Pith. sign in

REVIEW 3 major objections 4 minor 54 references

3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 3DZip keeps 94.7% of 3D question-answering accuracy with just 128 tokens

desk verdict Solid, honest token-compression paper for 3D VLMs; the central claim holds, but test-set-tuned hyperparameters and missing error bars mean the headline margins are optimistic. read the letter →

arxiv 2608.01185 v1 pith:75WRCGMQ submitted 2026-08-02 cs.CV cs.LG

Changwoo Baek , Kyeongbo Kong This is my paper · ORCID
classification cs.CVcs.LG
keywords tokencompression3Dquestionansweringvision-languagemodelsdeterminantalpointprocessfeaturediversityvoxelizationspatiallyconstrainedmergingefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that token compression for 3D vision-language models fails when it treats tokens as a flat 2D bag of features: multi-view projection creates two distinct forms of redundancy, point-level duplication and object-level imbalance, and spatial aggregation alone fixes only the first. 3DZip is a three-stage, training-free pipeline that voxelizes to remove duplicate points, selects anchor tokens with a Determinantal Point Process over feature similarity to spread tokens across distinct objects, and merges remaining tokens only within a spatial threshold. On ScanQA, SQA3D, and OpenEQA, it retains 94.7% of the uncompressed accuracy with only 128 tokens and runs 1.92x faster than the uncompressed LLaVA-3D, beating 2D-oriented pruning methods and voxel-based 3D baselines at every token budget. If correct, this shows that feature-space diversity, not spatial uniformity, is the right criterion for preserving reasoning ability under aggressive compression.

What carries the argument

The carrying mechanism is a three-stage compression pipeline: (1) axis-aligned voxelization with side length $\delta$, aggregating tokens per voxel by mean pooling; (2) greedy Determinantal Point Process anchor selection on the cosine-similarity kernel $L_{kl}=\hat{f}_k^\top \hat{f}_l$ of $\ell^2$-normalized voxel features, maximizing $\log\det(L_A)$; (3) spatially constrained merging that assigns each non-anchor token to its nearest anchor in feature space only if the grid-space distance $d_g \le \tau_g$, and otherwise discards it. The DPP step is the load-bearing novelty: it converts the selection criterion from geometric spread to semantic spread, which is what raises object coverage from

What would settle it

Compare 3DZip against a version with per-scene tuned $(\delta, \tau_g)$ on the large and dense scene quartile of SQA3D at a 32-token budget. If the tuned version gains more than about 1 EM point over the fixed settings, the claim that the fixed hyperparameters are the best overall trade-off (Supp. B.9) would be weakened; if the tuned version does not improve, scene-adaptivity is unnecessary.

Watch

Extended reading notes

Core claim

The central claim is that object-level redundancy in projection-based 3D VLMs is a long-tail token-allocation problem that survives pure spatial aggregation, and that it is best addressed by selecting anchors according to feature-space diversity. 3DZip demonstrates this by decomposing redundancy into three stages: coarse voxelization (point-level), DPP-based anchor selection using cosine similarity of normalized features (object-level), and spatially constrained merging of nearby non-anchor tokens into anchors (geometric consistency). The experiments show consistent gains over FastV, SparseVLM, VisionZip, VisPruner, voxelization-only, and DTC across three benchmarks and token budgets of 128,

Load-bearing premise

The fixed voxel size $\delta=0.2\text{m}$ and merging threshold $\tau_g=5$, chosen by ablations on SQA3D, are assumed to generalize across scenes and benchmarks; the paper's own Limitations section concedes that scene size and object density may warrant different values.

