While recent Transformer and Mamba architectures have advanced point cloud representation learning, they are typically developed for single-task or single-domain settings. Directly applying them to multi-task domain generalization (DG) leads to degraded performance. Transformers effectively model global dependencies but suffer from quadratic attention cost and lack explicit structural ordering, whereas Mamba offers linear-time recurrence yet often depends on coordinate-driven serialization, which is sensitive to viewpoint changes and missing regions, causing structural drift and unstable sequential modeling. In this paper, we propose Structure-Aware Domain Generalization (SADG), a Mamba-based In-Context Learning framework that preserves structural hierarchy across domains and tasks. We design structure-aware serialization (SAS) that generates transformation-invariant sequences using centroid-based topology and geodesic curvature continuity. We further devise hierarchical domain-aware modeling (HDM) that stabilizes cross-domain reasoning by consolidating intra-domain structure and fusing inter-domain relations. At test time, we introduce a lightweight spectral graph alignment (SGA) that shifts target features toward source prototypes in the spectral domain without updating model parameters, ensuring structure-preserving test-time feature shifting. In addition, we introduce MP3DObject, a real-scan object dataset for multi-task DG evaluation. Comprehensive experiments demonstrate that the proposed approach improves structural fidelity and consistently outperforms state-of-the-art methods across multiple tasks including reconstruction, denoising, and registration.
Figures & tables
Figure 1: Overview of our SADG . (a) During training, point clouds from multiple source domains are partitioned into local patches and serialized into structure-aware sequences using the Centroid Distance Spectrum (CDS) and Geodesic Curvature Spectrum (GCS), which preserve global topology and local geometric continuity. The serialized sequences are then processed by Mamba blocks under the Hierarchical Domain-Aware Modeling (HDM) mechanism, which stabilizes intra-domain structure and fuses inter-domain relations. (b) At test time, the Spectral Graph Alignment (SGA) performs structure-aware shifting that guides target features toward source prototypes in the spectral domain without updating model parameters, enabling robust generalization to unseen target domains across multiple tasks.
Figure 2: Comparison of different serialization strategies. The proposed CDS and GCS maintain transformation invariance and structural consistency across unaligned real-scan objects, providing a stable foundation for domain-generalized 3D understanding.
Table 1: Comparison on the multi-domain and multi-task benchmark. All models are trained on four source domains and directly evaluated on the remaining one. Evaluation metric: Chamfer Distance (CD, ×10−3 , lower is better).
Variant
Reconstruction
Denoising
Registration
Serialization Variants
Z-order Scanning
7.32
12.47
6.29
Hilbert Curve
6.23
11.13
7.68
w/o CDS
4.75
10.69
6.17
w/o GCS
5.82
8.92
4.43
Ours (full SAS)
3.55
6.61
2.84
Table 2: Ablation study on the proposed components.
Figure 4: Comparison between naive serialization variants and our method, highlighting stable topology and geometry in 3D objects.
Figure 5: Qualitative results on synthetic (ModelNet) and real-scan (MP3DObject) targets. Our method recovers both detailed geometry and smooth surfaces. For visualization, MP3DObject instances are rendered after aligning their non-canonical orientations.
Figure 6: t-SNE visualization of latent features. The Input plot shows encoder-only features, while the latter two depict representations after Mamba-based sequential modeling.
Variant
NPR ↑
Avg. CD ↓
Runtime / batch ↓
Naive BFS (CPU)
1.00
4.27
2.22s
Spectral CDS (GPU, ours)
0.97
4.33
0.75s
Table 3: Naive BFS vs. spectral CDS under MP3DObject as target.
Figure 7: Illustration of why Euclidean distance fails on curved surfaces: straight-line distances shortcut across the token, whereas geodesic distances correctly follow the intrinsic surface geometry.
Method
Topo-NPR ↑
Geo-NPR ↑
Z-order
0.59
0.48
Hilbert curve
0.62
0.48
SAS (CDS+GCS, ours)
0.69
0.67
Table 4: Neighborhood preservation rate under random rotations (higher is better). Our serialization preserves intrinsic neighborhoods more faithfully than coordinate-driven baselines.
Variant
Reconstruction
Denoising
Registration
Naive FPS order
7.73
13.64
7.54
Random traversal
8.17
12.92
7.24
SAS (CDS+GCS, ours)
3.55
6.61
2.84
Table 5: Ablations of serialization strategies on MP3DObject. Naive FPS order and random traversal are clearly inferior to our structure-aware serialization.
Kernel setting
Avg. CD
Comment
Fixed σ=γ=0.05
5.76
too local, sensitive to noise
Fixed σ=γ=0.10
5.02
tuned baseline
Fixed σ=γ=0.20
5.52
overly smooth, loses detail
Adaptive (median-based, ours)
4.33
robust across categories
Table 6: Fixed vs. data-adaptive kernel scales on MP3DObject. We report average Chamfer Distance (CD ×10−3 ) over reconstruction, denoising, and registration. Data-adaptive scales based on median distances provide robust performance without tuning.
