Pith.

Computer Vision and Pattern Recognition · arXiv:2610.20990 v1 · submitted 14 Oct 2026 · 44 references

DeSpatial: The devil of Agentic Spatial Reasoning Lies in 3D Itself

Yifei Liu, Xiaolong Zhou, Ziyang Gong, Yifan Zhou, Yuanyuan Gao, Yuning Gong, Yuchen Yang, Hongjie Zhang

show 3 more authors

Xue Yang, Xiao Sun, Zhihang Zhong

The paper says

“Spatial agents have access to reconstruction and segmentation tools, yet inconsistent predictions and fragmented object identities can prevent their outputs from becoming reliable 3D scene information.”

more of the abstract

A simple oracle experiment exposes this perception bottleneck: replacing tool-predicted perception with ground-truth 3D bounding boxes raises an agent to nearly full score across five object-centric VSI-Bench tasks, while retaining the same reasoning tools. Motivated by this finding, we believe the devil of spatial reasoning lies in 3D itself, and introduce DeSpatial(Devil of Spatial), which explicitly decomposes spatial understanding into perception and reasoning. DeSpatial has a shared 3D perception tool to reconcile independent perception outputs into a unified, metric, gravity-aligned 3D scene. This perception tool has two core components: a) a joint segmentation–depth co-refinement module that consolidates masks of the same object and removes projection artifacts caused by inconsistent segmentation and depth within a single frame; b) a cross-view instance fusion module that consolidate voxel-overlapping instances but not allowed to merge same-frame instances. This fusion module additionally fuses spatially disjoint fragments caused by insufficient multi-view observations, by checking their 3D structure and source mask appearance. After receiving the complete and unified 3D scene information, the agent chooses reasoning or computational tools to solve the spatial problem. DeSpatial achieves leading performance on VSI-Bench, MindCube-Tiny, and ViewSpatial, with additional long-video results on VSI-SUPER, revealing the bottleneck of spatial understanding lies in a precise 3D perception itself.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Oracle experiment is the real contribution; the method is competent engineering but hyperparameters tuned on the same benchmark without a validation set, and VLM judge calibration is missing. This line is wrong

3 major objections · 8 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 44 references say.

The paper says, at Section 4.3, Table 5; Section 3.3; Appendix B.3

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

Hyperparameter tuning without a separated validation set: The ablation study and all reported results use hyperparameters (voxel size 0.1 m, τ_ov=0.1, θ=20°, τ_c=0.08, τ_d=0.02, λ=5, k=1, etc.) that appear to be tuned on the same 504-question VSI-Bench subset used to measure ablation gains.

more of objection 1

The paper does not describe a held-out validation procedure, and the stratified subset is drawn from the full benchmark. If these thresholds were selected to maximize performance on this subset, the reported improvements (e.g., +11.27 for co-refinement, +2.01 for visual fragment association) may reflect overfitting to the particular scene distribution and question statistics of that subset. The authors should either (a) report a separate validation split or cross-validation used for threshold selection, (b) demonstrate that the method is robust to reasonable variations in each threshold, or (c) acknowledge this as a limitation with concrete evidence that the chosen thresholds are not critical.

The paper says, at Section 3.3, Appendix B.3

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Uncalibrated VLM judge: The visual fragment association and occlusion-resolution steps rely on a VLM judge that decides whether two mask candidates belong to the same physical object. The paper does not report any accuracy, precision, or recall of this judge on a held-out set of fragment pairs.

more of objection 2

Because the judge's decisions directly affect which observations are merged into a single instance, a systematic bias (e.g., preferring to merge for visually similar fragments) could inflate recall without genuine geometric consistency. The authors should report judge accuracy on a sample of ground-truth fragment pairs, and ideally provide the judge's prompt and decision criteria in full. At minimum, a sensitivity analysis should show how varying the judge's 'same' threshold affects downstream VSI-Bench scores.

The paper says, at Section 2, Table 1; Section 6

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.74):

“The oracle results concern five object-centric categories and answer-consistent annotations; they do not establish that perception is the only limitation across all spatial tasks. 9 Preprint REFERENCES Shuai Bai, Yuxuan Cai, Ruizhe Chen, et al. Qwen3-VL technical report, 2025a.”