Editorial extensions

If this is right

  • 3DZip is training-free and model-agnostic; it generalizes to Video-3D-LLM and SR-3D under identical token budgets (Table 5).
  • At 128 tokens it retains 94.7% relative accuracy and cuts FLOPs by 90.2% and KV cache by 86.0% (Table 4).
  • It preserves performance best on attribute recognition and object recognition in OpenEQA, categories that require local visual detail (Table 2).
  • Under aggressive compression to 32 tokens, it still retains 88.9% relative accuracy, outperforming all baselines.
  • The component ablations show each stage contributes: removing voxelization drops EM by about 1 point, removing the spatial constraint drops it by 0.5–0.8 points, and feature diversity beats spatial diversity by 2.1–2.7 points.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The fixed voxel size and merging threshold are tuned on SQA3D and flagged by the authors as scene-dependent; a scene-adaptive version that samples $\delta$ and $\tau_g$ from scene volume or object density would plausibly close the residual gap on small-object and fine-detail questions, the failure cases identified in Supp. B.8.
  • The object-coverage analysis suggests that any multi-view lifting VLM that pools by position before selection will inherit the same long-tail bias; feature-DPP could be inserted into other pooling schemes (ConceptFusion/OpenFusion++ styles) with similar gains, as the paper's own Supp. B.7 hints.
  • Because the method is purely geometric and feature-based, it may transfer to 3D dense captioning or open-vocabulary segmentation without retraining; the paper already reports captioning gains on Scan2Cap, so a natural next test is zero-shot transfer to unseen scene types such as HM3D, which the paper could not evaluate due to missing episode trajectories.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes 3DZip, a training-free three-stage token compression method for projection-based 3D vision-language models. The pipeline first applies coarse voxelization to reduce multi-view point-level redundancy, then selects a fixed number of anchor tokens via a Determinantal Point Process over normalized feature similarities, and finally merges nearby non-anchor tokens into anchors under a spatial grid-distance constraint. The method is evaluated on ScanQA, SQA3D, and OpenEQA using LLaVA-3D as the primary backbone, with additional results on Video-3D-LLM and SR-3D. The authors report consistent gains over existing 2D-based and 3D-aware token compression baselines at 128, 64, and 32 tokens, retaining 94.7% of uncompressed accuracy at 128 tokens and achieving a 1.92x inference speedup.

Significance. The problem addressed is timely and practical: projection-based 3D VLMs generate thousands of tokens, and a training-free, geometry-aware compression method could have broad utility. The paper makes a useful empirical observation that spatial aggregation alone leaves an object-level token imbalance, and that feature-space diversity selection improves object coverage. The experimental scope is strong: three benchmarks, three token budgets, component ablations, and generalization to two additional backbones. The DPP objective is a genuine diversity measure and not a relabeling of the target metric, so there is no circularity in the central claim. The main weakness is that the two hyperparameters (voxel size delta and merging threshold tau_g) are selected on the same SQA3D test split used for the headline numbers, and no uncertainty quantification is provided. If the reported gains survive a validation-based selection protocol, the paper would make a solid contribution.

