why PoinTriE is not evaluated on Mamba-based backbones for segmentation tasks (Tables 2a and 2b)? Additionally, could the authors include comparisons with PMA and PoinTriE in the few-shot classification experiments (Table 8), if technically feasible, to provide a more comprehensive evaluation?

3. Related work: Please consider citing Parameter-Efficient Fine-Tuning of State Space Models and State-offset Tuning: State-based Parameter-Efficient Fine-Tuning for State Space Models. If feasible, including experimental comparisons would further strengthen the evaluation.

Important Note: The questions raised in the Questions for Authors section, particularly those concerning scanning strategies and the fairness of comparisons, are critical to my assessment. I strongly encourage the authors to address these concerns carefully during the rebuttal, as their responses may materially affect my final recommendation.

3 comments

Thank you for your constructive feedback. We are deeply grateful for your recognition of our method's novelty, efficiency, and effectiveness across diverse 3D tasks. We hope your concerns will be well addressed below.

Q1: Clarification of scanning strategies and fairness of comparisons

A1: We thank the reviewer for emphasizing the importance of scanning strategies and fair comparisons. We clarify the following aspects.

(1) Scanning strategies. Mantis adopts Hilbert and Trans-Hilbert dual serialization across all three backbones, including PointMamba, Mamba3D, and ZigzagPointMamba, while preserving their original internal scanning mechanisms. This establishes a consistent input serialization protocol while retaining the pretrained backbone architectures.

(2) Fair comparisons. Importantly, as indicated by the * notation in Table 1, we reimplemented PMA, PoinTriE, and LoRA and evaluated them under the same Hilbert/Trans-Hilbert serialization protocol. For each backbone, all methods share the same input preprocessing, pretrained checkpoints, data splits, and training configurations. Consequently, the comparisons are conducted under matched conditions rather than different serialization strategies.

(3) Controlled evidence. To further eliminate the influence of serialization strategies, Appendix Table 8 compares SSM-targeted LoRA and Mantis under the same dual-serialization and DSCD training protocol. Both methods adopt identical Hilbert/Trans-Hilbert orders, input preprocessing, and pretrained checkpoints, differing in their adaptation mechanisms. Under these matched conditions, LoRA (B,C,ΔB,C,\Delta) + DSCD achieves 90.63%, whereas Mantis reaches 93.48% (+2.85 pp) on PB_T50_RS. This comparison provides stronger evidence that the additional gains arise from state-aware modulation rather than differences in serialization strategies.

Q2: Can PoinTriE and PMA comparisons be extended to segmentation and few-shot learning?

A2: We thank the reviewer for this valuable suggestion. We clarify the omission and extend the comparisons to segmentation and few-shot classification.

(1) Segmentation comparisons. As discussed in Appendix C.3, we selected PMA as a representative PEFT baseline for Mamba-based segmentation tasks. Given the space constraints, we prioritized PMA in Tables 2a and 2b, while reporting PoinTriE comparisons in object classification (Table 1) and Point-MAE semantic segmentation (Table 2b). We acknowledge that including PoinTriE on Mamba-based segmentation would further strengthen the evaluation.

To address this concern, we have expanded Tables 2a and 2b in the revised manuscript to include PoinTriE on Mamba-based backbones. The expanded comparisons are presented below.

Part segmentation on ShapeNetPart.

MethodsParam. (M)Inst. mIoU ↑Cls. mIoU ↑
PointMamba (Full Fine-tuning)12.285.382.6
+ PMA1.384.2 (-1.1)82.3 (-0.3)
+ PoinTriE0.884.7 (-0.6)82.5 (-0.1)
+ Mantis (Ours)0.685.1 (-0.2)82.9 (+0.3)
Mamba3D (Full Fine-tuning)15.585.683.6
+ PMA1.883.7 (-1.9)82.4 (-1.2)
+ PoinTriE1.183.5 (-2.1)82.7 (-0.9)
+ Mantis (Ours)0.786.1 (+0.5)84.1 (+0.5)
ZigzagPointMamba (Full Fine-tuning)12.285.884.2
+ PMA1.384.1 (-1.7)82.0 (-2.2)
+ PoinTriE0.884.9 (-0.9)82.4 (-1.8)
+ Mantis (Ours)0.685.9 (+0.1)84.5 (+0.3)

Indoor semantic segmentation on S3DIS Area 5 and ScanNet V2.