Pith says: Objection 3 of 3 · major

Oracle experiment scope: The oracle diagnostic covers only five object-centric categories (counting, absolute distance, object size, relative distance, relative direction) and uses annotations that are 'answer-consistent' with the benchmark's questions. The authors explicitly note that 'the oracle results concern five object-centric categories and answer-consistent annotations; they do not establish that perception is the only limitation across all spatial tasks' (Section 6).

more of objection 3

This is a properly stated limitation. However, the paper's title and abstract make a stronger claim ('the devil of spatial reasoning lies in 3D itself'). To avoid overstatement, the claim in the title and abstract should be qualified to reflect the diagnostic's scope (object-centric measurements) and reliance on annotation-consistent ground truth, rather than implying a universal bottleneck.

8 minor objections

The paper says, at Abstract

Pith says: minor objection 1 of 8

The abstract states 'DeSpatial achieves leading performance on VSI-Bench, MindCube-Tiny, and ViewSpatial'. However, the VSI-Bench average (69.22%) does not surpass all prior work listed in Table 2 (e.g., AlloSpatial at 61.0% is lower, but the comparison is not apples-to-apples due to different backbones and settings). 'Leading' is acceptable but should note the backbone difference. Consider rephrasing to 'competitive performance' or specify the setting.

The paper says, at Fig. 1, Fig. 2, Fig. 3

Pith says: minor objection 2 of 8

The figures are dense and the captions are long. Consider adding callout numbers or letters in the figures that correspond to the steps described in the text, so readers can more easily map the narrative to the visuals.

show 6 more

The paper says, at Section 3.2, Eq. (7)

Pith says: minor objection 3 of 8

The duplicate mask consolidation uses an overlap ratio threshold of 0.5. What is the justification for this value? A brief note or reference would be helpful, especially since the mask consolidation step affects downstream geometry.

The paper says, at Section 3.3, Eq. (5)

Pith says: minor objection 4 of 8

The overlap score w_ij is normalized by the smaller occupied set. This is a reasonable choice, but it means that a small partial observation can have high overlap with a complete observation even if the complete observation is much larger. A comment about the potential sensitivity to observation extent would be useful.

The paper says, at Appendix B.3

Pith says: minor objection 5 of 8

The fragment candidate retrieval uses 'fifth-percentile distances' and 'top-three neighbors'. The choice of percentile and number of neighbors seems arbitrary. Since these parameters affect recall of fragment pairs, some ablation or justification (or a note that results are not overly sensitive) would strengthen the exposition.

The paper says, at Table 2

Pith says: minor objection 6 of 8

In the VSI-Bench results, the row for VLM-3R shows 60.9% average with backbone LLaVA-NV-7B. Later models like DeSpatial use Qwen3-VL-8B. The backbone difference makes direct comparison difficult. The paper should discuss the extent to which the backbone contributes to gains vs. the perception tool itself. The oracle experiment already establishes the ceiling, but a controlled comparison (DeSpatial with the same backbone as prior agents, e.g., InternVL3-8B) would be informative.

The paper says, at Section 5 (Related Work)

Pith says: minor objection 7 of 8

The related work section is comprehensive but reads like a list. Consider grouping methods by their approach to resolving perception inconsistency (e.g., voxel-level fusion, program synthesis, interactive exploration) to better position DeSpatial's contribution.

The paper says, at Throughout

Pith says: minor objection 8 of 8

There are minor formatting issues: in the abstract 'DeSpatial(Devil ofSpatial)' is missing a space. Also, some references are inconsistently formatted (e.g., 'arXiv preprint' vs. 'arXiv preprint' capitalization). These should be corrected for camera-ready version.

What would settle it

Reserve 20% of VSI‑Bench scenes (stratified by category) as a validation set before any tuning. Re‑run the complete DeSpatial pipeline with the hyperparameters reported in the paper; compute the eight‑category average on that held‑out set. If the score drops by more than 3 percentage points relative to the reported 69.22, or if the per‑category ranking of ablations in Table 5 changes, the hyperparameters are overfit. Additionally, curate 200 paired fragment observations (100 merge positives, 100 negatives) and measure the VLM judge's binary accuracy; if below 80%, the visual fragment association step may introduce systematic error.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 16 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 29 s · 18,878 tokens read, 2,843 written
  3. Editor wrote the verdict above 24 s · 19,751 tokens read, 2,479 written
  4. Referee major revision: 3 major objections, 8 minor 20 s · 19,183 tokens read, 2,614 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The main thing to know: the oracle experiment in Section 2 is the real contribution. It cleanly shows that with ground-truth 3D boxes, an agent scores 98% on five VSI-Bench tasks versus 50% with raw perception tools, using the same reasoning pipeline. That diagnostic is novel, well-controlled, and worth having on record. The authors are honest about its limits—they say the annotations align with the benchmark answers, so it measures compatibility rather than perfect geometry (b52, b270).

