pith. sign in

arxiv: 2606.22476 · v1 · pith:J3OFR5D6new · submitted 2026-06-21 · 💻 cs.CV

CVSBench: A Comprehensive Benchmark for Cross-view Spatial Reasoning and Dreaming

Ruixun Liu , Lingyu Zhang , Lanxuan Xue , Kaiyu Li , Bowen Fu , Xiangyong Cao This is my paper

Pith reviewed 2026-06-26 10:52 UTC · model grok-4.3

classification 💻 cs.CV
keywords cross-view spatial reasoningvision-language modelsbenchmarksatellite-street pairs3D scene imaginationviewpoint variationspatial cognitionvisual spatial representations
0
0 comments X

The pith

VLMs struggle with cross-view spatial reasoning but gain substantially from 3D scene imagination.

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

The paper creates CVSBench to test whether vision-language models can reason spatially across extreme viewpoint shifts, using satellite and street-level image pairs. It shows that current models lose object-level and layout consistency under these changes. Two families of fixes are compared: language-only reasoning and methods that add visual spatial imagination. The results indicate that language-only methods give only small gains, while feeding in 3D-reconstructed scenes produces clear improvements. This points to the requirement for explicit visual-spatial representations if VLMs are to approach human-like cross-view spatial cognition.

Core claim

Advanced VLMs fail to maintain object-level and layout consistency when reasoning across satellite-street viewpoint changes. Language-only reasoning yields only marginal improvements, whereas a 3D scene imagination pipeline that supplies visual spatial representations produces substantial gains on cross-view VQA, grounding, and viewpoint identification tasks. The work therefore concludes that explicit visual-spatial representations are necessary for robust spatial cognition in these models.

What carries the argument

The 3D scene imagination pipeline that converts input views into explicit visual spatial representations for downstream reasoning.

If this is right

  • Language-only reasoning is insufficient for maintaining consistency across drastic viewpoint changes.
  • Explicit visual-spatial representations improve performance on object-level and layout tasks.
  • Spatially grounded reasoning and cognitive map inputs are promising directions for closing the gap with human performance.
  • The benchmark enables controlled, systematic measurement of progress on cross-view abilities.

Where Pith is reading between the lines

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

  • The same 3D imagination step could be tested on other multi-view domains such as indoor navigation or robotics.
  • Models that integrate 3D reconstruction as a default module may show broader gains on any task requiring viewpoint invariance.
  • Repeated use of the benchmark over time would track whether the field is moving toward visual-spatial representations.

Load-bearing premise

The satellite-street scene pairs and tasks in CVSBench form a valid and unbiased measure of cross-view spatial reasoning abilities.

What would settle it

A VLM that reaches high accuracy on the benchmark's cross-view VQA and grounding tasks without using any 3D scene imagination or other visual spatial input would falsify the necessity claim.

Figures

Figures reproduced from arXiv: 2606.22476 by Bowen Fu, Kaiyu Li, Lanxuan Xue, Lingyu Zhang, Ruixun Liu, Xiangyong Cao.