MethodsParam. (M)S3DIS mAcc ↑S3DIS mIoU ↑ScanNet VmIoU ↑ScanNet PmIoU ↑
Point-MAE (Full Fine-tuning)21.869.960.851.250.8
+ PointLoRA2.264.852.638.137.2
+ PMA2.465.053.239.337.9
+ PoinTriE1.966.154.943.442.1
PointMamba (Full Fine-tuning)12.270.160.950.350.1
+ PMA1.263.854.642.741.6
+ PoinTriE0.865.756.244.843.7
+ Mantis (Ours)0.669.359.849.849.5
Mamba3D (Full Fine-tuning)15.468.258.451.350.7
+ PMA1.864.154.643.142.3
+ PoinTriE1.165.855.945.244.0
+ Mantis (Ours)0.768.058.150.450.0

(2) Few-shot comparisons. Following the reviewer's suggestion, we have also expanded Appendix Table 9 in the revised manuscript to include PMA and PoinTriE in the few-shot experiments. The comparison covers 5-way/10-way classification with 10-shot/20-shot settings, following the original evaluation protocol of 10 independent experiments without voting.

Few-shot classification on ModelNet40.

Methods5-way 10-shot5-way 20-shot10-way 10-shot10-way 20-shot
PointMamba (Full Fine-tuning)96.9 ± 2.099.0 ± 1.193.0 ± 4.495.6 ± 3.2
+ PMA95.5 ± 2.597.5 ± 2.591.3 ± 4.694.7 ± 3.7
+ PoinTriE94.8 ± 2.297.8 ± 1.390.7 ± 3.994.1 ± 3.4
+ Mantis (Ours)97.2 ± 1.999.0 ± 1.093.4 ± 4.096.4 ± 3.1
Mamba3D (Full Fine-tuning)96.4 ± 2.298.2 ± 1.292.4 ± 4.195.2 ± 2.9
+ PMA94.4 ± 3.098.1 ± 1.991.5 ± 4.394.6 ± 3.5
+ PoinTriE93.8 ± 2.898.2 ± 1.792.0 ± 4.094.9 ± 3.1
+ Mantis (Ours)95.3 ± 3.598.9 ± 2.491.7 ± 3.295.3 ± 3.2
ZigzagPointMamba (Full Fine-tuning)96.0 ± 2.199.0 ± 1.290.0 ± 2.294.2 ± 1.0
+ PMA95.2 ± 2.798.4 ± 1.790.9 ± 3.594.5 ± 4.9
+ PoinTriE95.7 ± 2.498.7 ± 1.491.2 ± 1.993.9 ± 3.9
+ Mantis (Ours)96.2 ± 1.699.0 ± 1.192.8 ± 2.195.5 ± 4.6

We sincerely thank the reviewer for this constructive suggestion. These additional comparisons provide a more comprehensive evaluation of Mantis across segmentation and few-shot learning, further clarifying its effectiveness and applicability across different backbones and downstream tasks.

Q3: Could the authors provide additional experimental comparisons with Sparse Dimension Tuning and State-offset Tuning?

A3: We thank the reviewer for recommending these relevant works. In addition to citing these works, we have expanded the experimental evaluation to include Parameter-Efficient Fine-Tuning of State Space Models (ICML 2025), State-offset Tuning: State-based Parameter-Efficient Fine-Tuning for State Space Models (ACL 2025) and MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba (ICLR 2025). We implement these methods on PointMamba, Mamba3D, and ZigzagPointMamba under matched training protocols and comparable parameter budgets. The additional comparisons are presented in Table 4 of the revised manuscript, covering all three ScanObjectNN variants and ModelNet40.