The DeSpatial framework itself is competent engineering. The joint segmentation–depth co-refinement and cross-view instance fusion are sensible solutions to real problems (mask fragmentation, projection artifacts, cross-view identity). The ablations show each component adds something. The gains on VSI-Bench (69.22 avg), MindCube-Tiny, and ViewSpatial are real improvements over prior agentic approaches.

The soft spot the stress-test note hits: hyperparameters are tuned on the same benchmark without a held-out validation set. Voxel size, overlap threshold, angular threshold, confidence threshold, depth-edge threshold, Mahalanobis multiplier—all set to values that work on VSI-Bench. The ablation uses a 504-question subset of the same benchmark. The paper describes no validation split or cross-validation. So the reported numbers may be optimistic for new scenes or sensor setups. Related: the VLM judge used for fragment association has zero calibration; its accuracy is never measured, and its decisions feed directly into the scene representation.

These concerns are real but not fatal. The oracle experiment supports the central claim independently; hyperparameter sensitivity only affects the magnitude of the claimed gain, not the diagnostic. The authors acknowledge most other limitations in their Section 6. What they don't acknowledge is the validation blind spot.

This paper is for anyone building spatial agents or designing perception pipelines for 3D scene understanding. It deserves peer review. The authors should add a validation protocol (held-out scenes or cross-validation) and calibrate their VLM judge before the final version.

The strongest objection

The paper's ablations and cross-view fusion hyperparameters (voxel size 0.1 m, overlap threshold τ_ov=0.1, angular threshold θ=20°, etc.) are tuned on the same 504‑question VSI‑Bench subset used for the ablation study. No held‑out validation procedure is described. This means the reported improvements (e.g., +11.27 for co‑refinement, +2.01 for visual fragment association) may reflect overfitting to the particular scene distribution and question statistics of that subset. Furthermore, the VLM judge used for fragment association and occlusion resolution has zero calibration: its accuracy is never measured, and its decisions are accepted at face value. A systematic bias in the judge—e.g., preferring to merge for visually similar fragments—would inflate recall without genuine geometric improvement. Because the judge's decisions directly feed into the final scene representation, an uncalibrated judge weakens the claim that the perception bottleneck is being addressed in a principled way.

The reader's weakest assumption—that hyperparameters tuned on the same benchmark without a validation set may not generalize—is the most load‑bearing concern. I concur and propose a concrete out‑of‑sample test to settle it. Additionally, I highlight the uncalibrated VLM judge as a secondary but related risk, which the reader also noted. Neither concern invalidates the paper's central diagnostic (oracle experiment), but both affect the trustworthiness of the method's reported advantage. Since the reader already issued a CONDITIONAL verdict, and my analysis supports that judgment without compelling a stronger change, UNCHANGED is appropriate.

The referee's report

If the central claim holds—that 3D perception accuracy is the dominant bottleneck—then the paper provides a clear direction for future spatial agent design: invest in making perception outputs consistent and metric, rather than in more elaborate reasoning or tool-calling strategies. The oracle experiment is clean and persuasive, and the proposed method's components are well motivated by specific failure cases (fragmented identities, geometric contamination). The empirical results across multiple benchmarks (VSI-Bench, MindCube-Tiny, ViewSpatial, VSI-SUPER) are strong and largely consistent. The paper also releases practical engineering contributions (co-refinement, fusion rules, gravity alignment) that could be reused. However, the significance is somewhat tempered by the fact that the oracle study covers only five object-centric categories (not all eleven VSI-Bench categories) and uses answer-consistent annotations, which the authors acknowledge. The broader claim that the 'devil lies in 3D itself' is supported for those tasks, but generalization to more complex spatial tasks (room size, route planning, appearance order) remains less tested.

