Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Video spatial reasoning improves when a model uses geometric consistency to organize memory and to keep cross-view answers stable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:31 UTC pith:SQOPUHA7

load-bearing objection Solid architecture, strong ablations, but the VSI-Bench result hinges on an unaddressed train/test overlap question. the 3 major comments →

arxiv 2607.17599 v1 pith:SQOPUHA7 submitted 2026-07-20 cs.CV

ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning

Ting Huang , Zhenyu Zhang , Wenyuan Huang , Jian Yang , Hao Tang This is my paper
classification cs.CV
keywords video spatial reasoninggeometric consistencymultimodal large language modelsgeometry-consistent memoryself-supervised reinforcement learningviewpoint stabilitycamera pose and depthlong-video understanding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Video spatial reasoning—inferring where objects are, how far apart, and how to move between them—remains weak in current video-language models, which tend to read frames as independent semantic images. The paper proposes ConsiSpace, which treats geometric consistency as both a memory-management rule and a learning goal: predicted camera pose and depth decide which frames to store and fuse, and a post-training stage rewards the model for giving the same answer, metric, and relational structure from different views. On three spatial-reasoning benchmarks (VSI-Bench, OSI-Bench, MMSI-Video-Bench), ConsiSpace reports an average gain of 12.6 points over the strongest baselines, with bigger gains on metric and viewpoint-heavy tasks. The paper's claim is that consistency across viewpoints, not more semantic data, is what unlocks reliable video spatial reasoning.

Core claim

The paper argues that the bottleneck in video spatial reasoning is not recognition but aggregating consistent spatial evidence across redundant, viewpoint-varying frames. To that end, geometric consistency becomes the memory's organizing principle: camera pose and depth gate which frames are written, which stored entries are fused, and which frames are retrieved (semantic relevance plus direction consistency). A post-training stage, UC-SSRL, samples two views of the same question and rewards agreement in answer distributions, metric estimates, and topological relations. The authors report consistent gains on three benchmarks, with the largest on metric- and viewpoint-heavy tasks.

What carries the argument

The central machinery is the Geometry-Consistent Memory (GCM) plus the Unified Consistency SSRL (UC-SSRL). GCM stores per-frame evidence tokens in an implicit bank and camera pose/depth in an explicit bank; the explicit bank drives geometry-gated writing (skip frames with little motion/viewpoint change), geometry-consistent fusion (merge entries that are geometrically near and semantically similar), and geometry-filtered retrieval (semantic chunk recall, then frame selection with a direction-consistency term). UC-SSRL samples two views of the same question and uses a reward mixing answer KL, metric absolute difference, and topology KL, optimized with policy-gradient on LoRA-style parameters.

Load-bearing premise

The results hinge on the indoor training set VSI-590K and the indoor evaluation set VSI-Bench not sharing videos; the paper asserts disjointness only for the outdoor nuScenes-10K versus OSI-Bench, not for the indoor pair.

What would settle it

Compare ConsiSpace trained on VSI-590K with all VSI-Bench videos removed, or directly measure video/scene overlap between the two collections; if removing overlapping videos erases the reported VSI gain, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is right, video spatial reasoning gains do not require larger models or more annotated spatial data; they can come from reorganizing evidence around geometry and from self-supervised consistency signals.
  • Long-video efficiency improves sharply: at 200 frames, ConsiSpace uses 35.0 GB and 3.10 s versus 78.5 GB and 10.80 s for a strong dense-evidence baseline, while scoring higher on all three benchmarks (Table 4).
  • The gains transfer to navigation-oriented perception: on Habitat ObjectNav next-waypoint prediction, ConsiSpace with UC-SSRL reaches 50.8% direction accuracy versus 42.6% for the base model (Table 5).
  • Predicted geometry does not need to be perfect: adding 30% noise to geometry cues degrades performance gradually, and ConsiSpace stays above a geometry-aware baseline (Table 8).
  • Each consistency reward contributes: ablations show answer, metric, and topological rewards are complementary, and the full objective gives the best scores (Table 7).