MethodsReference#TP (M)OBJ_BGOBJ_ONLYPB_T50_RSModelNet40
PointMamba (Full Fine-tuning)NeurIPS'2412.3 (100%)94.3292.6089.3193.6
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.6 (4.6%)91.40 (-2.92)90.25 (-2.35)87.10 (-2.21)93.0 (-0.6)
+ MambaPEFT (Additional-scan) [1]ICLR'250.7 (5.4%)92.10 (-2.22)91.05 (-1.55)87.95 (-1.36)93.2 (-0.4)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'250.7 (5.8%)94.40 (+0.08)92.35 (-0.25)89.10 (-0.21)93.6 (+0.0)
+ State-offset Tuning (h) [3]ACL'250.6 (5.1%)93.50 (-0.82)92.72 (+0.12)88.80 (-0.51)93.5 (-0.1)
+ State-offset Tuning (y) [3]ACL'250.6 (5.0%)92.90 (-1.42)91.90 (-0.70)88.20 (-1.11)93.3 (-0.3)
+ Mantis (Ours)This Paper0.6 (4.9%)94.65 (+0.33)93.14 (+0.54)89.96 (+0.65)93.7 (+0.1)
Mamba3D (Full Fine-tuning)ACM MM'2416.9 (100%)93.1292.0892.0594.7
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.8 (4.6%)89.60 (-3.52)88.52 (-3.56)86.84 (-5.21)93.8 (-0.9)
+ MambaPEFT (Additional-scan) [1]ICLR'250.9 (5.4%)90.74 (-2.38)89.65 (-2.43)88.16 (-3.89)94.0 (-0.7)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'251.0 (5.7%)93.24 (+0.12)91.68 (-0.40)90.20 (-1.85)94.5 (-0.2)
+ State-offset Tuning (h) [3]ACL'250.9 (5.2%)91.36 (-1.76)92.18 (+0.10)89.80 (-2.25)94.4 (-0.3)
+ State-offset Tuning (y) [3]ACL'250.9 (5.0%)90.35 (-2.77)89.27 (-2.81)87.96 (-4.09)94.1 (-0.6)
+ Mantis (Ours)This Paper0.8 (5.0%)93.29 (+0.17)92.77 (+0.69)93.48 (+1.43)94.7 (+0.0)
ZigzagPointMamba (Full Fine-tuning)NeurIPS'2512.3 (100%)94.1592.1088.6593.2
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.6 (4.6%)91.25 (-2.90)90.20 (-1.90)86.60 (-2.05)92.6 (-0.6)
+ MambaPEFT (Additional-scan) [1]ICLR'250.7 (5.4%)92.00 (-2.15)91.00 (-1.10)87.40 (-1.25)92.8 (-0.4)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'250.7 (5.8%)94.22 (+0.07)91.85 (-0.25)88.40 (-0.25)93.2 (+0.0)
+ State-offset Tuning (h) [3]ACL'250.6 (5.1%)93.45 (-0.70)92.20 (+0.10)88.10 (-0.55)93.1 (-0.1)
+ State-offset Tuning (y) [3]ACL'250.6 (5.0%)92.95 (-1.20)91.40 (-0.70)87.60 (-1.05)92.9 (-0.3)
+ Mantis (Ours)This Paper0.6 (4.9%)94.02 (-0.13)93.27 (+1.17)90.11 (+1.46)93.7 (+0.5)

Compared with these SSM-specific PEFT approaches, Mantis introduces input-conditioned modulation of the native selective coefficients (B,C,ΔB,C,\Delta), complemented by DSCD to regularize serialization-dependent representations.

We sincerely thank the reviewer for this suggestion, which has helped us strengthen the discussion of existing SSM-specific PEFT methods and provide a more comprehensive evaluation of Mantis.

References

[1] M. Yoshimura, T. Hayashi, and Y. Maeda, "MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba," Proceedings of the International Conference on Learning Representations (ICLR), 2025.

[2] K. Galim, W. Kang, Y. Zeng, H. I. Koo, and K. Lee, "Parameter-Efficient Fine-Tuning of State Space Models," Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025.

[3] W. Kang, K. Galim, Y. Zeng, M. Lee, H. I. Koo, and N. I. Cho, "State-offset Tuning: State-based Parameter-Efficient Fine-Tuning for State Space Models," Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025.

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

Mantis proposes PEFT of pretrained Mamba-based point cloud models. Its State-Aware Adapter modulates selective SSM coefficients through sparse, input-conditioned controls, while Dual-Serialization Consistency Distillation aligns representations and predictions across serialization orders. Experiments cover classification, part segmentation, and semantic segmentation.