Pith's reading of the paper (a machine's summary, not the authors')

This paper argues that the primary obstacle to agentic spatial reasoning is not the reasoning ability of the agent but the accuracy of the 3D perception that feeds it. The authors prove this with an oracle experiment: when ground-truth 3D oriented bounding boxes replace raw perception tools, the same agent jumps from 50.38% to 98.23% on five object-centric benchmark tasks while keeping all reasoning tools identical. To close this gap, the paper introduces DeSpatial, a shared 3D perception tool that reconciles inconsistent segmentation and depth outputs from independent models. It uses a joint segmentation–depth co-refinement module to clean per-frame observations and a cross-view instance fusion module to assign consistent object identities across frames. The result is a unified, metric, gravity-aligned 3D scene that downstream reasoning tools can directly use, yielding leading scores on multiple spatial reasoning benchmarks.

What carries the argument. The key machinery is a shared 3D perception tool with two core components. The first, joint segmentation–depth co-refinement, consolidates duplicate masks and merges parts of occluded objects within a single frame, then applies geometry, depth-edge, and 3D outlier filters to remove projection artifacts. The second, cross-view instance fusion, associates observations of the same object across frames using voxel-overlap matching while preventing merges of distinct objects through a same-frame exclusion rule; it further links spatially disjoint fragments via a VLM-based visual identity judge. Gravity alignment via GeoCalib ensures the scene has a consistent up direction, making all geometric computations reference-frame independent.

Core claim. The core discovery is that the accuracy of 3D perception is the decisive bottleneck in agentic spatial understanding on object-centric tasks. The oracle experiment (Table 1) demonstrates that with perfect 3D bounding boxes, the agent's reasoning tools are nearly sufficient (98.23%), while with real perception tools the same reasoning pipeline performs poorly (50.38%). This isolates the problem as one of reconciling fragmented and inconsistent perception outputs rather than one of reasoning complexity. DeSpatial's architecture embodies this insight by explicitly separating perception and reasoning: a fixed, shared perception tool constructs a reliable scene representation, and separate reasoning tools operate on it without needing to reconcile intermediate perception outputs themselves.

The premise it rests on. The oracle experiment only covers five object-centric tasks whose annotations are designed to align with the ground-truth boxes, so it does not prove that perception is the only bottleneck for all spatial tasks, such as room-size estimation or navigation.

If the paper is right (Pith's inferences, not the authors')

Spatial agent designs should prioritize improving 3D perception quality over building more elaborate reasoning or tool-calling strategies.

A unified, metric, gravity-aligned 3D scene representation enables simpler reasoning tools (counting, distance, direction) to achieve near-perfect accuracy, reducing the need for complex multi-step inference.

The oracle diagnostic can be applied to other spatial tasks (e.g., room size, route planning) to reveal whether perception or reasoning is the limiting factor in each case.

The modular perception–reasoning split allows each component to be improved independently, potentially accelerating progress in embodied AI and robotics.

If the perception bottleneck is universal, then significant gains in spatial AI could come from improving depth estimation, segmentation, and multi-view fusion, rather than from scaling language models or reasoning chain lengths.

The paper's hyperparameters (voxel size 0.1 m, overlap threshold 0.1, angle threshold 20°, etc.) were tuned on the same benchmarks without a clearly separated validation set; their generalizability to new indoor or outdoor scenes remains untested.

For non-object-centric tasks (e.g., room-size estimation, path planning), the oracle experiment was not conducted, so it is plausible that reasoning limitations also contribute there, contrary to the paper's broad claim about the 'devil' lying solely in 3D.