major comments (3)
  1. [Sec. 5.2/5.4, Supp. B.5/B.9, Sec. 6] The two free hyperparameters, delta=0.2m and tau_g=5, are selected by ablating on the SQA3D test split (Table 3b, Supp. B.5), and the same split produces the headline numbers in Table 1. Supp. B.9 shows that optimal values are scene-size dependent: small scenes favor delta=0.1m/tau_g=1 and large scenes favor tau_g=7, with EM differences up to about 1.3 points at 32 tokens. Since the reported margins over the best baselines are only 0.6-2.6 EM points at 128 tokens, selection on the test set could materially inflate the 'consistently outperforms' claim. The Limitations section itself concedes that fixed hyperparameters may be suboptimal across scene types. Please provide a validation-based or pre-specified hyperparameter protocol, and report results under that protocol rather than the best test-set configuration.
  2. [Table 1] No error bars, confidence intervals, or significance tests are reported. Some margins are small (e.g., ScanQA at 128 tokens: 24.2 vs. 23.6 for Voxelization; OpenEQA at 32 tokens: 55.7 vs. 53.5 for VisPruner), and the claim of consistent superiority across all budgets and datasets would be substantially stronger with scene-level bootstrap or per-dataset variance estimates. The use of temperature 0 removes sampling noise but does not account for scene variability.
  3. [Sec. 5.2 and Supp. D.2] The main text and Table 1 report results under the label 'OpenEQA', while Supp. D.2 states that only the ScanNet subset is evaluated because HM3D trajectory states are no longer publicly available. This is a scope reduction that should be stated in the main text and reflected in the abstract's 'three benchmarks' claim. The method may perform differently on HM3D scenes, and the current wording overstates the evidence.
minor comments (4)
  1. [Eq. (15)] The voxel grid index c_k appears to use a floor operation, but the notation is rendered ambiguously. Please write c_k = floor(p_k / delta) explicitly in the text.
  2. [Table 1] The 'Rel.' column is defined only implicitly. State in the caption or text that it is the average of per-dataset retention ratios relative to the LLaVA-3D baseline.
  3. [Sec. 5.2] For OpenEQA, please clarify in the main text that the HM3D portion is excluded and that all reported OpenEQA numbers are for the ScanNet subset, even if the project page uses the broader benchmark name.
  4. [General] The paper would benefit from a code release or a detailed pseudocode implementation of the DPP greedy update (Supp. C.1 provides equations but no reference implementation). This would improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity; method is self-contained against external benchmarks, with a minor non-load-bearing self-citation and a test-set hyperparameter tuning caveat.

full rationale

The central derivation is not circular. 3DZip's three stages are defined directly on token geometry and CLIP features (Eqs. 10-19): mean-pooled voxelization, greedy DPP maximization of a cosine-similarity kernel, and spatially constrained merging. The DPP objective (Eq. 13) is a diversity measure over token features and is never defined in terms of the QA accuracy it is claimed to preserve, so the reported EM/LLM-Match numbers (Table 1) are not equal to the objective by construction. The only self-citation is to the authors' AgilePruner [3] in the Related Works survey of 2D token pruning; it is not used to justify any load-bearing premise or to exclude alternatives, so it does not make the argument circular. The main caveat is that the two hyperparameters delta=0.2 m and tau_g=5 are selected via ablations on the SQA3D test split (Sec. 5.1, Table 3b, Supp. B.5), and the same test split is used for the headline SQA3D comparison; the Limitations section (Sec. 6) and Supp. B.9 concede that fixed hyperparameters are scene-size dependent. This is a test-set selection/overfitting concern, not a self-definitional reduction: the method then applies the same fixed hyperparameters to independent ScanQA/OpenEQA test sets and to Video-3D-LLM/SR-3D backbones (Supp. A.1), where no tuning was performed. Those external comparisons give the central claim independent content. Accordingly, no circular step is exhibited, and the score reflects only the minor self-citation and the tuning caveat.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on domain assumptions about the informativeness of CLIP features and the safety of mean pooling, plus two free hyperparameters selected via ablations on the SQA3D benchmark. No new physical or conceptual entities are introduced.

free parameters (2)
  • Voxel size δ = 0.2 m
    Selected via SQA3D ablations (Table 3b); EM peaks at δ=0.2m for both 32 and 64 token budgets.
  • Grid-space merging threshold τg = 5
    Selected via SQA3D ablations (Supp. Table 9c / B.5); best at 64 tokens, tied at 32. Fixed across all scenes and datasets.
assumptions (3)
  • domain assumption CLIP features of lifted 3D tokens preserve enough semantic and spatial discriminability for cosine-similarity diversity selection to aid QA.
    Stage 2's DPP kernel is built from these features (Sec. 4.2, Eq. 13). If features did not distinguish objects, coverage would not translate to accuracy.
  • domain assumption Mean pooling within voxels and at anchors preserves task-relevant information.
    Stage 1 (Eq. 10) and Stage 3 (Eq. 18) average features. Supp. B.8 shows small-object and fine-detail questions degrade, validating this is a real assumption with known failure modes.
  • standard math The greedy DPP approximation yields a near-optimal diverse subset.
    Uses Chen et al. [5]. Supp. B.3 shows similar results with FPS and MMDP, so the approximation choice is not the main driver, but it remains an unproved reliance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering." pith.science (2026). https://pith.science/paper/75WRCGMQ