Figure 1
Figure 1. Figure 1: Overview of CVSBench. The left panel reviews representative tasks from ex￾isting spatial benchmarks suffering from several limitations [22,47,63]. The right panel displays our benchmark specifications, a radar chart comparing model performance, and core tasks with sample questions, demonstrating the superiority of our benchmark in task diversity, scale, and difficulty. The indices (e.g., A–E) shown after e… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset annotation pipeline (Red: ground truth (GT)). VQA: Multi-type prompts covering diverse question formats and instruction styles. Viewpoint Lo￾calization: Red arrows denote camera poses. Grounding: Red bboxes, arrows, and dots represent objects, viewpoints, and camera locations. currently lack the capacity for “urban-scale” cognition [25,65,70], particularly in processing cross-view data from satelli… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of task categories in CVSBench. The left panel shows the detailed categorization of task types in CVSBench, with counts or proportions indicated in parentheses. The right panel presents a word cloud of the QA composition and the distribution of typical answer categories. 3.1 Data Annotation Pipeline CVSBench is constructed from two cross-view datasets: CVUSA [54] and Uni￾versity1652 [69]. We curat… view at source ↗
Figure 4
Figure 4. Figure 4: Framework for enhancing spatial reasoning in VLMs. The center shows the SFT dataset annotation pipeline for textual CoT reasoning, while the right illustrates the generation of complementary visual information to support explicit spatial imagination. Training with GRPO. Based on the annotated datasets, we employ standard SFT followed by RL using the Group Relative Policy Optimization (GRPO) [45] framework.… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples. A comparison of methods for enhancing the spatial understanding capabilities of VLMs. Green indicates GT or correct prediction. Red indicates incorrect prediction. 5 Conclusion We present CVSBench, a large-scale benchmark for evaluating cross-view spa￾tial reasoning and imagination in VLMs under the satellite–street view setting. Our study reveals that current VLMs struggle to maintai… view at source ↗
Figure 6
Figure 6. Figure 6: Task-specific annotation interfaces used in dataset construction. (a) Viewpoint annotation interface for the FOV-subset, where annotators label camera locations and viewing directions using arrows on the satellite image. (b) Cross-view bounding-box annotation interface for the FOV-subset, built upon the previously annotated viewpoint arrows. (c) Bounding-box verification interface for the CVUSA-subset, whe… view at source ↗
Figure 7
Figure 7. Figure 7: Representative cross-view VQA examples from the CVUSA-subset, including Ground-to-Satellite and Satellite-to-Ground question types. A.4 Spatial Imagination Generation with Nanobanana To provide additional spatial imagination cues for cross-view reasoning, we gen￾erate auxiliary 3D miniature views using a Nanobanana-style image generation pipeline. The generated image is designed as a clean isometric archit… view at source ↗
Figure 8
Figure 8. Figure 8: Representative cross-view VQA examples from the FOV-subset, including fa￾cade visibility, roof form, symmetry, occlusion reasoning, relative position, facade color, ground material, height comparison, and vegetation reasoning [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative examples of the grounding and viewpoint localization tasks. The figure shows one representative cross-view grounding example together with the View-Arrow and View-Image tasks. For brevity, only one grounding example is shown, since the grounding task format is consistent across subsets. A high-fidelity 3D isometric miniature model reconstructed from a satel￾lite map layout, blending accurate… view at source ↗
Figure 10
Figure 10. Figure 10: Illustration of the three spatial imagination settings used in our Nanobanana￾style generation pipeline. Left top: satellite-view input and the corresponding gener￾ated isometric miniature. Left bottom: single street-view input and the corresponding generated miniature. Right: multi-view street-view inputs used for symmetry-related questions and the resulting unified isometric miniature. The generated min… view at source ↗
Figure 11
Figure 11. Figure 11: comparison of similarity heatmaps across different datasets. Brightness in￾dicates cosine similarity between image features extracted by CLIP. Compared with existing spatial reasoning benchmarks such as MindCube and NLVR2, our datasets (FOV and CVUSA) exhibit significantly lower off-diagonal similarity, indicating larger cross-view discrepancies and stronger viewpoint variations. A similar phenomenon can … view at source ↗
Figure 12
Figure 12. Figure 12: Reward curves during GRPO training for the two reasoning variants. Both models show steady reward improvement, while Imagination CoT maintains a slightly higher reward in the early and middle stages. The curves converge in later training steps, indicating stable optimization. C.2 Additional Qualitative Results We provide additional qualitative comparisons between the Cross-View Imagi￾nation Chain-of-Thoug… view at source ↗
Figure 13
Figure 13. Figure 13: Additional qualitative comparisons between Imagination-CoT and Structured￾CoT on representative CVSBench examples [PITH_FULL_IMAGE:figures/full_fig_p047_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional qualitative comparisons between Imagination-CoT and Structured￾CoT on representative CVSBench examples [PITH_FULL_IMAGE:figures/full_fig_p048_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional qualitative comparisons between Imagination-CoT and Structured￾CoT on representative CVSBench examples [PITH_FULL_IMAGE:figures/full_fig_p049_15.png] view at source ↗
read the original abstract