The visual identity judge depends on a VLM, which introduces computational cost and potential hallucination; a learned geometric appearance model might be more reliable and faster.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Spatial agents have access to reconstruction and segmentation tools, yet inconsistent predictions and fragmented object identities can prevent their outputs from becoming reliable 3D scene information. A simple oracle experiment exposes this perception bottleneck: replacing tool-predicted perception with ground-truth 3D bounding boxes raises an agent to nearly full score across five object-centric VSI-Bench tasks, while retaining the same reasoning tools. Motivated by this finding, we believe the devil of spatial reasoning lies in 3D itself, and introduce DeSpatial(Devil of Spatial), which explicitly decomposes spatial understanding into perception and reasoning. DeSpatial has a shared 3D perception tool to reconcile independent perception outputs into a unified, metric, gravity-aligned 3D scene. This perception tool has two core components: a) a joint segmentation–depth co-refinement module that consolidates masks of the same object and removes projection artifacts caused by inconsistent segmentation and depth within a single frame; b) a cross-view instance fusion module that consolidate voxel-overlapping instances but not allowed to merge same-frame instances. This fusion module additionally fuses spatially disjoint fragments caused by insufficient multi-view observations, by checking their 3D structure and source mask appearance. After receiving the complete and unified 3D scene information, the agent chooses reasoning or computational tools to solve the spatial problem. DeSpatial achieves leading performance on VSI-Bench, MindCube-Tiny, and ViewSpatial, with additional long-video results on VSI-SUPER, revealing the bottleneck of spatial understanding lies in a precise 3D perception itself.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-16

The authors' main claim, as the classifier picked it out 0.88

“DeSpatial achieves leading performance on VSI-Bench, MindCube-Tiny, and ViewSpatial, with additional long-video results on VSI-SUPER, revealing the bottleneck of spatial understanding lies in a precise 3D perception itself.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.17 · human participants 0.02 · controlled comparison 0.78 · randomized assignment 0.02 · states quantitative result 0.36 · compares with prior results 0.72 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Spatial agents have access to reconstruction and segmentation tools, yet inconsistent predictions and fragmented object identities can prevent their outputs from becoming reliable 3D scene information. background 0.86 0.11 0.08 asserted 0.56
A simple oracle experiment exposes this perception bottleneck: replacing tool-predicted perception with ground-truth 3D bounding boxes raises an agent to nearly full score across five object-centric VSI-Bench tasks, while retaining the same reasoning tools. result 0.93 0.86 0.21 asserted 1.00
Motivated by this finding, we believe the devil of spatial reasoning lies in 3D itself, and introduce DeSpatial(Devil of Spatial), which explicitly decomposes spatial understanding into perception and reasoning. contribution 0.82 0.49 0.68 speculative 0.46
DeSpatial has a shared 3D perception tool to reconcile independent perception outputs into a unified, metric, gravity-aligned 3D scene. method 0.67 0.84 0.11 asserted 1.00
This perception tool has two core components: a) a joint segmentation--depth co-refinement module that consolidates masks of the same object and removes projection artifacts caused by inconsistent segmentation and depth within a single frame; b) a cross-view instance fusion module that consolidate voxel-overlapping instances but not allowed to merge same-frame instances. method 0.99 0.80 0.04 asserted 1.00
This fusion module additionally fuses spatially disjoint fragments caused by insufficient multi-view observations, by checking their 3D structure and source mask appearance. method 0.99 0.74 0.04 asserted 1.00
After receiving the complete and unified 3D scene information, the agent chooses reasoning or computational tools to solve the spatial problem. method 0.99 0.39 0.05 asserted 0.97
DeSpatial achieves leading performance on VSI-Bench, MindCube-Tiny, and ViewSpatial, with additional long-video results on VSI-SUPER, revealing the bottleneck of spatial understanding lies in a precise 3D perception itself. result 0.91 0.94 0.71 asserted 0.99

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-16: 1 request, 6,497 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. Overview of DeSpatial. The VLM receives the video and question and calls our shared perception tool. Joint segmentation–depth co-refinement cleans per-frame observations; cross-view instance fusion associates them into distinguished individual objects. Colors identify individual masks before fusion and persistent instances afterward. Gravity estimation establishes a common vertical direction. The VLM then reasons and answers using reasoning and computational tools
Figure 2 of the paper
Figure 2. Joint segmentation–depth co-refinement. (a) Same-class mask consolidation unifies highly overlapping masks and visually verified parts of an occluded object. Negative or uncertain identity judgments keep masks separate. (b) Confidence, depth-edge, and 3D outlier filters remove unreliable geometry. Component selection then removes residual detached points that distort the bounding box, retaining the top-k components for k confirmed object parts, usually one. where pt is the reconstructed world-coordinate point map, sR applies metric scaling and gravity alignment, and Vt is the set of…
Figure 3 of the paper
Figure 3. Cross-view instance fusion. (a) Voxel overlap links partial observations, while same￾frame exclusions prevent distinct objects from merging, including through indirect connections. (b) Visual fragment association handles observations without shared voxels. Multiple source views and a common 3D view provide complementary appearance and structural evidence to an identity judge. A positive decision combines the observed fragments and their provenance. To prevent distinct objects from merging, we impose same-frame exclusion on entire groups of observations. For a group G, let F(G) = {ti…
Figure 4 of the paper
Figure 4. Human annotation interface for gravity estimation. Both panels show the same scene and

Figures are the authors', reproduced from the paper.