@misc{pith2026260801185,
  author       = {Pith},
  title        = {Pith review of: 3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/75WRCGMQ}},
  note         = {Machine review of arXiv:2608.01185}
}
abstract

Recent 3D vision-language models (3D VLMs) construct geometry aware tokens by projecting 2D visual features into world coordinates, enabling spatial reasoning for tasks such as 3D question answering. However, this design generates thousands of tokens per scene, resulting in substantial computational and memory overhead. While token compression has been extensively studied in 2D VLMs, existing approaches rely on semantic relevance or attention-based selection that overlook the structured spatial nature of 3D tokens. Moreover, redundancy in 3D representations cannot be resolved by spatial proximity alone, as object-level token imbalance persists even after spatial aggregation. To address this, we propose 3DZip, a three-stage token compression framework that first applies coarse voxelization to remove point-level redundancy, then selects anchor tokens based on feature-space diversity via a Determinantal Point Process, and finally merges remaining tokens under spatial constraints to preserve geometric coherence. Experiments on three 3D question answering benchmarks demonstrate that 3DZip consistently outperforms existing compression methods, retaining 94.7% of the original performance with only 128 tokens, achieving a $1.92\times$ faster inference speed.

Figures

Figures reproduced from arXiv: 2608.01185 by the authors.

Figure 1
Figure 1. (a) Per-object token allocation under different selection strategies. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of geometry-aware 3D token construction and the com [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed three-stage token compression pipeline. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of retained 3D tokens on OpenEQA. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Additional qualitative examples on SQA3D. [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Additional qualitative examples on OpenEQA. [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Additional qualitative examples on ScanQA. [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Object-level token allocation across selection strategies on the full [PITH_FULL_IMAGE:figures/full_fig_p028_8.png]
Figure 9
Figure 9. Figure 9: Prompt used for LLM-based evaluation in OpenEQA. ScanQA We evaluate on the validation split of ScanQA [2], which contains 4,675 questions across 71 ScanNet scenes. Each question asks about object attributes, spatial relationships, or scene properties, and is paired wit…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 29 canonical work pages

  1. [1]

    Alvar, S.R., Singh, G., Akbari, M., Zhang, Y.: Divprune: Diversity-based visual tokenpruningforlargemultimodalmodels.In:ProceedingsoftheComputerVision and Pattern Recognition Conference. pp. 9392–9401 (2025)

    2025

  2. [2]

    In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: Scanqa: 3d question answer- ing for spatial scene understanding. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19129–19139 (2022)

    2022

  3. [3]

    In: The Fourteenth International Conference on Learning Representations (2026)

    Baek, C., Song, J., Kim, S., Kong, K.: Agilepruner: An empirical study of attention and diversity for adaptive visual token pruning in large vision-language models. In: The Fourteenth International Conference on Learning Representations (2026)

    work page 2026

  4. [4]

    arXiv preprint arXiv:2309.16609 (2023)

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

    arXiv 2023

  5. [5]

    Advances in neural information pro- cessing systems31(2018)

    Chen, L., Zhang, G., Zhou, E.: Fast greedy map inference for determinantal point process to improve recommendation diversity. Advances in neural information pro- cessing systems31(2018)

    work page 2018

  6. [6]

    In: European Conference on Computer Vision

    Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., Chang, B.: An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In: European Conference on Computer Vision. pp. 19–35. Springer (2024)

    2024

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., Chen, T.: Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 26428–26438 (2024)

    2024

  8. [8]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Chen, Z., Gholami, A., Nießner, M., Chang, A.X.: Scan2cap: Context-aware dense captioning in rgb-d scans. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 3193–3203 (2021)

    2021

Show all 54 references
  1. [9]

    In: The Fourteenth International Conference on Learning Representations (2026),https://openreview.net/forum?id=GTpf2NuwtR

    Cheng, A.C., Fu, Y., Chen, Y., Liu, Z., Li, X., Radhakrishnan, S., Han, S., Lu, Y., Kautz, J., Molchanov, P., Yin, H., Wang, X., Liu, S.: 3d aware region prompted vision language model. In: The Fourteenth International Conference on Learning Representations (2026),https://open...