Critical strengths
  1. Addresses a relevant problem: adapting frozen Mamba point cloud backbones with limited trainable parameters.
  2. Evaluates three backbones across multiple downstream tasks, with competitive performance against full fine-tuning.
  3. Provides useful controlled experiments, including SSM-oriented LoRA, shared consistency regularization, and practical resource measurements.
Critical weaknesses
  1. Important Mamba/SSM-specific methods, including MambaPEFT, Sparse Dimension Tuning, and State-offset Tuning, are missing. Although ProDiaL and SSM-oriented LoRA are evaluated, the omitted alternatives weaken the evidence for novelty and superiority over existing architecture-aware adaptation approaches.

  2. Table 7 improves LoRA from 83.19% to 89.41% simply by targeting native SSM projections. This substantially weakens the claimed fundamental token/state granularity mismatch. The remaining advantage requires clearer attribution to contextual conditioning, nonlinear modulation, or sparsity.

  3. Eq. (8) specifies exact ZOH input discretization, whereas the standard Mamba selective scan uses a different input-injection formulation. The correspondence to the executed recurrence, the placement of modulation relative to softplus, coefficient tensor layouts, and initialization require clarification.

Questions for authors
  1. How does Mantis compare with the omitted Mamba methods under matched parameter budgets, training protocols, and hyperparameter-search budgets?

  2. Can you provide forward pseudocode clarifying the actual discretization, positivity of the modulated time step, coefficient shapes, and initialization?

  3. Do the gains over SSM LoRA + DSCD persist across other backbones and segmentation tasks under identical preprocessing and order conditioning?

  4. Can you report uncertainty across matched seeds and evaluate unseen serialization orders using both unprojected feature consistency and task accuracy?

5 comments

Thank you for your constructive feedback and for recognizing the relevance of our research problem, competitive performance, and thorough experiments. We hope your concerns will be well addressed below.

Q1: How does Mantis compare with the omitted Mamba methods under matched parameter budgets, training protocols, and hyperparameter-search budgets?

A1: We sincerely thank the reviewer for pointing out these important omissions. We agree that comparisons with existing SSM-specific PEFT methods are necessary to further establish the effectiveness of Mantis. To address this concern, we have added a new comparison in Table 4 of the revised manuscript, covering three representative approaches: MambaPEFT [1] (Affix-tuning and Additional-scan), Sparse Dimension Tuning [2], and State-offset Tuning [3] (h and y).

Specifically, we extend the evaluation to three point cloud backbones, namely PointMamba, Mamba3D, and ZigzagPointMamba. All methods were evaluated using consistent data splits, training configurations, and validation protocols, with comparable trainable parameter budgets and hyperparameter-tuning efforts.

MethodsReference#TP (M)OBJ_BGOBJ_ONLYPB_T50_RSModelNet40
PointMamba (Full Fine-tuning)NeurIPS'2412.3 (100%)94.3292.6089.3193.6
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.6 (4.6%)91.40 (-2.92)90.25 (-2.35)87.10 (-2.21)93.0 (-0.6)
+ MambaPEFT (Additional-scan) [1]ICLR'250.7 (5.4%)92.10 (-2.22)91.05 (-1.55)87.95 (-1.36)93.2 (-0.4)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'250.7 (5.8%)94.40 (+0.08)92.35 (-0.25)89.10 (-0.21)93.6 (+0.0)
+ State-offset Tuning (h) [3]ACL'250.6 (5.1%)93.50 (-0.82)92.72 (+0.12)88.80 (-0.51)93.5 (-0.1)
+ State-offset Tuning (y) [3]ACL'250.6 (5.0%)92.90 (-1.42)91.90 (-0.70)88.20 (-1.11)93.3 (-0.3)
+ Mantis (Ours)This Paper0.6 (4.9%)94.65 (+0.33)93.14 (+0.54)89.96 (+0.65)93.7 (+0.1)
Mamba3D (Full Fine-tuning)ACM MM'2416.9 (100%)93.1292.0892.0594.7
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.8 (4.6%)89.60 (-3.52)88.52 (-3.56)86.84 (-5.21)93.8 (-0.9)
+ MambaPEFT (Additional-scan) [1]ICLR'250.9 (5.4%)90.74 (-2.38)89.65 (-2.43)88.16 (-3.89)94.0 (-0.7)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'251.0 (5.7%)93.24 (+0.12)91.68 (-0.40)90.20 (-1.85)94.5 (-0.2)
+ State-offset Tuning (h) [3]ACL'250.9 (5.2%)91.36 (-1.76)92.18 (+0.10)89.80 (-2.25)94.4 (-0.3)
+ State-offset Tuning (y) [3]ACL'250.9 (5.0%)90.35 (-2.77)89.27 (-2.81)87.96 (-4.09)94.1 (-0.6)
+ Mantis (Ours)This Paper0.8 (5.0%)93.29 (+0.17)92.77 (+0.69)93.48 (+1.43)94.7 (+0.0)
ZigzagPointMamba (Full Fine-tuning)NeurIPS'2512.3 (100%)94.1592.1088.6593.2
+ MambaPEFT (Affix-tuning, w/ Proj.) [1]ICLR'250.6 (4.6%)91.25 (-2.90)90.20 (-1.90)86.60 (-2.05)92.6 (-0.6)
+ MambaPEFT (Additional-scan) [1]ICLR'250.7 (5.4%)92.00 (-2.15)91.00 (-1.10)87.40 (-1.25)92.8 (-0.4)
+ Sparse Dimension Tuning (w/ Proj. LoRA) [2]ICML'250.7 (5.8%)94.22 (+0.07)91.85 (-0.25)88.40 (-0.25)93.2 (+0.0)
+ State-offset Tuning (h) [3]ACL'250.6 (5.1%)93.45 (-0.70)92.20 (+0.10)88.10 (-0.55)93.1 (-0.1)
+ State-offset Tuning (y) [3]ACL'250.6 (5.0%)92.95 (-1.20)91.40 (-0.70)87.60 (-1.05)92.9 (-0.3)
+ Mantis (Ours)This Paper0.6 (4.9%)94.02 (-0.13)93.27 (+1.17)90.11 (+1.46)93.7 (+0.5)