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the geometry-consistency principle could be lifted to other domains where viewpoint invariance is the ground-truth signal, such as object tracking, 3D reconstruction, or embodied exploration; the paper's memory design suggests a generic recipe rather than a video-only fix.
  • Editorial inference: the reported VSI gain may depend on the train/eval split between VSI-590K and VSI-Bench; the paper only verifies disjointness for the outdoor nuScenes-10K versus OSI-Bench, not for the indoor pair, so the VSI numbers should be read with that caveat.
  • Editorial inference: replacing the fixed write/fusion thresholds with a learned, scene-adaptive gating policy is a natural next step; the sensitivity analysis shows the method is stable around the defaults, which leaves room to test whether adaptive thresholds improve further.
  • Editorial inference: because UC-SSRL needs only paired views of the same scene, it could be applied to unlabeled or weakly labeled video streams for continual self-improvement, something the authors do not explore.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The manuscript proposes ConsiSpace, a video spatial reasoning framework built on Qwen3-VL-8B-Instruct with a frozen SigLIP2 visual encoder and a frozen VGGT geometry encoder. Spatial evidence is managed by a Geometry-Consistent Memory that gates writes by camera motion, fuses geometrically consistent entries, and retrieves query-relevant chunks/frames with a direction-consistency term. After SFT, Unified Consistency SSRL optimizes answer, metric, and topological consistency across two sampled views using only LoRA parameters. Experiments on VSI-Bench, OSI-Bench, and MMSI-Video-Bench report large gains over prior specialized models, with the abstract claiming an average improvement of 12.6 points, together with efficiency and ablation analyses.

Significance. If the results hold, the contribution is a practical and well-motivated mechanism for making video spatial reasoning both more compact and more viewpoint-stable. The ablations are thorough: the three memory operations are tested individually, threshold sensitivity is explored, noisy geometry robustness is measured, and the UC-SSRL reward components are decomposed. A particular strength is that the paper reports entry counts and efficiency numbers, so the memory mechanism's benefit is not only measured by accuracy. The principal risk to the paper's central claim is not architectural but empirical: the indoor training set VSI-590K may overlap with the VSI-Bench evaluation videos, and this is not addressed. The advertised average gain also does not exactly match the tables. These issues need to be resolved before the headline results can be certified.

major comments (3)
  1. [§4.1, §4.2, Supp. A.3] The manuscript never states that VSI-590K training videos are disjoint from VSI-Bench evaluation videos. Section 4.2 says VSI-590K is the primary indoor instruction-tuning source; Section 4.1 reports on VSI-Bench. The only explicit disjointness statement in the paper is for nuScenes-10K vs OSI-Bench (Section 4.2 and Supp. A.3). VSI-590K is attributed to the Cambrian-S paper [46], and its name suggests a direct relationship with VSI-Bench. If VSI-590K contains VSI-Bench source videos, the +7.0 VSI gain and the 12.6-point average are inflated. Please provide a video-level overlap analysis and, if overlap exists, re-train/evaluate on a held-out subset.
  2. [§4.3, §B.2] SFT is said to use VSI-590K and nuScenes-10K, but the training configuration—8 GPUs, per-device batch size 1, gradient accumulation 2, 200 steps—corresponds to only 3,200 optimizer updates' worth of samples. That is under 1% of VSI-590K alone, so it cannot represent instruction-tuning on that dataset as described. The same step count is reused for UC-SSRL. Please state the total number of examples seen (or epochs), or correct the step count; otherwise the training setup is not reproducible and the reported gains cannot be connected to the claimed data.
  3. [Abstract, §4.4, Tables 1–3] The abstract claims an average gain of 12.6 points over the strongest baselines. From the tables, the deltas are +7.0 on VSI-Bench (Table 1), +12.7 on OSI-Bench (Table 2), and +14.9/+15.0 on MMSI-Video-Bench under the two settings (Table 3). These values average to 12.4 (four settings) or about 11.6 (one MMSI setting per benchmark). Please clarify how 12.6 was computed; this is a claim-level reporting inconsistency.
minor comments (4)
  1. [§3.4, §B.4] The per-reward weights α, η, β in Eqs. (14)–(16) are not reported; only the overall UC-SSRL weight is given. Without values, the reward balance is not reproducible.
  2. [§3.3, Eq. (9) vs §3.4, Eq. (14)] α and β denote retrieval weights in Eq. (9) and reward weights in Eqs. (14)/(16); the reuse of α (and β) creates ambiguity. Rename one set.
  3. [§4.5, Table 4] ConsiSpace accuracy drops as the frame budget grows (e.g., VSI 76.6→71.6→72.4; OSI 53.0→45.8→46.8; MMSI-SC 57.5→53.1→54.4). Since this section argues long-video scaling, please explain why additional frames do not help the proposed method.
  4. [§3.3] The sentence 'In practice, geometry filtering can be implemented by downweighting inconsistent entries during scoring or by masking them before concatenation' is vague; specify which implementation is used in the main experiments.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, or forces. The free parameters are hyperparameters of the memory and RL stages, several of which are not fully reported. The most important domain assumption is the invariance of spatial answers under viewpoint changes, which is the foundation of the UC-SSRL objective. The potential data overlap between VSI-590K and VSI-Bench is an unstated selection assumption and is tracked as a red flag.