References (44)
  1. [1] Shuai Bai, Yuxuan Cai, Ruizhe Chen, et al. Qwen3-VL technical report, 2025 a . URL https://arxiv.org/abs/2511.21631 on PithPith's reviewarXiv
  2. [2] Shuai Bai, Keqin Chen, Xuejing Liu, et al. Qwen2.5-VL technical report, 2025 b . URL https://arxiv.org/abs/2502.13923 on PithPith's reviewarXiv
  3. [3] Zhongang Cai, Ruisi Wang, Chenyang Gu, Fanyi Pu, Junxiang Xu, Yubo Wang, Wanqi Yin, Zhitao Yang, Chen Wei, Qingping Sun, Tongxi Zhou, Jiaqi Li, Hui En Pang, Oscar Qian, Yukun Wei, Zhiqian Lin, Xuanke Shi, Kewang Deng, Xiaoyang Han, Zukai Chen, Xiangyu Fan, Hanming Deng, Lewei ... on PitharXiv
  4. [4] Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R \"a dle, Tria... on PithPith's reviewarXiv
  5. [5] Zeren Chen, Xiaoya Lu, Zhijie Zheng, Pengrui Li, Lehan He, Yijin Zhou, Jing Shao, Bohan Zhuang, and Lu Sheng. Geometrically-constrained agent for spatial reasoning. arXiv preprint arXiv:2511.22659, 2025. URL https://arxiv.org/abs/2511.22659 on PitharXiv
  6. [6] Zhe Chen, Weiyun Wang, Yue Cao, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling, 2024. URL https://arxiv.org/abs/2412.05271 on PithPith's reviewarXiv
  7. [7] Yalun Dai, Hao Li, Shulin Tian, Runmao Yao, Yuhao Dong, Fangzhou Hong, Zhaoxi Chen, Fangfu Liu, Tao Wang, Kim-Hui Yap, and Ziwei Liu. S-Agent : Spatial tool-use elicits reasoning for spatial intelligence, 2026. URL https://arxiv.org/abs/2606.20515v3 on PithPith's reviewarXiv
  8. [8] Martin Ester, Hans-Peter Kriegel, J \"o rg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pp.\ 226--231, 1996. URL h... on Pith
  9. [9] Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Peihao Wang, Huaizhi Qu, Shijie Zhou, Dilin Wang, Zhicheng Yan, Hongyu Xu, Justin Theiss, Tianlong Chen, Jiachen Li, Zhengzhong Tu, Zhangyang Wang, and Rakesh Ranjan. Vlm-3r: Vision-language mo... on Pith
  10. [10] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24 0 (6): 0 381--395, 1981. doi:10.1145/358669.358692. URL https://www.sri.com/wp-content/up... on Pithdoi:10.1145/358669.358692
  11. [11] Gemini Team . Gemini 2.5 : Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities, 2025. URL https://arxiv.org/abs/2507.06261 on PithPith's reviewarXiv
  12. [12] Google DeepMind . Gemini 3 Pro model card, 2025. URL https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf on Pith
  13. [13] Yejie Guo, Yunzhong Hou, Wufei Ma, Meng Tang, and Ming-Hsuan Yang. Pursuing minimal sufficiency in spatial reasoning. arXiv preprint arXiv:2510.16688, 2025. URL https://arxiv.org/abs/2510.16688 on PitharXiv
  14. [14] Jixuan He, Xueting Li, Chieh Hubert Lin, and Ming-Hsuan Yang. Reasmory : 3D reconstruction as explicit memory for VLMs spatial reasoning, 2026. URL https://arxiv.org/abs/2606.00963 on PithPith's reviewarXiv
  15. [15] Hsiang-Wei Huang, Junbin Lu, Kuang-Ming Chen, Jianxu Shangguan, Cheng-Yen Yang, and Jenq-Neng Hwang. CaMo : Camera motion grounded evaluation and training for vision-language models, 2026. URL https://arxiv.org/abs/2605.20165 on PithPith's reviewarXiv
  16. [16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Doll \'a r, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vis... on Pith
  17. [17] Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. ViewSpatial-Bench : Evaluating multi-perspective spatial localization in vision-language models, 2025 a . URL htt... on PitharXiv