The results demonstrate that Mantis achieves consistently strong transfer performance across all three backbones, with particularly notable gains on the challenging PB_T50_RS split. This suggests that generic SSM adaptation strategies may be insufficient to capture the geometric structures and serialization sensitivity of 3D point clouds. By jointly incorporating input-conditioned state modulation and cross-serialization consistency, Mantis provides a more effective solution for parameter-efficient 3D adaptation.

Q2: Could you provide forward-pass pseudocode

A2: We thank the reviewer for this valuable suggestion. To improve clarity and reproducibility, we have added two algorithms in Appendix C.4 (Algorithm) of the revised manuscript.

Specifically, Algorithm 2 (SAA-Modulated Selective Scan) details the forward procedure, including coefficient shapes, sparse control and low-rank modulation, positive time-step enforcement via Softplus, discretization, and hidden-state initialization and recurrence.

For better readability, Algorithm 1 (Overall Training Pipeline of Mantis) further summarizes the complete workflow, covering dual serialization, SAA adaptation, and DSCD optimization.

We hope these additions will help the reviewer better understand both the overall framework and its core state-space adaptation mechanism.

Q3: Do the gains over SSM LoRA + DSCD generalize across backbones and tasks under matched settings?

A3: We thank the reviewer for this insightful question. To answer this question, we extend the comparison between Mantis and SSM LoRA + DSCD to three backbones and three downstream benchmarks, covering both classification and segmentation. All comparisons use matched preprocessing, serialization orders, order conditioning, and training protocols, ensuring a controlled evaluation of the adaptation mechanisms.

MethodsPB_T50_RS (OA)ShapeNetPart (Cls. mIoU)S3DIS Area 5 (mIoU)
PointMamba (Full Fine-tuning)89.3182.660.9
+ LoRA (B,C,ΔB,C,\Delta)87.94 (-1.37)81.4 (-1.2)57.8 (-3.1)
+ LoRA (B,C,ΔB,C,\Delta) + DSCD88.71 (-0.60)82.2 (-0.4)59.1 (-1.8)
+ Mantis (Ours)89.96 (+0.65)82.9 (+0.3)59.8 (-1.1)
Mamba3D (Full Fine-tuning)92.0583.658.4
+ LoRA (B,C,ΔB,C,\Delta)89.41 (-2.64)82.8 (-0.8)56.6 (-1.8)
+ LoRA (B,C,ΔB,C,\Delta) + DSCD90.63 (-1.42)83.3 (-0.3)57.4 (-1.0)
+ Mantis (Ours)93.48 (+1.43)84.1 (+0.5)58.1 (-0.3)
ZigzagPointMamba (Full Fine-tuning)88.6584.259.2
+ LoRA (B,C,ΔB,C,\Delta)87.85 (-0.80)83.2 (-1.0)57.1 (-2.1)
+ LoRA (B,C,ΔB,C,\Delta) + DSCD88.76 (+0.11)83.9 (-0.3)58.1 (-1.1)
+ Mantis (Ours)90.11 (+1.46)84.5 (+0.3)58.9 (-0.3)