Variant
Reconstruction
Denoising
Registration
w/ concat
5.40
7.11
4.23
w/ interleave (ours)
3.55
6.61
2.84
Table 7: Interleaving vs. concatenation in HDM on MP3DObject. Interleaving tokens from both domains enables more effective cross-domain information flow.
Variant
Reconstruction
Denoising
Registration
No SGA
11.95
17.38
9.43
Simple feature shift
7.62
12.56
7.96
SGA (ours)
3.55
6.61
2.84
Table 8: SGA vs. simple feature shifting on MP3DObject. Spectral alignment yields consistently better domain generalization.
Setting
Avg. CD
Comment
α=0.0 (no SGA)
12.92
no alignment applied
α=0.5 (fixed)
6.77
improves over no alignment
α=1.0 (fixed)
10.20
over-alignment, less stable
Adaptive (cosine, ours)
4.33
similarity-based, best overall
Table 9: Effect of alignment strength in SGA on MP3DObject. We report average Chamfer Distance (CD ×10−3 ) over three tasks.
Method
SAS (s)
Fwd. (s)
SGA (s)
Total (s)
FLOPs (G)
Params (M)
DG-PIC [ 28 ]
–
0.94
–
0.94
21.07
27.57
SADG (ours)
0.33
0.25
0.17
0.75
14.89
18.87
Table 10: Runtime breakdown and model complexity.
Dataset
Curvature
Missing
Noise
Orientation
ModelNet
low
low
low
low
ShapeNet
low–mid
low
low
low
ScanNet
mid
mid–high
mid
mid
ScanObjectNN
mid–high
high
high
high
MP3DObject
high
high
high
high
Table 11: Qualitative characterization of datasets along four difficulty dimensions. “Low / mid / high” indicate relative levels to highlight trends. MP3DObject sits at the most challenging end across all dimensions.
Figure 8: Visual comparison of datasets. MP3DObject instances present complex indoor objects in cluttered scenes with substantial occlusions and highly varied poses. For visualization clarity, MP3DObject samples are shown in a manually aligned canonical view; all training and evaluation use the original unaligned scans.
Figure 9: MP3DObject per-class visualization in original and aligned poses. For each class, we show several instances in their original unaligned pose (left of each pair) and in a manually aligned pose (right of each pair) to facilitate visual inspection. The original pose distribution is highly diverse, reflecting realistic scanning conditions. Alignment is applied only for visualization; all training and evaluation in our experiments use the original unaligned MP3DObject scans.
Figure 10: Qualitative comparisons on different target domains.
Recent advances in point cloud In-Context Learning (ICL) have demonstrated strong multitask capabilities. Existing approaches typically adopt a Masked Point Modeling (MPM)-based paradigm for point cloud ICL. However, MPM-based methods directly predict the target point cloud from masked tokens without leveraging geometric priors, requiring the model to infer spatial structure and geometric details solely from token-level correlations via transformers. Additionally, these methods suffer from a training-inference objective mismatch, as the model learns to predict the target point cloud using target-side information that is unavailable at inference time. To address these challenges, we propose DeformPIC, a deformation-based framework for point cloud ICL. Unlike existing approaches that rely on masked reconstruction, DeformPIC learns to deform the query point cloud under task-specific guidance from prompts, enabling explicit geometric reasoning and consistent objectives. Extensive experiments demonstrate that DeformPIC consistently outperforms previous state-of-the-art methods, achieving reductions of 1.6, 1.8, and 4.7 points in average Chamfer Distance on reconstruction, denoising, and registration tasks, respectively. Furthermore, we introduce a new out-of-domain benchmark to evaluate generalization across unseen data distributions, where DeformPIC achieves state-of-the-art performance.
Chengxing Lin, Jinhong Deng, Yinjie Lei +1
Shenzhen Institute for Advanced Study, UESTC · School of Computer Science and Engineering, UESTC · Sichuan University
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.
Zihao Guo, Yiding Sun, Jihua Zhu +3
Xi’an Jiaotong University, Xi’an, China · Peking University, Beijing, China · School of Artificial Intelligence and Robotics, Hunan University, China +1
We dream of a future where point clouds from all domains can come together to shape a single model that benefits them all. Toward this goal, we present Utonia, a first step toward training a single self-supervised point transformer encoder across diverse domains, spanning remote sensing, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds lifted from RGB-only videos. Despite their distinct sensing geometries, densities, and priors, Utonia learns a consistent representation space that transfers across domains. This unification improves perception capability while revealing intriguing emergent behaviors that arise only when domains are trained jointly. Beyond perception, we observe that Utonia representations can also benefit embodied and multimodal reasoning: conditioning vision-language-action policies on Utonia features improves robotic manipulation, and integrating them into vision-language models yields gains on spatial reasoning. We hope Utonia can serve as a step toward foundation models for sparse 3D data, and support downstream applications in AR/VR, robotics, and autonomous driving.
Yujia Zhang, Xiaoyang Wu, Yunhan Yang +6
The University of Hong Kong · Xiaomi · The Chinese University of Hong Kong