  18. [18] Haoyuan Li, Zhengdong Hu, Jun Wang, Hehe Fan, and Yi Yang. Skill-3d: Evolving scene-aware skills for agentic 3d spatial reasoning, 2026. URL https://arxiv.org/abs/2606.07436 on PithPith's reviewarXiv
  19. [19] Hongxing Li, Dingming Li, Zixuan Wang, Yuchen Yan, Hang Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu, Jun Xiao, and Yueting Zhuang. SpatialLadder : Progressive training for spatial reasoning in vision-language models, 2025 b . URL https://arxiv.org/abs/2510.08531 on PitharXiv
  20. [20] Pengteng Li, Pinhao Song, Wuyang Li, Huizai Yao, Weiyu Guo, Yijie Xu, Dugang Liu, and Hui Xiong. See&trek: Training-free spatial prompting for multimodal large language model. In Advances in Neural Information Processing Systems, volume 38, 2025 c . URL https://papers.neurips.... on Pith
show 24 more
  1. [21] Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth Anything 3 : Recovering the visual space from any views. arXiv preprint arXiv:2511.10647, 2025. URL https://arxiv.org/abs/2511.10647 on PithPith's reviewarXiv
  2. [22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO : Marrying DINO with grounded pre-training for open-set object detection. In Computer Vision -- ECCV 2024, pp.\ 38--55, ... on Pithdoi:10.1007/978-3-031-72970-6_3
  3. [23] Zhanpeng Luo, Ce Zhang, Silong Yong, Cunxi Dai, Qianwei Wang, Haoxi Ran, Guanya Shi, Katia Sycara, and Yaqi Xie. pySpatial : Generating 3d visual programs for zero-shot spatial reasoning. arXiv preprint arXiv:2603.00905, 2026. URL https://arxiv.org/abs/2603.00905 on PitharXiv
  4. [24] Chaofan Ma, Zhenjie Mao, Yuhuan Yang, Fanqin Zeng, Yue Shi, Yingjie Zhou, Xiaofeng Cao, and Jiangchao Yao. Reason, then re-reason: Cross-view revisiting improves spatial reasoning. In International Conference on Machine Learning, 2026. URL https://arxiv.org/abs/2606.11683 on PithPith's reviewarXiv
  5. [25] Damiano Marsili, Rohun Agrawal, Yisong Yue, and Georgia Gkioxari. Visual agentic AI for spatial reasoning with a dynamic API . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19446--19455, 2025. URL https://openaccess.thecvf.com/conte... on Pith
  6. [26] Tingshu Mou, Jiabo He, Renying Wang, Ce Liu, Hao Yang, Tiehua Zhang, Jingjing Chen, and Xingjun Ma. ViSRA : A video-based spatial reasoning agent for multi-modal large language models, 2026. URL https://arxiv.org/abs/2605.10106 on PithPith's reviewarXiv
  7. [27] OpenAI . GPT-4o system card, 2024. URL https://arxiv.org/abs/2410.21276 on PithPith's reviewarXiv
  8. [28] OpenAI . Introducing GPT-5.4 , 2026. URL https://openai.com/index/introducing-gpt-5-4/ on Pith
  9. [29] Shouwei Ruan, Bin Wang, Zhenyu Wu, Qihui Zhu, Yuxiang Zhang, Jingzhi Li, Yubin Wang, and Xingxing Wei. AlloSpatial : Agentic harness framework for spatial reasoning in foundation models, 2026. URL https://arxiv.org/abs/2606.08952 on PithPith's reviewarXiv
  10. [30] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, Yan Lu, Jenq-Neng Hwang, and Gaoang Wang. MovieChat : From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Co... on Pith
  11. [31] Shi-Yu Tian, Zhuo-Xia Wang, Xuan-Yi Zhu, Zhi Zhou, Xinwei Yang, Kun-Yang Yu, Ming Yang, Yang Chen, and Yu-Feng Li. Self-evolving neuro-symbolic skills for tool-augmented spatial reasoning, 2026. URL https://arxiv.org/abs/2608.07955 on PithPith's reviewarXiv