The results show that Mantis consistently outperforms SSM LoRA + DSCD across all three backbones and both classification and segmentation tasks. Specifically, Mantis achieves gains of 1.25, 2.85, and 1.35 pp on PB_T50_RS, alongside consistent improvements of 0.6–0.8 pp on segmentation benchmarks.

Q4: Can you report uncertainty across matched seeds and evaluate unseen serialization orders using both unprojected feature consistency and task accuracy?

A4: We thank the reviewer for this insightful question. We evaluate Mantis and SAA w/o DSCD on Mamba3D with PB_T50_RS using three matched random seeds. Both methods are trained with Hilbert and Trans-Hilbert serialization and evaluated, without retraining, on Hilbert (seen), Z-order (unseen), and Trans-Z-order (unseen).

We report two metrics: (1) OA, the classification accuracy under each serialization order; and (2) Raw Cos, the average cosine similarity between the same point clouds encoded using Hilbert and another order, computed on globally pooled encoder features before the DSCD projection head. Higher Raw Cos. indicates stronger consistency of the unprojected representations. All values are reported as mean ± standard deviation across three matched seeds.

MethodsHilbert OA ↑Z-order OA ↑Trans-Z OA ↑Raw Cos. (H, Z) ↑Raw Cos. (H, Trans-Z) ↑
SAA w/o DSCD91.87 ± 0.2488.92 ± 0.3889.35 ± 0.340.812 ± 0.0090.826 ± 0.008
Mantis (Ours)93.42 ± 0.1891.72 ± 0.2792.03 ± 0.250.891 ± 0.0060.903 ± 0.005

The results demonstrate stable performance across matched random seeds and stronger unprojected feature consistency under unseen serialization orders, accompanied by consistently high task accuracy. These findings indicate that the benefits of DSCD extend beyond the projection space, improving encoder-level consistency and preserving task accuracy under unseen serialization orders.

References

[1] M. Yoshimura, T. Hayashi, and Y. Maeda, "MambaPEFT: Exploring Parameter-Efficient Fine-Tuning for Mamba," Proceedings of the International Conference on Learning Representations (ICLR), 2025.

[2] K. Galim, W. Kang, Y. Zeng, H. I. Koo, and K. Lee, "Parameter-Efficient Fine-Tuning of State Space Models," Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025.

[3] W. Kang, K. Galim, Y. Zeng, M. Lee, H. I. Koo, and N. I. Cho, "State-offset Tuning: State-based Parameter-Efficient Fine-Tuning for State Space Models," Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025.

The authors have provided very thorough supplementary experiments, and I am now inclined to give a score of 3 or higher. Finally, what are your LoRA configuration parameters? Could you also demonstrate, using the Mamba3D baseline, that LoRA remains ineffective even when tested across a wider range of configurations and combinations?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Epilogue by Authors

We would like to express our sincere gratitude to all reviewers and the Area Chair for their time and effort in evaluating our work. We particularly appreciate the reviewers who actively engaged in the discussion and provided thoughtful follow-up questions. Their constructive feedback has been invaluable in improving the clarity, rigor, and completeness of our manuscript.

We are pleased to note that two reviewers who actively participated in the discussion expressed their willingness to raise their scores following our responses, indicating that our rebuttals have successfully addressed many of their initial concerns.

  • Reviewer wUPK recognized the thoroughness of our supplementary experiments and expressed an intention to raise the score to 3 or higher.
  • Reviewer 918X maintained a positive score of 3, reflecting continued recognition of our work.
  • Reviewer 3ona carefully re-evaluated our manuscript following extensive discussions and assigned a score of 3.

Once again, we extend our heartfelt thanks to all reviewers and the Area Chair for their expertise and dedication in evaluating our submission. Your collective insights have been crucial in enhancing the quality and impact of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.