    2026

  2. [10]

    In: Proc

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proc. Computer Vision and Pattern Recognition (CVPR), IEEE (2017)

    2017

  3. [11]

    Advances in neural information pro- cessing systems35, 16344–16359 (2022)

    Dao, T., Fu, D., Ermon, S., Rudra, A., Ré, C.: Flashattention: Fast and memory- efficient exact attention with io-awareness. Advances in neural information pro- cessing systems35, 16344–16359 (2022)

    2022

  4. [12]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Deng, J., He, T., Jiang, L., Wang, T., Dayoub, F., Reid, I.: 3d-llava: Towards gener- alist 3d lmms with omni superpoint transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 3772–3782 (2025) 3DZip 17

    2025

  5. [13]

    In: Proceedings of the IEEE conference on com- puter vision and pattern recognition workshops

    DeTone, D., Malisiewicz, T., Rabinovich, A.: Superpoint: Self-supervised interest point detection and description. In: Proceedings of the IEEE conference on com- puter vision and pattern recognition workshops. pp. 224–236 (2018)

    2018

  6. [14]

    IEEE transactions on image processing6(9), 1305– 1315 (1997)

    Eldar, Y., Lindenbaum, M., Porat, M., Zeevi, Y.Y.: The farthest point strategy for progressive image sampling. IEEE transactions on image processing6(9), 1305– 1315 (1997)

    1997

  7. [15]

    In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV)

    Fu, R., Liu, J., Chen, X., Nie, Y., Xiong, W.: Scene-llm: Extending language model for 3d visual reasoning. In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV). pp. 2195–2206 (February 2025)

    2025

  8. [16]

    In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

    Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: Injecting the 3d world into large language models. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 20482...

    2023

  9. [17]

    Advances in Neural Information Processing Systems 37, 113991–114017 (2024)

    Huang, H., Chen, Y., Wang, Z., Huang, R., Xu, R., Wang, T., Liu, L., Cheng, X., Zhao, Y., Pang, J., et al.: Chat-scene: Bridging 3d scene and large language models with object identifiers. Advances in Neural Information Processing Systems 37, 113991–114017 (2024)

    2024

  10. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Huang, H.W., Chen, F.C., Chai, W., Su, C.C., Xia, L., Jung, S., Yang, C.Y., Hwang, J.N., Sun, M., Kuo, C.H.: Zero-shot 3d question answering via voxel- based dynamic token compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)...

    2025

  11. [19]

    In: International Conference on Machine Learning

    Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., Li, Q., Zhu, S.C., Jia, B., Huang, S.: An embodied generalist agent in 3d world. In: International Conference on Machine Learning. pp. 20413–20451. PMLR (2024)

    2024

  12. [20]

    arXiv preprint arXiv:2410.21276 (2024)

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Os- trow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

    2024 arXiv

  13. [21]

    arXiv preprint arXiv:2302.07241 (2023)

    Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Maalouf, A., Li, S., Iyer, G., Saryazdi, S., Keetha, N., et al.: Conceptfusion: Open-set mul- timodal 3d mapping. arXiv preprint arXiv:2302.07241 (2023)

    2023 arXiv

  14. [22]

    In: 2025 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS)

    Jin, X., Frosi, M., Matteucci, M.: Openfusion++: An open-vocabulary real-time scene understanding system. In: 2025 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS). pp. 634–641. IEEE (2025)