free parameters (6)
  • Geometry-gated writing thresholds (τ_p, τ_θ) = 0.15 m, 15°
    Control when new frames are written to memory; chosen by hand and validated only by a small sensitivity sweep in Table 10.
  • Fusion neighborhood thresholds (r, θ) = 0.40 m, 30°
    Define which entries are considered geometrically consistent for fusion; chosen by hand and tested in Table 11.
  • UC-SSRL overall weight = 1×10^-2
    Scalar controlling the reinforcement objective during post-SFT training; set without a detailed search.
  • Reward combination weights α, η, β = not reported
    Weights of answer, metric, and topology consistency rewards in Equations 14–17 are not specified in the paper.
  • Retrieval top-K (chunk and frame) = 8 and 8
    Number of retrieved chunks and frames; chosen as a design decision without a reported search.
  • Memory token count and chunk count = M=64, n_chunks=4
    Memory capacity hyperparameters; no ablation on these values is reported.
axioms (4)
  • domain assumption Correct spatial relations are invariant under smooth viewpoint changes of a static scene (Section 1).
    The whole consistency-training objective assumes that two views of the same scene and same question should yield the same answer. This fails for egocentric questions whose reference frame is tied to the current camera pose, and for dynamic scenes where object relations change over time.
  • domain assumption VGGT provides sufficiently accurate camera pose and depth for memory gating, fusion, and retrieval.
    All geometry-guided operations depend on VGGT outputs; the robustness test in Table 8 suggests some tolerance to noise, but the method still assumes the cues are usable.
  • domain assumption Standard benchmark protocols (VSI-Bench, OSI-Bench, MMSI-Video-Bench) are accepted as ground truth.
    The paper evaluates against benchmark metrics without questioning the benchmarks themselves; this is standard practice.
  • standard math Cosine similarity and KL divergence are appropriate for semantic and distributional alignment.
    Used in memory retrieval and consistency rewards; these are standard mathematical tools.

pith-pipeline@v1.3.0-alltime-deepseek · 20402 in / 12518 out tokens · 111995 ms · 2026-08-01T17:31:28.787333+00:00 · methodology

0 comments
read the original abstract

Video spatial reasoning is essential for navigation-oriented perception and long-video question answering, where models must infer spatial relations across long horizons under changing viewpoints. However, existing multimodal large language models (MLLMs) remain largely semantic-centric, and often fail to reliably aggregate consistent spatial evidence from redundant video observations, leading to inefficient or unstable reasoning. To address these issues, we propose ConsiSpace, a geometry-consistency-aware framework for geometry-sensitive video spatial reasoning that turns spatial consistency into both an evidence organization principle and an explicit post-SFT learning signal. We build a geometry-consistent memory (GCM) including implicit evidence tokens and explicit geometric cues, and leverage efficient organization strategies to compactly preserve task-related spatial evidence. Furthermore, we utilize unified consistency self-supervised reinforcement learning (UC-SSRL) after supervised fine-tuning to improve cross-view stability, with answer-, metric-, and topology-consistency rewards. Extensive experiments on three spatial-reasoning benchmarks, VSI-Bench, OSI-Bench, and MMSI-Video-Bench, show consistent gains, improving the average score by 12.6 points over the strongest baselines.

Figures

Figures reproduced from arXiv: 2607.17599 by Hao Tang, Jian Yang, Ting Huang, Wenyuan Huang, Zhenyu Zhang.

Figure 1
Figure 1. Figure 1: ConsiSpace learning geometric consistency for video spatial reason [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ConsiSpace. ConsiSpace organizes visual-spatial tokens and pose-depth cues with a geometry-consistent memory, retrieves query-relevant evidence for answer generation, and uses UC-SSRL to improve viewpoint stability. priors are mostly used as input features or training signals, while their role in long-video evidence management remains less explored. We instead treat geom￾etry as a consistency c… view at source ↗
Figure 3
Figure 3. Figure 3: Geometry-Consistent Memory (GCM) evidence lifecycle. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative effect of geometry￾gated writing and fusion. Dense updates introduce redundant points, while GCM produces cleaner and more compact point clouds. Red numbers denote normalized re￾tained entries [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative visualization of geometry-filtered retrieval. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Consistency under revisit views. Given the same query, baseline models produce inconsistent counting answers across two revisited viewpoints, while ConsiS￾pace with UC-SSRL gives consistent predictions. and View B. This suggests that UC-SSRL refines viewpoint stability after SFT, rather than learning task correctness from consistency alone. Additional visualizations. We provide more qualitative examples on… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative results on VSI-Bench [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative results on MMSI-Video-Bench. [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative results on OSI-Bench [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 18 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:1903.11027 (2019)

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)

    Pith/arXiv arXiv 1903

  3. [3]

    In: The Thirteenth International Conference on Learning Represen- tations (2025)

    Chen, G., Liu, Y., Huang, Y., Pei, B., Xu, J., He, Y., Lu, T., Wang, Y., Wang, L.: CG-bench: Clue-grounded question answering benchmark for long video un- derstanding. In: The Thirteenth International Conference on Learning Represen- tations (2025)

    2025

  4. [4]

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

    Chen, Y., Qi, Z., Zhang, W., Jin, X., Zhang, L., Liu, P.: Reasoning in space via grounding in the world. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  5. [5]

    In: International Conference on Learning Representations

    Chen, Y., Xue, F., Li, D., Hu, Q., Zhu, L., Li, X., Fang, Y., Tang, H., Yang, S., Liu, Z., et al.: Longvila: Scaling long-context visual language models for long videos. In: International Conference on Learning Representations. vol. 2025, pp. 18227–18246 (2025)

    2025

  6. [6]

    Dongfang, Z., Zheng, X., Weng, Z., Lyu, Y., Paudel, D.P., Van Gool, L., Yang, K., Hu, X.: Are multimodal large language models ready for omnidirectional spatial reasoning? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9759–9769 (2026)

    2026

  7. [7]

    arXiv preprint arXiv:2505.20279 (2025)

    Fan, Z., Zhang, J., Li, R., Zhang, J., Chen, R., Hu, H., Wang, K., Qu, H., Wang, D., Yan, Z., et al.: Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. arXiv preprint arXiv:2505.20279 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    arXiv preprint arXiv:2601.11442 (2026)

    Gao, X., Zhang, Z., Chen, D.Z., Xu, S., Quan, L., Pérez-Pellitero, E., Jang, Y.: Map2thought: Explicit 3d spatial reasoning via metric cognitive maps. arXiv preprint arXiv:2601.11442 (2026)

    arXiv 2026

  9. [9]

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

    Hu, W., Lin, J., Long, Y., Ran, Y., Jiang, L., Wang, Y., Zhu, C., Xu, R., Wang, T., Pang, J.: G2vlm: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9535–9546 (2026)

    2026

  10. [10]

    arXiv preprint arXiv:2511.17889 (2025)

    Huang, T., Li, D., Yang, R., Zhang, Z., Yang, Z., Tang, H.: Mobilevla-r1: Rein- forcing vision-language-action for mobile robots. arXiv preprint arXiv:2511.17889 (2025)

    arXiv 2025

  11. [11]

    arXiv preprint arXiv:2507.23478 (2025)

    Huang, T., Zhang, Z., Tang, H.: 3d-r1: Enhancing reasoning in 3d vlms for unified scene understanding. arXiv preprint arXiv:2507.23478 (2025)

    Pith/arXiv arXiv 2025

  12. [12]

    In: Thirteenth International Conference on 3D Vision (2026)

    Huang, T., Zhang, Z., Wang, Y., Tang, H.: 3d coca: Contrastive learners are 3d captioners. In: Thirteenth International Conference on 3D Vision (2026)

    2026

  13. [13]

    arXiv preprint arXiv:2505.15232 (2025)

    Huang, T., Zhang, Z., Zhang, R., Zhao, Y.: Dc-scene: Data-centric learning for 3d scene understanding. arXiv preprint arXiv:2505.15232 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    arXiv preprint arXiv:2506.01946 (2025)

    Huang, X., Wu, J., Xie, Q., Han, K.: 3drs: Mllms need 3d-aware representation supervision for scene understanding. arXiv preprint arXiv:2506.01946 (2025)

    arXiv 2025

  15. [15]

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

    Ji, Y., Tan, H., Shi, J., Hao, X., Zhang, Y., Zhang, H., Wang, P., Zhao, M., Mu, Y., An, P., et al.: Robobrain: A unified brain model for robotic manipulation from abstract to concrete. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1724–1734 (2025)

    2025

  16. [16]

    In: European Confer- ence on Computer Vision

    Krantz, J., Wijmans, E., Majumdar, A., Batra, D., Lee, S.: Beyond the nav-graph: Vision-and-language navigation in continuous environments. In: European Confer- ence on Computer Vision. pp. 104–120. Springer (2020) ConsiSpace 17

    2020

  17. [17]

    In: European conference on computer vision

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: European conference on computer vision. pp. 71–91. Springer (2024)

    2024

  18. [18]

    arXiv preprint arXiv:2505.21500 (2025)

    Li, D., Li, H., Wang, Z., Yan, Y., Zhang, H., Chen, S., Hou, G., Jiang, S., Zhang, W., Shen, Y., et al.: Viewspatial-bench: Evaluating multi-perspective spatial local- ization in vision-language models. arXiv preprint arXiv:2505.21500 (2025)

    arXiv 2025

  19. [19]

    arXiv preprint arXiv:2510.12276 (2025)

    Li, F., Song, W., Zhao, H., Wang, J., Ding, P., Wang, D., Zeng, L., Li, H.: Spatial forcing: Implicit spatial representation alignment for vision-language-action model. arXiv preprint arXiv:2510.12276 (2025)

    arXiv 2025

  20. [20]

    arXiv preprint arXiv:2602.06037 (2026)

    Li, H., Cao, Q., Tang, T., Xiang, K., Guo, Z., Han, J., Xu, H., Liang, X.: Thinking with geometry: Active geometry integration for spatial reasoning. arXiv preprint arXiv:2602.06037 (2026)

    Pith/arXiv arXiv 2026

  21. [21]

    arXiv preprint arXiv:2510.08531 (2025)

    Li, H., Li, D., Wang, Z., Yan, Y., Wu, H., Zhang, W., Shen, Y., Lu, W., Xiao, J., Zhuang, Y.: Spatialladder: Progressive training for spatial reasoning in vision- language models. arXiv preprint arXiv:2510.08531 (2025)

    arXiv 2025

  22. [22]

    arXiv preprint arXiv:2511.10647 (2025)

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    arXiv preprint arXiv:2512.10863 (2025)

    Lin, J., Xu, R., Zhu, S., Yang, S., Cao, P., Ran, Y., Hu, M., Zhu, C., Xie, Y., Long, Y., et al.: Mmsi-video-bench: A holistic benchmark for video-based spatial intelligence. arXiv preprint arXiv:2512.10863 (2025)

    arXiv 2025

  24. [24]

    arXiv preprint arXiv:2509.10884 (2025)

    Liu, Q., Huang, T., Zhang, Z., Tang, H.: Nav-r1: Reasoning and navigation in embodied scenes. arXiv preprint arXiv:2509.10884 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    arXiv preprint arXiv:2510.27606 (2025)

    Liu, Y., Zhang, B., Zang, Y., Cao, Y., Xing, L., Dong, X., Duan, H., Lin, D., Wang, J.: Spatial-ssrl: Enhancing spatial understanding via self-supervised reinforcement learning. arXiv preprint arXiv:2510.27606 (2025)

    arXiv 2025

  26. [26]

    arXiv preprint arXiv:2504.01805 (2025)

    Ouyang, K., Liu, Y., Wu, H., Liu, Y., Zhou, H., Zhou, J., Meng, F., Sun, X.: Spacer: Reinforcing mllms in video spatial reasoning. arXiv preprint arXiv:2504.01805 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    arXiv preprint arXiv:2109.08238 (2021)

    Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., et al.: Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for em- bodied ai. arXiv preprint arXiv:2109.08238 (2021)

    Pith/arXiv arXiv 2021

  28. [28]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al.: Habitat: A platform for embodied ai research. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9339–9347 (2019)

    2019

  29. [29]

    arXiv preprint arXiv:2601.03267 (2025)

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025)

    Pith/arXiv arXiv 2025

  30. [30]

    arXiv preprint arXiv:2601.06496 (2026)

    Tang, H., Huang, T., Zhang, Z.: 3d coca v2: Contrastive learners with test-time search for generalizable spatial intelligence. arXiv preprint arXiv:2601.06496 (2026)

    arXiv 2026

  31. [31]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., Millican, K., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    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: Multilingual vision-language encoders with improved semantic understanding, lo- calization, and dense features. arXiv preprint arXiv:2502.14786 (2025)

    Pith/arXiv arXiv 2025

  33. [33]

    arXiv preprint arXiv:2604.05695 (2026) 18 T

    Wang, C., Huang, T., Sun, C., Ning, X., Wang, D., Tang, H.: Let geometry guide: Layer-wise unrolling of geometric priors in multimodal llms. arXiv preprint arXiv:2604.05695 (2026) 18 T. Huang et al

    Pith/arXiv arXiv 2026

  34. [34]

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

    Wang, H., Zhao, Y., Wang, T., Fan, H., Zhang, X., Zhang, Z.: Ross3d: Reconstruc- tive visual instruction tuning with 3d-awareness. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9275–9286 (2025)

    2025

  35. [35]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025)

    2025

  36. [36]

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

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20697–20709 (2024)

    2024

  37. [37]

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

    Wang,W.,He,Z.,Hong,W.,Cheng,Y.,Zhang,X.,Qi,J.,Ding,M.,Gu,X.,Huang, S., Xu, B., et al.: Lvbench: An extreme long video understanding benchmark. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22958–22967 (2025)

    2025

  38. [38]

    In: European Conference on Computer Vision

    Wang, X., Zhang, Y., Zohar, O., Yeung-Levy, S.: Videoagent: Long-form video understanding with large language model as agent. In: European Conference on Computer Vision. pp. 58–76. Springer (2024)

    2024

  39. [39]

    arXiv preprint arXiv:2603.25399 (2026)

    Wang, X., Wang, C., Xu, Y., Ye, M., Zhang, F.C., Tian, J., Zhan, X., Zhu, L., Lu, C., Yang, L.: Lamp: Learning vision-language-action policies with 3d scene flow as latent motion prior. arXiv preprint arXiv:2603.25399 (2026)

    Pith/arXiv arXiv 2026

  40. [40]

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

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π 3: Permutation-equivariant visual geometry learning. In: The Fourteenth International Conference on Learning Representations (2026)

    2026

  41. [41]

    arXiv preprint arXiv:2512.16561 (2025)

    Wang, Y., Ke, L., Zhang, B., Qu, T., Yu, H., Huang, Z., Yu, M., Xu, D., Yu, D.: N3d-vlm: Native 3d grounding enables accurate spatial reasoning in vision- language models. arXiv preprint arXiv:2512.16561 (2025)

    arXiv 2025

  42. [42]

    arXiv preprint arXiv:2505.23747 (2025)

    Wu, D., Liu, F., Hung, Y.H., Duan, Y.: Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence. arXiv preprint arXiv:2505.23747 (2025)

    Pith/arXiv arXiv 2025

  43. [43]

    arXiv preprint arXiv:2506.09965 (2025)

    Wu, J., Guan, J., Feng, K., Liu, Q., Wu, S., Wang, L., Wu, W., Tan, T.: Reinforcing spatial reasoning in vision-language models with interwoven thinking and visual drawing. arXiv preprint arXiv:2506.09965 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    arXiv preprint arXiv:2512.19683 (2025)

    Wu, M., Wang, Z., Wang, F., Yang, J., Pollefeys, M., Zhang, T.: From indoor to open world: Revealing the spatial reasoning gap in mllms. arXiv preprint arXiv:2512.19683 (2025)

    arXiv 2025

  45. [45]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    Yang, J., Yang, S., Gupta, A.W., Han, R., Fei-Fei, L., Xie, S.: Thinking in space: How multimodal large language models see, remember, and recall spaces. In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 10632– 10643 (2025)

    2025

  46. [46]

    arXiv preprint arXiv:2511.04670 (2025)

    Yang, S., Yang, J., Huang, P., Brown, E., Yang, Z., Yu, Y., Tong, S., Zheng, Z., Xu, Y., Wang, M., et al.: Cambrian-s: Towards spatial supersensing in video. arXiv preprint arXiv:2511.04670 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    Advances in Neural Information Processing Systems 38(2026)

    Zhang, J., Chen, Y., Xu, Y., Huang, Z., Mei, J., Chen, C., Zhou, Y., Yuan, Y.J., Cai, X., Huang, G., et al.: From flatland to space: Teaching vision-language models to perceive and reason in 3d. Advances in Neural Information Processing Systems 38(2026)

    2026

  48. [48]

    Expert Systems with Applications p

    Zhang, L., Wan, W., Huang, T., Fang, Z., Zhang, H.: Multigranularity-3dqa: Dy- namic multi-granularity perception and gated dual-stage encoder-decoder for 3d question answering. Expert Systems with Applications p. 131676 (2026)

    2026

  49. [49]

    arXiv preprint arXiv:2601.13029 (2026) ConsiSpace 19

    Zhang, Z., Wu, Y., Jia, L., Wang, Y., Zhang, Z., Li, Y., Ran, B., Zhang, F., Sun, Z., Yin, Z., et al.: Think3d: Thinking with space for spatial reasoning. arXiv preprint arXiv:2601.13029 (2026) ConsiSpace 19

    arXiv 2026

  50. [50]

    arXiv preprint arXiv:2511.23075 (2025)

    Zhao, R., Zhang, Z., Xu, J., Chang, J., Chen, D., Li, L., Sun, W., Wei, Z.: Space- mind: Camera-guided modality fusion for spatial reasoning in vision-language mod- els. arXiv preprint arXiv:2511.23075 (2025)

    arXiv 2025

  51. [51]

    arXiv preprint arXiv:2505.24625 (2025)

    Zheng, D., Huang, S., Li, Y., Wang, L.: Learning from videos for 3d world: En- hancing mllms with 3d vision geometry priors. arXiv preprint arXiv:2505.24625 (2025)

    arXiv 2025

  52. [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

  53. [53]

    In: Proceedings of the Computer Vision and Pattern Recognition Confer- ence

    Zhi, H., Chen, P., Li, J., Ma, S., Sun, X., Xiang, T., Lei, Y., Tan, M., Gan, C.: Lscenellm: Enhancing large 3d scene understanding using adaptive visual prefer- ences. In: Proceedings of the Computer Vision and Pattern Recognition Confer- ence. pp. 3761–3771 (2025) 20 T. Huang et al. A Outdoor Data Synthesis This section describes our automated pipeline ...

    2025