Humans can effortlessly reason about scenes across different viewpoints, yet it remains unclear whether Vision-Language Models (VLMs) possess similar cross-view spatial abilities. Satellite-street scene pairs, with their complex contexts and extreme viewpoint variations, provide an ideal testbed. Motivated by this, we introduce CVSBench, a large-scale benchmark for evaluating cross-view spatial reasoning through satellite-street pairs. This benchmark supports multiple tasks, including cross-view VQA, cross-view grounding, and viewpoint identification. CVSBench comprises 3,297 cross-view image groups with 9,468 object-level annotations and 40,679 question-answer (QA) pairs, enabling systematic and controlled evaluation of cross-view spatial reasoning. Extensive evaluations reveal that advanced VLMs struggle to maintain object-level and layout consistency under drastic viewpoint changes. To bridge this gap towards human-like spatial cognition, we investigate two categories of approaches: spatially grounded reasoning and the incorporation of cognitive map inputs. Our findings demonstrate that language-only reasoning yields marginal improvements, while incorporating visual spatial imagination via a 3D scene imagination pipeline substantially improves cross-view reasoning. These results highlight the necessity of explicit visual-spatial representations for robust spatial cognition in VLMs. Our data and code are released at https://huggingface.co/datasets/zlyzlyzly/CVSBench.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces CVSBench, a benchmark using 3,297 satellite-street image groups, 9,468 object annotations, and 40,679 QA pairs to evaluate cross-view spatial reasoning in VLMs via tasks such as cross-view VQA, grounding, and viewpoint identification. Evaluations show VLMs struggle with consistency under viewpoint changes; language-only approaches yield marginal gains while a 3D scene imagination pipeline yields substantial improvements, supporting the conclusion that explicit visual-spatial representations are required for robust spatial cognition. The dataset and code are released.

Significance. If the benchmark tasks require genuine viewpoint-invariant spatial understanding rather than non-spatial patterns, the work would be significant for identifying VLM limitations in spatial reasoning and motivating visual imagination mechanisms, with the public release aiding reproducibility and follow-on research.

major comments (1)
  1. [Abstract] Abstract: The central claim that the 3D scene imagination pipeline 'substantially improves cross-view reasoning' (and thus demonstrates necessity of explicit visual-spatial representations) rests on CVSBench tasks genuinely measuring cross-view spatial abilities. However, the provided description of the 3,297 image groups and 40,679 QA pairs does not specify controls or validation against non-spatial shortcuts such as object co-occurrence statistics, lighting correlations, or phrasing patterns that could differ systematically by view type.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for this constructive comment on the abstract and the underlying validity of the benchmark. We agree that stronger documentation of controls against non-spatial shortcuts is needed to support the central claims and will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the 3D scene imagination pipeline 'substantially improves cross-view reasoning' (and thus demonstrates necessity of explicit visual-spatial representations) rests on CVSBench tasks genuinely measuring cross-view spatial abilities. However, the provided description of the 3,297 image groups and 40,679 QA pairs does not specify controls or validation against non-spatial shortcuts such as object co-occurrence statistics, lighting correlations, or phrasing patterns that could differ systematically by view type.

    Authors: We acknowledge that the manuscript does not currently provide explicit validation or controls against the listed non-spatial shortcuts. In the revised version we will add a dedicated subsection under 'Benchmark Construction' that details the QA generation protocol: questions were authored to require explicit cross-view object correspondence and layout reasoning (e.g., relative positioning, occlusion, and scale changes that are viewpoint-specific), with multiple annotators and automated checks to avoid view-dependent phrasing or lighting cues. We will also report an ablation in which non-spatial cues are randomized while preserving spatial relations, showing that model accuracy drops substantially, indicating reliance on spatial rather than statistical shortcuts. These additions will directly support the claim that CVSBench measures genuine cross-view spatial abilities. revision: yes

Circularity Check

0 steps flagged

No circularity: benchmark and empirical results are independent of self-referential fitting

full rationale

The paper introduces CVSBench as a new dataset with satellite-street pairs, annotations, and QA tasks, then reports empirical evaluations of VLMs with and without 3D imagination pipelines. No equations, parameter fittings, or derivations are present. Benchmark construction and performance comparisons do not reduce to self-definitions, fitted inputs renamed as predictions, or load-bearing self-citations. The central claim rests on experimental outcomes that remain falsifiable against external models and data splits, making the work self-contained against the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract contains no mathematical derivations, fitted parameters, or postulated entities; work is purely empirical benchmark creation and evaluation.

pith-pipeline@v0.9.1-grok · 5780 in / 1024 out tokens · 34030 ms · 2026-06-26T10:52:16.990079+00:00 · methodology

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

90 extracted references · 13 linked inside Pith

  1. [1]

    Advances in neural information processing systems35, 23716– 23736 (2022) 4

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Men- sch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems35, 23716– 23736 (2022) 4

    2022

  2. [2]

    In: arxiv (2025) 10

    An, X., Xie, Y., Yang, K., et al.: Llava-onevision-1.5: Fully open framework for democratized multimodal training. In: arxiv (2025) 10

    2025

  3. [3]

    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) 2, 4

    2022

  4. [4]

    arXiv preprint arXiv:2511.21631 (2025) 1, 4, 10, 11

    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) 1, 4, 10, 11

    Pith/arXiv arXiv 2025

  5. [5]

    arXiv preprint arXiv:2502.13923 (2025) 1, 10

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., et al.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025) 1, 10

    Pith/arXiv arXiv 2025

  6. [6]

    International Research in Geographical and environmental education28(4), 262–280 (2019) 2, 4, 9

    Bednarz, R., Lee, J.: What improves spatial thinking? evidence from the spatial thinking abilities test. International Research in Geographical and environmental education28(4), 262–280 (2019) 2, 4, 9

    2019

  7. [7]

    arXiv preprint arXiv:2308.06595 (2023) 7

    Bitton, Y., Bansal, H., Hessel, J., Shao, R., Zhu, W., Awadalla, A., Gardner, J., Taori, R., Schmidt, L.: Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595 (2023) 7

    arXiv 2023

  8. [8]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA)

    Cai, W., Ponomarenko, I., Yuan, J., Li, X., Yang, W., Dong, H., Zhao, B.: Spa- tialbot: Precise spatial understanding with vision language models. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 9490–9498. IEEE (2025) 10

    2025

  9. [9]

    In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition

    Chen, B., Xu, Z., Kirmani, S., Ichter, B., Sadigh, D., Guibas, L., Xia, F.: Spa- tialvlm: Endowing vision-language models with spatial reasoning capabilities. In: ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecog- nition. pp. 14455–14465 (2024) 1, 3, 4, 10

    2024

  10. [10]

    Advances in Neural Information Processing Systems37, 135062–135093 (2024) 1, 3, 4, 7

    Cheng, A.C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., Liu, S.: Spatialrgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems37, 135062–135093 (2024) 1, 3, 4, 7

    2024

  11. [11]

    arXiv preprint arXiv:2507.06261 (2025) 1, 6, 8, 9

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025) 1, 6, 8, 9

    Pith/arXiv arXiv 2025

  12. [12]

    Advances in neural information processing systems36, 49250–49267 (2023) 4 16 Ruixun Liu et al

    Dai, W., Li, J., Li, D., Tiong, A., Zhao, J., Wang, W., Li, B., Fung, P.N., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning. Advances in neural information processing systems36, 49250–49267 (2023) 4 16 Ruixun Liu et al

    2023

  13. [13]

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

    Ding,X.,Han,J.,Xu,H.,Liang,X.,Zhang,W.,Li,X.:Holisticautonomousdriving understanding by bird’s-eye-view injected multi-modal large models. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13668–13677 (2024) 2

    2024

  14. [14]

    arXiv preprint arXiv:2306.13394 (2023) 4, 7

    Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) 4, 7

    Pith/arXiv arXiv 2023

  15. [15]

    In: European Conference on Computer Vision

    Fu, X., Hu, Y., Li, B., Feng, Y., Wang, H., Lin, X., Roth, D., Smith, N.A., Ma, W.C., Krishna, R.: Blink: Multimodal large language models can see but not per- ceive. In: European Conference on Computer Vision. pp. 148–166. Springer (2024) 4, 7

    2024

  16. [16]

    arXiv preprint arXiv:2509.06266 (2025) 2

    Gholami, M., Rezaei, A., Weimin, Z., Mao, S., Zhou, S., Zhang, Y., Akbari, M.: Spatial reasoning with vision-language models in ego-centric multi-view scenes. arXiv preprint arXiv:2509.06266 (2025) 2

    arXiv 2025

  17. [17]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., et al.: Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14375–14385 (2024) 4

    2024

  18. [18]

    Advances in Neural Information Processing Systems36, 20482–20494 (2023) 4

    Hong, Y., Zhen, H., Chen, P., Zheng, S., Du, Y., Chen, Z., Gan, C.: 3d-llm: In- jecting the 3d world into large language models. Advances in Neural Information Processing Systems36, 20482–20494 (2023) 4

    2023

  19. [19]

    ISPRS Journal of Photogrammetry and Remote Sensing224, 272–286 (2025) 5, 7

    Hu, Y., Yuan, J., Wen, C., Lu, X., Liu, Y., Li, X.: Rsgpt: A remote sensing vision language model and benchmark. ISPRS Journal of Photogrammetry and Remote Sensing224, 272–286 (2025) 5, 7

    2025

  20. [20]

    arXiv preprint arXiv:2410.21276 (2024) 1, 4, 10

    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) 1, 4, 10

    Pith/arXiv arXiv 2024

  21. [21]

    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) 2, 4

    2025

  22. [22]

    arXiv preprint arXiv:2506.03135 (2025) 2, 4, 7

    Jia, M., Qi, Z., Zhang, S., Zhang, W., Yu, X., He, J., Wang, H., Yi, L.: Omnispatial: Towards comprehensive spatial reasoning benchmark for vision language models. arXiv preprint arXiv:2506.03135 (2025) 2, 4, 7

    arXiv 2025

  23. [23]

    Cognitive science4(1), 71–115 (1980) 4

    Johnson-Laird, P.N.: Mental models in cognitive science. Cognitive science4(1), 71–115 (1980) 4

    1980

  24. [24]

    49–54 (2025) 10

    Khandoga, M., Kostiuk, Y., Polishko, A., Kozlov, K., Filipchuk, Y., Kiulian, A.: Framing the language: Fine-tuning gemma 3 for manipulation detection pp. 49–54 (2025) 10

    2025

  25. [25]

    arXiv preprint arXiv:2406.09246 (2024) 5

    Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., Sanketi, P., et al.: Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246 (2024) 5

    Pith/arXiv arXiv 2024

  26. [26]

    Kuckreja, K., Danish, M.S., Naseer, M., Das, A., Khan, S., Khan, F.S.: Geochat: Grounded large vision-language model for remote sensing (2023) 10

    2023

  27. [27]

    Journal of geography111(1), 15–26 (2012) 4, 9

    Lee, J., Bednarz, R.: Components of spatial thinking: Evidence from a spatial thinking ability test. Journal of geography111(1), 15–26 (2012) 4, 9

    2012

  28. [28]

    Lee, P.Y., Je, J., Park, C., Uy, M.A., Guibas, L., Sung, M.: Perspective-aware reasoninginvision-languagemodelsviamentalimagerysimulation.In:Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9241–9251 (2025) 2, 3, 4 CVSBench 17

    2025

  29. [29]

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

    Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: Seed- bench: Benchmarking multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13299– 13308 (2024) 7

    2024

  30. [30]

    arXiv preprint arXiv:2505.21500 (2025) 2

    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) 2

    arXiv 2025

  31. [31]

    In: International conference on machine learning

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: International conference on machine learning. pp. 19730–19742. PMLR (2023) 4

    2023

  32. [32]

    Advances in Neural Information Processing Systems37, 3229–3242 (2024) 5, 7

    Li, X., Ding, J., Elhoseiny, M.: Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. Advances in Neural Information Processing Systems37, 3229–3242 (2024) 5, 7

    2024

  33. [33]

    In: Proceedings of the 2023 conference on empirical methods in natural language processing

    Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hal- lucination in large vision-language models. In: Proceedings of the 2023 conference on empirical methods in natural language processing. pp. 292–305 (2023) 4

    2023

  34. [34]

    2: Pushing the frontier of open large language models (2025) 10

    Liu, A., Mei, A., Lin, B., Xue, B., Wang, B., Xu, B., Wu, B., Zhang, B., Lin, C., Dong, C., et al.: Deepseek-v3. 2: Pushing the frontier of open large language models (2025) 10

    2025

  35. [35]

    Transactions of the Association for Computational Linguistics11, 635–651 (2023) 1, 4, 7

    Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics11, 635–651 (2023) 1, 4, 7

    2023

  36. [36]

    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) 4

    2024

  37. [37]

    Advances in neural information processing systems36, 34892–34916 (2023) 4

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023) 4

    2023

  38. [38]

    Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? In: European conference on computer vision. pp. 216–233. Springer (2024) 7

    2024

  39. [39]

    IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555–8566 (2020) 5, 7

    Lobry, S., Marcos, D., Murray, J., Tuia, D.: Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing 58(12), 8555–8566 (2020) 5, 7

    2020

  40. [40]

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

    Ma, W., Chen, H., Zhang, G., Chou, Y.C., Chen, J., de Melo, C., Yuille, A.: 3dsr- bench: A comprehensive 3d spatial reasoning benchmark. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6924–6934 (2025) 4, 7

    2025

  41. [41]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Mitra, C., Huang, B., Darrell, T., Herzig, R.: Compositional chain-of-thought prompting for large multimodal models. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 14420–14431 (2024) 4

    2024

  42. [42]

    In: European Conference on Computer Vision

    Muhtar, D., Li, Z., Gu, F., Zhang, X., Xiao, P.: Lhrs-bot: Empowering remote sens- ing with vgi-enhanced large multimodal language model. In: European Conference on Computer Vision. pp. 440–457. Springer (2024) 5, 7

    2024

  43. [43]

    In: Studying visual and spatial reasoning for design creativity, pp

    Newcombe, N.S., Shipley, T.F.: Thinking about spatial thinking: New typology, new assessments. In: Studying visual and spatial reasoning for design creativity, pp. 179–192. Springer (2014) 2, 4, 9

    2014

  44. [44]

    Ramakrishnan, S.K., Wijmans, E., Kraehenbuehl, P., Koltun, V.: Does spatial cognition emerge in frontier models? arXiv preprint arXiv:2410.06468 (2024) 4

    arXiv 2024

  45. [45]

    arXiv preprint arXiv:2402.03300 (2024) 9 18 Ruixun Liu et al

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024) 9 18 Ruixun Liu et al

    Pith/arXiv arXiv 2024

  46. [46]

    arXiv preprint arXiv:2601.03267 (2025) 10

    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) 10

    Pith/arXiv arXiv 2025

  47. [47]

    In: European Conference on Computer Vision

    Szymańska, E., Dusmanu, M., Buurlage, J.W., Rad, M., Pollefeys, M.: Space3d- bench: Spatial 3d question answering benchmark. In: European Conference on Computer Vision. pp. 68–85. Springer (2024) 2

    2024

  48. [48]

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

    Toker, A., Zhou, Q., Maximov, M., Leal-Taixé, L.: Coming down to earth: Satellite- to-street view synthesis for geo-localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6488–6497 (2021) 2, 5

    2021

  49. [49]

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

    Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9568–9578 (2024) 4

    2024

  50. [50]

    arXiv preprint arXiv:2409.12191 (2024) 4

    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) 4

    Pith/arXiv arXiv 2024

  51. [51]

    5: Advancing open-source multimodal models in versatility, reasoning, and efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025) 1, 10

    Pith/arXiv arXiv 2025

  52. [52]

    Wang, Y., Liu, Z., Wang, Z., Hu, H., Liu, P., Rao, Y.: Geovista: Web-augmented agenticvisualreasoningforgeolocalization.arXivpreprintarXiv:2511.15705(2025) 2, 5

    arXiv 2025

  53. [53]

    arXiv preprint arXiv:2308.08769 (2023) 4

    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) 4

    arXiv 2023

  54. [54]

    In: Proceedings of the IEEE International Conference on Com- puter Vision

    Workman, S., Souvenir, R., Jacobs, N.: Wide-area image geolocalization with aerial reference imagery. In: Proceedings of the IEEE International Conference on Com- puter Vision. pp. 3961–3969 (2015) 6

    2015

  55. [55]

    arXiv preprint arXiv:2505.23747 (2025) 4

    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) 4

    Pith/arXiv arXiv 2025

  56. [56]

    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 (2025) 10

    2025

  57. [57]

    Advances in Neural Information Processing Systems37, 90277–90317 (2024) 3

    Wu, W., Mao, S., Zhang, Y., Xia, Y., Dong, L., Cui, L., Wei, F.: Mind’s eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems37, 90277–90317 (2024) 3

    2024

  58. [58]

    arXiv preprint arXiv:2601.14339 (2026) 5

    Xu, H., Hu, Y., Zhu, Z., Gao, C., Wang, Z., Rao, J., Lu, W., Li, W., Yin, Q., Li, Y.: Citycube: Benchmarking cross-view spatial reasoning on vision-language models in urban environments. arXiv preprint arXiv:2601.14339 (2026) 5

    arXiv 2026

  59. [59]

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

    Xu,Y.,Zhu,L.,Yang,Y.:Mc-bench:Abenchmarkformulti-contextvisualground- ing in the era of mllms. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17675–17687 (2025) 7

    2025

  60. [60]

    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) 2, 4, 7

    2025

  61. [61]

    Advances in Neural Information Processing Systems37, 21875–21911 (2024) 9

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. Advances in Neural Information Processing Systems37, 21875–21911 (2024) 9

    2024

  62. [62]

    arXiv preprint arXiv:2506.17218 (2025) 3 CVSBench 19

    Yang, Z., Yu, X., Chen, D., Shen, M., Gan, C.: Machine mental imagery: Empower multimodal reasoning with latent visual tokens. arXiv preprint arXiv:2506.17218 (2025) 3 CVSBench 19

    Pith/arXiv arXiv 2025

  63. [63]

    In: Structural Priors for Vision Workshop at ICCV’25 (2025) 2, 4, 7

    Yin, B., Wang, Q., Zhang, P., Zhang, J., Wang, K., Wang, Z., Zhang, J., Chan- drasegaran, K., Liu, H., Krishna, R., et al.: Spatial mental modeling from limited views. In: Structural Priors for Vision Workshop at ICCV’25 (2025) 2, 4, 7

    2025

  64. [64]

    arXiv preprint arXiv:2509.18905 (2025) 1, 2, 4, 7

    Yu, S., Chen, Y., Ju, H., Jia, L., Zhang, F., Huang, S., Wu, Y., Cui, R., Ran, B., Zhang, Z., et al.: How far are vlms from visual spatial intelligence? a benchmark- driven perspective. arXiv preprint arXiv:2509.18905 (2025) 1, 2, 4, 7

    arXiv 2025

  65. [65]

    arXiv preprint arXiv:2406.10721 (2024) 5

    Yuan, W., Duan, J., Blukis, V., Pumacay, W., Krishna, R., Murali, A., Mousavian, A., Fox, D.: Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv preprint arXiv:2406.10721 (2024) 5

    arXiv 2024

  66. [66]

    IEEE transactions on geoscience and remote sensing 61, 1–13 (2023) 5, 7

    Zhan, Y., Xiong, Z., Yuan, Y.: Rsvg: Exploring data and models for visual ground- ing on remote sensing data. IEEE transactions on geoscience and remote sensing 61, 1–13 (2023) 5, 7

    2023

  67. [67]

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

    Zhang, L., Zhai, X., Zhao, Z., Zong, Y., Wen, X., Zhao, B.: What if the tv was off? examining counterfactual reasoning abilities of multi-modal language mod- els. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21853–21862 (2024) 3

    2024

  68. [68]

    arXiv preprint arXiv:2302.00923 (2023) 4

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain- of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 4

    Pith/arXiv arXiv 2023

  69. [69]

    In: Proceedings of the 28th ACM interna- tional conference on Multimedia

    Zheng, Z., Wei, Y., Yang, Y.: University-1652: A multi-view multi-source bench- mark for drone-based geo-localization. In: Proceedings of the 28th ACM interna- tional conference on Multimedia. pp. 1395–1403 (2020) 6

    2020

  70. [70]

    arXiv preprint arXiv:2409.18125 (2024) 5

    Zhu, C., Wang, T., Zhang, W., Pang, J., Liu, X.: Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125 (2024) 5

    arXiv 2024

  71. [71]

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

    Zhu, S., Yang, T., Chen, C.: Vigor: Cross-view image geo-localization beyond one- to-one retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3640–3649 (2021) 2, 5

    2021

  72. [72]

    sample_id

    Zhu, Z., Wang, X., Li, Y., Zhang, Z., Ma, X., Chen, Y., Jia, B., Liang, W., Yu, Q., Deng, Z., et al.: Move to understand a 3d scene: Bridging visual grounding and exploration for efficient and versatile embodied navigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8120–8132 (2025) 2 20 Ruixun Liu et al. A Dataset Co...

    2025

  73. [73]

    distance_to_camera

    “distance_to_camera” 2) “building_footprint_area” 3) “build- ing_connection” 4) “roof_form” ========================================= QUESTION COUNT Generate EXACTLY 3 questions per image pair. If distance/area cannot meet EXTREME-ONLY rule -> do NOT force them. Use roof_form and/or building_connection instead. ========================================= MU...

  74. [74]

    ========================================= CATEGORY RULES

    vegetation_sector 2) occlusion_binary 3) relative_position 4) height_comparison 5) facade_color 6) ground_material ========================================= TARGET DEFINITION REQUIREMENT (CRITICAL) Every target MUST be: - clearly identifiable in satellite - uniquely disam- biguated using satellite-only cues such as: distance to camera, alignment with the ...

  75. [75]

    What is the domi- nant vegetation type in the [sector] along the arrow direction?

    vegetation_sector (4 options) Question template: “What is the domi- nant vegetation type in the [sector] along the arrow direction?” Options EXACTLY: A) Trees B) Shrubs / bushes C) Grass / lawn D) No obvious vegetation

  76. [76]

    Is [satellite- defined target] occluded in Street-View?

    occlusion_binary (2 options ONLY) Question template: “Is [satellite- defined target] occluded in Street-View?” Options EXACTLY: A) Oc- cluded (blocked) B) Not occluded (clear view)

  77. [77]

    Facing forward (the arrow direction), where does [satellite-defined target] appear in Street-View?

    relative_position (4 options) Question template EXACTLY: “Facing forward (the arrow direction), where does [satellite-defined target] appear in Street-View?” Options EXACTLY: A) Left side B) Right side C) Near the center (mostly straight ahead) D) Not visible in the forward view

  78. [78]

    Com- pare the heights of (1) [Building A defined by satellite geometry] and (2) [Building B defined by satellite geometry]. Which is taller in Street- View?

    height_comparison (4 options, STRICT) Question template: “Com- pare the heights of (1) [Building A defined by satellite geometry] and (2) [Building B defined by satellite geometry]. Which is taller in Street- View?” Options EXACTLY: A) The first building is taller B) The second building is taller C) They are about the same height D) Only one of them is vi...

  79. [79]

    What is the domi- nant ground material in the forward sector along the arrow direction?

    ground_material (4 options) Question template: “What is the domi- nant ground material in the forward sector along the arrow direction?” Options EXACTLY: A) Asphalt B) Concrete C) Grass D) Bare soil / dirt ========================================= FINAL VALIDITY CHECK (MANDATORY) For EACH question verify ALL: 1) Target is uniquely localizable using satell...

  80. [80]

    I stand at the satellite center and face the TOP of the satellite image

    Correct answer is determined by checking Street-View. 4) Street-View clearly confirms the correct option. If ANY check fails -> DISCARD that question. 30 Ruixun Liu et al. S2G-CVUSA Prompt.The CVUSA S2G prompt uses the same general logic as the FOV version, but replaces the observation-arrow geometry with a fixed panorama mapping centered at the satellite...

Showing first 80 references.