    2025

  15. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kang, W., Huang, H., Shang, Y., Shah, M., Yan, Y.: Robin3d: Improving 3d large language model via robust instruction tuning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3905–3915 (2025)

    2025

  16. [24]

    arXiv preprint arXiv:2505.18227 (2025)

    Kong, Z., Li, Y., Zeng, F., Xin, L., Messica, S., Lin, X., Zhao, P., Kellis, M., Tang, H., Zitnik, M.: Token reduction should go beyond efficiency in generative models–from vision, language to multimodality. arXiv preprint arXiv:2505.18227 (2025)

    2025

  17. [25]

    In: Proceedings of the AAAI Confer- ence on Artificial Intelligence

    Lin, Z., Lin, M., Lin, L., Ji, R.: Boosting multimodal large language models with visual tokens withdrawal for rapid inference. In: Proceedings of the AAAI Confer- ence on Artificial Intelligence. vol. 39, pp. 5334–5342 (2025)

    2025

  18. [26]

    In: Proceedings of the IEEE/CVF international conference on com- puter vision

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the IEEE/CVF international conference on com- puter vision. pp. 17627–17638 (2023) 18 Baek and Kong

    2023

  19. [27]

    arXiv preprint arXiv:2412.19437 (2024)

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  20. [28]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tun- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 26296–26306 (2024)

    2024

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, Z., Zhu, L., Shi, B., Zhang, Z., Lou, Y., Yang, S., Xi, H., Cao, S., Gu, Y., Li, D., et al.: Nvila: Efficient frontier visual language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4122– 4134 (2025)

    2025

  22. [30]

    Advances in Neural Information Processing Systems37, 50898–50924 (2024)

    Lyu, R., Lin, J., Wang, T., Yang, S., Mao, X., Chen, Y., Xu, R., Huang, H., Zhu, C., Lin, D., et al.: Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. Advances in Neural Information Processing Systems37, 50898–50924 (2024)

    2024

  23. [31]

    In: The Eleventh International Conference on Learning Representations (2023)

    Ma,X.,Yong,S.,Zheng,Z.,Li,Q.,Liang,Y.,Zhu,S.C.,Huang,S.:Sqa3d:Situated question answering in 3d scenes. In: The Eleventh International Conference on Learning Representations (2023)

    2023

  24. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Majumdar, A., Ajay, A., Zhang, X., Putta, P., Yenamandra, S., Henaff, M., Silwal, S., Mcvay, P., Maksymets, O., Arnaud, S., et al.: Openeqa: Embodied question answering in the era of foundation models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern r...

    2024

  25. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition

    Man, Y., Gui, L.Y., Wang, Y.X.: Situational awareness matters in 3d vision lan- guage reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 13678–13688 (2024)

    2024

  26. [34]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Qi, Z., Fang, Y., Sun, Z., Wu, X., Wu, T., Wang, J., Lin, D., Zhao, H.: Gpt4point: A unified framework for point-language understanding and generation. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26417–26427 (2024)

    2024

  27. [35]

    In: Thirty-fifth Conference on Neural Infor- mation Processing Systems Datasets and Benchmarks Track (Round 2) (2021), https://openreview.net/forum?id=-v4OuqNs5P

    Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J.M., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In: Thirty-fif...

    2021

  28. [36]

    Computers & Operations Research37(3), 498–508 (2010)

    Resende, M.G., Martí, R., Gallego, M., Duarte, A.: Grasp and path relinking for the max–min diversity problem. Computers & Operations Research37(3), 498–508 (2010)

    2010

  29. [37]

    In: ICCV (2025)

    Shang, Y., Cai, M., Xu, B., Lee, Y.J., Yan, Y.: Llava-prumerge: Adaptive token reduction for efficient large multimodal models. In: ICCV (2025)

    2025

  30. [38]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Shi, X., Wu, Z., Lee, S.: Aware visual grounding in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14056–14065 (2024)

    2024

  31. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Song, J., Kim, S., Kong, K.: Uncertainty-guided graph formulation via mwis for token pruning in lvlms. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9510–9519 (2026)

    2026

  32. [40]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  33. [41]

    arXiv preprint arXiv:2502.14786 (2025)

    Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al.: Siglip 2: 3DZip 19 Multilingual vision-language encoders with improved semantic understanding, lo- calization, and dense features. arX...

    2025 arXiv

  34. [42]

    arXiv preprint arXiv:2409.12191 (2024)

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

    2024 arXiv

  35. [43]

    arXiv preprint arXiv:2308.08769 (2023)

    Wang, Z., Huang, H., Zhao, Y., Zhang, Z., Zhao, Z.: Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769 (2023)

    2023 arXiv

  36. [44]

    arXiv preprint arXiv:2410.17247 (2024)

    Xing, L., Huang, Q., Dong, X., Lu, J., Zhang, P., Zang, Y., Cao, Y., He, C., Wang, J., Wu, F., et al.: Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247 (2024)

    2024 arXiv

  37. [45]

    In: European Conference on Computer Vision

    Xu, R., Wang, X., Wang, T., Chen, Y., Pang, J., Lin, D.: Pointllm: Empowering large language models to understand point clouds. In: European Conference on Computer Vision. pp. 131–147. Springer (2024)

    2024

  38. [46]

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., Dong, G., Wei, H., Lin, H., Tang, J., Wang, J., Yang, J., Tu, J., Zhang, J., Ma, J., Yang, J., Xu, J., Zhou, J., Bai, J., He, J., Lin, J., Dang, K., Lu, K., Chen, K., Yang, K., Li, M....

    2024 arXiv

  39. [47]

    In: Proceedings of the Com- puter Vision and Pattern Recognition Conference

    Yang, S., Chen, Y., Tian, Z., Wang, C., Li, J., Yu, B., Jia, J.: Visionzip: Longer is better but not necessary in vision language models. In: Proceedings of the Com- puter Vision and Pattern Recognition Conference. pp. 19792–19802 (2025)

    2025

  40. [48]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Ye, W., Wu, Q., Lin, W., Zhou, Y.: Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 22128–22136 (2025)

    2025

  41. [49]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, Q., Cheng, A., Lu, M., Zhang, R., Zhuo, Z., Cao, J., Guo, S., She, Q., Zhang, S.: Beyond text-visual attention: Exploiting visual cues for effective token pruning in vlms. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20857–20867 (2025)

    2025

  42. [50]

    arXiv e-prints pp

    Zhang, Q., Cheng, A., Lu, M., Zhuo, Z., Wang, M., Cao, J., Guo, S., She, Q., Zhang, S.: [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv e-prints pp. arXiv–2412 (2024)

    2024

  43. [51]

    In: International Conference on Machine Learning (2025)

    Zhang, Y., Fan, C.K., Ma, J., Zheng, W., Huang, T., Cheng, K., Gudovskiy, D., Okuno, T., Nakata, Y., Keutzer, K., et al.: Sparsevlm: Visual token sparsifica- tion for efficient vision-language model inference. In: International Conference on Machine Learning (2025)

    2025

  44. [52]

    Zheng, D., Huang, S., Wang, L.: Video-3d llm: Learning position-aware video repre- sentationfor3dsceneunderstanding.In:ProceedingsoftheIEEE/CVFConference on Computer Vision and Pattern Recognition. pp. 8995–9006 (2025)

    2025

  45. [53]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhu, C., Wang, T., Zhang, W., Pang, J., Liu, X.: Llava-3d: A simple yet effec- tive pathway to empowering lmms with 3d capabilities. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4295–4305 (2025)

    2025

  46. [54]

    Black.” ✘ DTC: “Black

    Zhu, Z., Ma, X., Chen, Y., Deng, Z., Huang, S., Li, Q.: 3d-vista: Pre-trained trans- former for 3d vision and text alignment. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 2911–2921 (2023) 20 Baek and Kong Supplementary Material Overview Th...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.