  12. [32] Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys. GeoCalib : Learning single-image calibration with geometric optimization. In European Conference on Computer Vision, 2024. URL https://arxiv.org/abs/2409.06704 on PithPith's reviewarXiv
  13. [33] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT : Visual geometry grounded transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5294--5306, 2025 a . URL https://openacce... on Pith
  14. [34] Qineng Wang, Baiqiao Yin, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, Saining Xie, Jiajun Wu, Fei-Fei Li, and Manling Li. MindCube : Spatial mental modeling from limited views, 2025 b . URL https://ar... on PitharXiv
  15. [35] Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. ^3 : Permutation-equivariant visual geometry learning. arXiv preprint arXiv:2507.13347, 2025 c . URL https://arxiv.org/abs/2507.13347 on PithPith's reviewarXiv
  16. [36] Jin Xu, Xiaojian Huang, Zhuodong Luo, Zhihong Zhang, Xin Liu, Jiansheng Wei, Xinzhi Wang, Jie Zhao, and Xuejin Chen. MV-STRIDE : Enabling MLLMs to master multi-view spatial reasoning via hierarchical capability modeling, 2026. URL https://arxiv.org/abs/2609.07258v1 on PithPith's reviewarXiv
  17. [37] Jihan Yang, Shusheng Yang, Anjali W. Gupta, Rilyn Han, Fei-Fei Li, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025 a . URL http... on PithPith's reviewarXiv
  18. [38] Shusheng Yang, Jihan Yang, Pinzhi Huang, Ellis Brown, Zihao Yang, Yue Yu, Shengbang Tong, Zihan Zheng, Yifan Xu, Muhan Wang, Daohan Lu, Rob Fergus, Yann LeCun, Fei-Fei Li, and Saining Xie. Cambrian-S : Towards spatial supersensing in video, 2025 b . URL https://arxiv.org/abs/2... on PithPith's reviewarXiv
  19. [39] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3D visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20623--20633, 2024. URL https://o... on Pith
  20. [40] Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-VStream : Memory-based real-time understanding for long video streams, 2024. URL https://arxiv.org/abs/2406.08085 on PithPith's reviewarXiv
  21. [41] Yiming Zhang, Jiacheng Chen, Jiaqi Tan, Yongsen Mao, Wenhu Chen, and Angel X. Chang. ReVSI : Rebuilding visual spatial intelligence evaluation for accurate assessment of VLM 3D reasoning, 2026 a . URL https://arxiv.org/abs/2604.24300 on PithPith's reviewarXiv
  22. [42] Zaibin Zhang, Yuhan Wu, Lianjie Jia, Yifan Wang, Zhongbo Zhang, Yijiang Li, Binghao Ran, Fuxi Zhang, Zhuohan Sun, Yizhuang Peng, Zhenfei Yin, Lijun Wang, and Huchuan Lu. Think3D : Thinking with space for spatial reasoning, 2026 b . URL https://arxiv.org/abs/2601.13029 on PithPith's reviewarXiv
  23. [43] Yang Zhou, Zixuan Huang, Sunzhu Li, Zhuo Yang, Chen Zhang, Shunian Chen, Caijun Yan, Jianyao Xu, Shunyu Liu, Weijie Fu, Peiliang Li, Xiaozhi Chen, and Yuxiang Cai. SpatialCLI : Learning to reason with spatial tools, then without them. arXiv preprint arXiv:2607.27703, 2026. URL... on PithPith's reviewarXiv
  24. [44] Jinguo Zhu, Weiyun Wang, Zhe Chen, et al. InternVL3 : Exploring advanced training and test-time recipes for open-source multimodal models, 2025. URL https://arxiv.org/abs/2504.10479 on PithPith's reviewarXiv
Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. Space-X: Self-Evolving Spatial Agents Weaving Clues into Evidence Chains cs.AI · Nov 2026

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment