Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Spatial reasoning errors in multimodal AI drop sharply when 3D evidence is extracted from the image and handed back as structured cards.

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 08:29 UTC pith:7A2GRGHX

load-bearing objection Training-free 3D card augmentation gives real gains on spatial benchmarks, but 'beats GPT-5' rests on a 0.04-point average and the orientation estimator's ~75% accuracy is the main thing that could break in the wild. the 4 major comments →

arxiv 2607.21085 v1 pith:7A2GRGHX submitted 2026-07-23 cs.CV

Geo3R: Mitigating Spatial Reasoning Hallucination in Multimodal Large Language Models

Mingyu Wang , Weilin Jin , Wenbo Li , Haoyang Huang , Tong Jia , Ying Li This is my paper
classification cs.CV
keywords spatial reasoning hallucinationmultimodal large language modelshallucination mitigation3D geometric reasoningsingle-image 3D estimationobject orientationviewpoint reasoning
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.

This paper argues that the spatial mistakes made by multimodal large language models (MLLMs) are a distinct failure mode — spatial reasoning hallucination — that standard hallucination fixes cannot touch, because those fixes operate on 2D visual-semantic alignment while the errors come from a gap between 2D pixels and 3D reality. To close that gap, Geo3R extracts metric depth, camera pose, gravity direction, and object orientation from a single image, then presents the recovered 3D measurements to the MLLM as structured 'geometric cards' organized by coordinate frame. On 18 tasks spanning perspective effects, object orientation, and viewpoint changes, the framework raises the 18-task average accuracy of Qwen3-VL from 55.07% to 65.97%, of Gemini-3-Flash from 65.34% to 72.40%, and of the much weaker LLaVA-1.5 from 43.23% to 46.10% — gains that outperform existing hallucination-mitigation methods and the concurrent tool-augmented APC-VLM, with no training. The core claim is that the missing ingredient in spatial reasoning is explicit, structured 3D evidence, and that it can be supplied at inference time.

Core claim

The central discovery is that spatial reasoning hallucination in MLLMs is caused by insufficient spatial structure modeling — the model reasons from 2D image cues that systematically diverge from 3D geometry (perspective foreshortening, object facing, observer viewpoint) — rather than by visual-textual misalignment. Geo3R lifts 2D pixels through four coordinate frames — image, camera, gravity-aligned world, and object-local — using monocular estimators, and packages the resulting measurements as three task-agnostic cards: an Object Card that anchors names to image regions, a World Card with depth, 3D extents, distances, and headings, and an Object-Local Card with offsets and directions in ea

What carries the argument

The load-bearing mechanism is the structured geometric card: a fixed set of fields, computed identically for every question, that decomposes 3D evidence by coordinate frame and pairs each continuous measurement with a short natural-language summary, plus field notes that state coordinate conventions. The cards let the MLLM select the relevant frame (world for global depth/height, object-local for egocentric left/right/front/behind) without a task-type classifier, and the structured pairing lets the model cross-check raw numbers against verbal summaries. The pipeline is carried by a chain of four pretrained monocular estimators that back-project pixels to 3D using depth and focal length, rota

Load-bearing premise

The framework is only as trustworthy as the pretrained monocular estimators that supply depth, camera pose, and object orientation; if those estimates are wrong, especially the orientation estimates, the cards feed the model misleading evidence.

What would settle it

Take a held-out set of scenes with known 3D ground truth, deliberately corrupt the Object-Local Card's facing directions (e.g., rotate every object's orientation by 180°) and measure Geo3R's accuracy; if it stays near the uncorrupted level, the MLLM is not actually using the orientation evidence, and the framework's gains must come from something else.

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

If this is right

  • Any MLLM can be augmented at inference time without training, so spatial reasoning improvements can be decoupled from model scale.
  • Existing hallucination-mitigation methods — attention reallocation, decoding penalties, triplet self-alignment — are insufficient for spatial tasks; a structured 3D evidence layer is required.
  • The same geometric evidence improves different base models by different amounts, indicating the bottleneck moves from evidence availability to the model's ability to integrate structured cues.
  • The card format itself matters: removing field names or unstructured text costs up to 6.9%, so how evidence is presented is part of the effect.
  • Geo3R's gains extend to the newest and strongest models (e.g., consistent improvements on Gemini-3-Pro and GPT-5.5 in the appendix), suggesting the approach is not saturated.

Where Pith is reading between the lines

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

  • If the central claim holds, spatial benchmarks may be better seen as tests of a model's access to geometric evidence rather than its innate spatial intelligence; a similar evidence-card layer could be built for other hallucination types, such as counting or attribute grounding, where 2D evidence is also insufficient.
  • Because the framework's ceiling is set by monocular estimators, the method will improve automatically as those estimators improve; a testable prediction is that Geo3R's gain will correlate with per-image orientation estimation accuracy.
  • The finding that removing the image in Stage 2 slightly helps on one benchmark (3DSR with Qwen3-VL) hints that, for some models, geometric cards can partially substitute for pixel input — a property that might extend to low-vision or compressed-input settings.
  • One could probe whether the discrete summaries or the raw numbers carry the signal: feeding only quantized directions (without extents/offsets) would isolate which card field drives the improvement.

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

4 major / 6 minor

Summary. The paper proposes Geo3R, a training-free, plug-and-play framework that augments MLLMs with explicit 3D geometric evidence extracted from a single image. It identifies a subcategory of relation hallucination it calls spatial reasoning hallucination, grouped into perspective, orientation, and viewpoint scenarios. Geo3R grounds task-relevant objects, back-projects them into camera, gravity-aligned world, and object-local coordinate frames using monocular depth, camera calibration, segmentation, and orientation estimators, and then presents the evidence as structured Object, World, and Object-Local cards in a second-stage prompt. The method is evaluated on 17,493 samples across 18 tasks from 3DSRBench, ViewSpatial-Bench, and CV-Bench, on LLaVA-1.5, Qwen3-VL, and Gemini-3-Flash. The reported average accuracies improve from 43.23% to 46.10%, 55.07% to 65.97%, and 65.34% to 72.40%, respectively, with the Qwen3-VL result surpassing GPT-5 by 0.04 points on the 18-task average. The paper also reports ablations, comparisons with hallucination-mitigation methods and APC-VLM, tool success/accuracy audits, and noise-robustness experiments.

Significance. If the reported results hold, Geo3R is a practically useful contribution: it demonstrates that explicit monocular 3D evidence can reduce spatial reasoning errors in MLLMs without training, and it provides a unifying treatment of perspective, orientation, and viewpoint failures. The paper has notable strengths: evaluation uses official benchmark splits and a standard accuracy metric; the component ablations in Table 4 and card-format ablations in Appendix F are meaningful; the robustness analysis in Appendix E, including tool accuracy auditing and synthetic noise injection, is more transparent than is typical; and the additional baselines in Appendix G strengthen the comparison. The central claim is nevertheless load-bearing on the reliability of upstream monocular estimators, especially the orientation estimator, and on the statistical significance of small average differences. These issues are addressable and do not in my view undermine the core methodology.

major comments (4)
  1. [§5.1.3, Tables 2 and 3] No error bars, confidence intervals, or significance tests are reported for any of the 18-task averages. This matters directly for the abstract's claim that Geo3R+Qwen3-VL 'surpasses GPT-5': the difference is 65.97% vs. 65.93%, i.e., 0.04 percentage points on a macro average over 17,493 samples. The paper should provide paired/bootstrap confidence intervals or a paired significance test across tasks or samples for at least the headline comparisons, and should report per-task variance. Without this, the GPT-5 superiority claim is not statistically supported.
  2. [§4.2.4, Table 11, Table 12, Table 4(d)] The paper's own robustness data show that the orientation estimator is the weakest link: Table 11 reports 75.6% own accuracy on a hand-annotated 100-image sample, while Table 4(d) shows that removing the Object-Local Card costs 12.7 points on Viewpoint tasks. Table 12 shows that flipping the facing direction (O.Flip) drops the 3DSR average from 68.1% to 63.9%, leaving only 5.8 points above the no-card baseline of 57.1%. Thus roughly one in four Object-Local card entries is wrong, and wrong orientation evidence can actively mislead rather than merely fail to help. The limitation statement in §6 concedes that tool accuracy upper-bounds framework performance, but the paper does not quantify the net expected benefit of the Object-Local Card conditional on estimator correctness. I recommend adding a per-task breakdown of orientation-estimator accuracy on the evaluation benchmarks, and either
  3. [§5.3, Table 3; Abstract] The comparison with existing hallucination-mitigation methods is restricted to LLaVA-1.5 (OPERA, Tri-HE, Reefknot, AdaptVis), and the tool-augmented baseline APC-VLM is evaluated only on Qwen3-VL. This supports the claim that Geo3R works on LLaVA-1.5 and on Qwen3-VL against APC-VLM, but the abstract's generalization 'outperforming existing models and methods' is broader than the evidence. The paper should either qualify the claim to the compared settings or, where implementable, run the strongest baselines on additional base models. At minimum, the abstract and §5.3 should state the base-model restriction explicitly.
  4. [Appendix E, Table 11] GeoCalib is explicitly excluded from the tool-validation study because its outputs are difficult to verify without ground-truth calibration. Yet GeoCalib's focal length and roll/pitch estimates enter Eqs. (1)–(4) and therefore affect all subsequent back-projection, world-frame extents, and object-local directions. The paper should provide at least an indirect validation of GeoCalib on images with known calibration or a sensitivity analysis in which focal length and gravity alignment are perturbed, to show that the headline results are not driven by unvalidated camera parameters.
minor comments (6)
  1. [§1] Typo: 'we definespatial reasoning hallucinationas' is missing spaces around 'defines'.
  2. [Figure 2] The example captions contain grammatical errors ('A apple on a armchair', 'A apple under a armchair'); these should be corrected.
  3. [§5.2, Figure 4] The text says Perspective improves by 10.24% on Qwen3-VL while Figure 4 labels the gain '+10.2'; please make the precision consistent.
  4. [Appendix E] The appendix says 'as shown in Section G' but does not define Section G in the main text or appendix; the cross-reference should be fixed.
  5. [Appendix D.2, Table 10] Inference-time measurements are reported for 720 samples but without detailing prompt length or hardware variance beyond a single H20 GPU; a sentence on the protocol would improve reproducibility.
  6. [Appendix C.4] The fallback 'use the answer choice texts directly as object names' could create duplicate or ambiguous object entries; a brief example of when this fallback triggers would help.

Circularity Check

0 steps flagged

No material circularity: Geo3R's gains are measured against held-out benchmark labels; card design is fixed and task-agnostic, with the only load-bearing dependency being upstream monocular tool accuracy, which the paper explicitly concedes.

full rationale

Geo3R is an empirical augmentation pipeline rather than a derivation whose conclusions are encoded in its inputs. The geometric cards are produced by fixed pretrained estimators (DepthPro, GeoCalib, Orient Anything V2, SAM) with no parameters fitted to the evaluation benchmarks; Section 5.1.3 states "we use the official evaluation splits as the ground truth," and Section 4.3 emphasizes that "all cards are task-agnostic, meaning the same set of fields is provided for every question." The reported accuracy gains are therefore not forced by construction. The definition of "spatial reasoning hallucination" in Section 3 is a conceptual taxonomy motivated by the authors' preliminary failure analysis, but the success metric is held-out multiple-choice accuracy, not the definition itself; no equation in the paper reduces a predicted quantity to a fitted parameter, a self-citation, or an imported uniqueness theorem. The identified dependency—monocular estimator accuracy, especially the 75.6% own accuracy of Orient Anything V2 in Table 11—is a robustness limitation, and the paper states it directly: "Geo3R relies on existing geometric estimation tools whose accuracy upper-bounds the framework's performance." Dependence on external tools is a correctness/generalization risk, not circularity. No load-bearing self-citations or ansatz-smuggling chains appear. Accordingly, the paper is self-contained against external benchmarks and merits a circularity score of 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

No new physical entity is posited. The central scientific load is carried by the reliability of off-the-shelf estimators and by the untested claim that explicit geometry in prompt text is faithfully used by MLLMs.

free parameters (2)
  • Trimmed quantile range for 3D bounding extent = 10th–90th percentile
    Used in Eq. (5) to define per-object 3D extents; chosen by hand and not fitted to evaluation labels.
  • Eight-way azimuth quantization = 45° bins centered at 0°
    Appendix C.3: quantizes directions for both camera-relative and object-local cues; a design choice that affects the card text.
axioms (5)
  • domain assumption Pinhole camera model with focal length f and principal point at image center; Eq. (2) assumes per-pixel metric depth z.
    The back-projection from 2D pixels to camera space uses DepthPro's metric depth and GeoCalib's focal length; if depth is not metric or calibration is wrong, the derived camera-space coordinates are inaccurate.
  • domain assumption DepthPro's monocular depth map is metric and aligned with the camera.
    Section 4.2.2 uses z directly in back-projection; depth errors propagate to all downstream cards.
  • domain assumption GeoCalib's roll/pitch aligns camera frame to gravity; gravity-aligned world Y is upward.
    Section 4.2.3 constructs world frame from Eq. (3); wrong gravity direction changes height comparisons and object-local transforms.
  • domain assumption Orient Anything V2 provides object orientation in camera frame that can be converted to world frame.
    Section 4.2.4 uses R_C_obj to build object-local frames; Table 11 shows own accuracy only ~75.6%, so this assumption is the weakest tool link.
  • domain assumption The base MLLM can reason correctly with structured cards and will not treat numeric fields as adversarial or ignore them.
    Geo3R is prompt-based; its effectiveness depends on instruction following. Ablations show cards help on average, but per-task effects vary.

pith-pipeline@v1.3.0-alltime-deepseek · 23263 in / 13069 out tokens · 115878 ms · 2026-08-01T08:29:56.111475+00:00 · methodology

0 comments
read the original abstract

Despite remarkable progress in visual understanding, Multimodal Large Language Models (MLLMs) remain prone to hallucinations when reasoning about spatial relationships, often producing judgments that contradict the true 3D structure of the scene. Though several existing works have proposed to mitigate hallucinations, our analysis indicates that they show limited effectiveness in spatial reasoning, as they fail to bridge the fundamental gap between 2D visual representations and 3D spatial reality. Based on this finding, we define hallucinations arising from insufficient spatial structure modeling as spatial reasoning hallucination, a subcategory of relation hallucination that existing mitigation methods fail to address. We further identify three typical scenarios where such hallucinations frequently occur: perspective effects, object orientation, and viewpoint changes. To this end, we propose Geo3R, a training-free, plug-and-play framework that incorporates geometric evidence and structured 3D reasoning to mitigate spatial reasoning hallucination. Experiments on three benchmarks, covering 18 tasks across all three scenarios, show that Geo3R substantially reduces spatial reasoning hallucination across diverse MLLMs without additional training, outperforming existing models and methods.

Figures

Figures reproduced from arXiv: 2607.21085 by Haoyang Huang, Mingyu Wang, Tong Jia, Weilin Jin, Wenbo Li, Ying Li.

Figure 1
Figure 1. Figure 1: The gap between 2D representations and 3D reality [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: General spatial relations (left) versus three spatial [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Geo3R. Specifically, DepthPro [3] produces a per-pixel metric depth map 𝐷, from which the depth 𝑧𝑖 for each object 𝑜𝑖 is computed. GeoCalib [27] estimates the focal length 𝑓 , from which we construct the intrinsic matrix 𝐾: 𝐾 = © ­ « 𝑓 0 𝑐𝑥 0 𝑓 𝑐𝑦 0 0 1 ª ® ¬ , (1) where (𝑐𝑥 , 𝑐𝑦) is the image center. Combined with segmentation masks refined by SAM [9], each pixel (𝑢, 𝑣) with depth 𝑧 is back￾pr… view at source ↗
Figure 4
Figure 4. Figure 4: Scenario-level average accuracy comparison across [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples of Geo3R mitigating spatial reasoning hallucinations across the three scenarios. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Scenario-level average accuracy of hallucination [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Geo3R two-stage pipeline on a height_higher example from 3DSRBench, with key prompt instructions shown inline. All outputs are from an actual Qwen3-VL run [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Geometric report produced for the example in Figure [PITH_FULL_IMAGE:figures/full_fig_p014_8.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

44 extracted references · 12 linked inside Pith

  1. [1]

    Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. 2024. Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930(2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Hunar Batra, Haoqin Tu, Hardy Chen, Yuanze Lin, Cihang Xie, and Ronald Clark

  3. [3]

    Aleksei Bochkovskii, AmaãG, l Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. 2024. Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073(2024)

    Pith/arXiv arXiv 2024

  4. [4]

    Zhongang Cai, Ruisi Wang, Chenyang Gu, Fanyi Pu, Junxiang Xu, Yubo Wang, Wanqi Yin, Zhitao Yang, Chen Wei, Qingping Sun, et al. 2025. Scaling spatial intelligence with multimodal foundation models.arXiv preprint arXiv:2511.13719 (2025)

    arXiv 2025

  5. [5]

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. 2024. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. 14455–14465

    2024

  6. [6]

    Shiqi Chen, Tongyao Zhu, Ruochen Zhou, Jinghan Zhang, Siyang Gao, Juan Car- los Niebles, Mor Geva, Junxian He, Jiajun Wu, and Manling Li. 2025. Why is spatial reasoning hard for vlms? an attention mechanism perspective on focus areas.arXiv preprint arXiv:2503.01773(2025)

    arXiv 2025

  7. [7]

    An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. 2024. Spatialrgpt: Grounded spatial reasoning in vision-language models.Advances in Neural Information Processing Systems37 (2024), 135062–135093

    2024

  8. [8]

    Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2024. Opera: Alleviating hal- lucination in multi-modal large language models via over-trust penalty and retrospection-allocation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13418–13427

    2024

  9. [9]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  10. [10]

    Phillip Y Lee, Jihyeon Je, Chanho Park, Mikaela Angelina Uy, Leonidas Guibas, and Minhyuk Sung. 2025. Perspective-aware reasoning in vision-language models via mental imagery simulation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 9241–9251

    2025

  11. [11]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. 2024. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326(2024)

    Pith/arXiv arXiv 2024

  12. [12]

    Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, et al. 2025. Viewspatial-bench: Evaluating multi-perspective spatial localization in vision- language models.arXiv preprint arXiv:2505.21500(2025)

    arXiv 2025

  13. [13]

    Pengteng Li, Pinhao Song, Wuyang Li, Weiyu Guo, Huizai Yao, Yijie Xu, Dugang Liu, and Hui Xiong. 2025. See&trek: Training-free spatial prompting for multi- modal large language model.arXiv preprint arXiv:2509.16087(2025)

    arXiv 2025

  14. [14]

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 conference on empirical methods in natural language processing. 292–305

    2023

  15. [15]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 26296–26306

    2024

  16. [16]

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. 2024. Grounding dino: Marry- ing dino with grounded pre-training for open-set object detection. InEuropean conference on computer vision. Springer, 38–55

    2024

  17. [17]

    Yuhong Liu, Beichen Zhang, Yuhang Zang, Yuhang Cao, Long Xing, Xiaoyi Dong, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Spatial-ssrl: Enhancing spatial understanding via self-supervised reinforcement learning.arXiv preprint arXiv:2510.27606(2025)

    arXiv 2025

  18. [18]

    Wufei Ma, Haoyu Chen, Guofeng Zhang, Yu-Cheng Chou, Jieneng Chen, Celso de Melo, and Alan Yuille. 2025. 3dsrbench: A comprehensive 3d spatial reasoning benchmark. InProceedings of the IEEE/CVF International Conference on Computer Vision. 6924–6934

    2025

  19. [19]

    Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, and Xu Sun. 2025. Spacer: Reinforcing mllms in video spatial reasoning. arXiv preprint arXiv:2504.01805(2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 4035–4045

    2018

  21. [21]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al

  22. [22]

    Yihong Tang, Ao Qu, Zhaokai Wang, Dingyi Zhuang, Zhaofeng Wu, Wei Ma, Shenhao Wang, Yunhan Zheng, Zhan Zhao, and Jinhua Zhao. 2024. Sparkle: Mas- tering basic spatial capabilities in vision language models elicits generalization to composite spatial reasoning. (2024)

    2024

  23. [23]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805(2023)

    Pith/arXiv arXiv 2023

  24. [24]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267(2025)

    Pith/arXiv arXiv 2025

  25. [25]

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai C Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. 2024. Cambrian-1: A fully open, vision-centric exploration of multimodal llms.Ad- vances in Neural Information Processing Systems37 (2024), 87310–87356

    2024

  26. [26]

    Yahan Tu, Rui Hu, and Jitao Sang. 2025. Ode: Open-set evaluation of hallucina- tions in multimodal large language models. InProceedings of the Computer Vision and Pattern Recognition Conference. 19836–19845

    2025

  27. [27]

    Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, et al . 2025. Kimi-vl technical report.arXiv preprint arXiv:2504.07491(2025)

    Pith/arXiv arXiv 2025

  28. [28]

    Hengyi Wang, Ruiqiang Zhang, Chang Liu, Guanjie Wang, Zehua Ma, Han Fang, and Weiming Zhang. 2026. Allocentric Perceiver: Disentangling Allocentric Reasoning from Egocentric Visual Priors via Frame Instantiation.arXiv preprint arXiv:2602.05789(2026)

    arXiv 2026

  29. [29]

    Zehan Wang, Ziang Zhang, Jiayang Xu, Jialei Wang, Tianyu Pang, Chao Du, Heng- shuang Zhao, and Zhou Zhao. 2026. Orient Anything V2: Unifying Orientation and Rotation Understanding.arXiv preprint arXiv:2601.05573(2026)

    arXiv 2026

  30. [30]

    Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys

  31. [31]

    Junjie Wu, Tsz Ting Chung, Kai Chen, and Dit-Yan Yeung. 2024. Unified triplet- level hallucination evaluation for large vision-language models.arXiv preprint arXiv:2410.23114(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    Jiarui Wu, Zhuo Liu, and Hangfeng He. 2025. Mitigating hallucinations in multimodal spatial relations through constraint-aware prompting. InFindings of the Association for Computational Linguistics: NAACL 2025. 3450–3468

    2025

  33. [33]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  34. [34]

    Haoning Wu, Xiao Huang, Yaohui Chen, Ya Zhang, Yanfeng Wang, and Weidi Xie

  35. [35]

    Yuncong Yang, Jiageng Liu, Zheyuan Zhang, Siyuan Zhou, Reuben Tan, Jianwei Yang, Yilun Du, and Chuang Gan. 2025. MindJourney: Test-Time Scaling with World Models for Spatial Reasoning.arXiv preprint arXiv:2507.12508(2025)

    arXiv 2025

  36. [36]

    Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. 2025. Mllms know where to look: Training-free perception of small visual details with multimodal llms.arXiv preprint arXiv:2502.17422(2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Kening Zheng, Junkai Chen, Yibo Yan, Xin Zou, Huiyu Zhou, and Xuming Hu

  38. [38]

    B extends higher in world Y, is farther from the camera, and is offset in top-down view compared with A

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. 2025. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479(2025). Geo3R: Mitigating Spatial Reasoning Hallucination in Multimodal Large Language Models Confer...

    Pith/arXiv arXiv 2025

  39. [39]

    Rui Yang, Ziyu Zhu, Yanwei Li, Jingjia Huang, Shen Yan, Siyuan Zhou, Zhe Liu, Xiangtai Li, Shuangye Li, Wenqian Wang, et al. 2025. Visual spatial tuning.arXiv preprint arXiv:2511.05491(2025)

    arXiv 2025

  40. [43]

    InFindings of the Association for Computational Linguistics: ACL 2025

    Reefknot: A comprehensive benchmark for relation hallucination evalua- tion, analysis and mitigation in multimodal large language models. InFindings of the Association for Computational Linguistics: ACL 2025. 6193–6212. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Wang et al

    2025

  41. [2023]

    InProceedings of the IEEE/CVF international conference on computer vision

    Segment anything. InProceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  42. [2024]

    InEuropean Conference on Computer Vision

    Geocalib: Learning single-image calibration with geometric optimization. InEuropean Conference on Computer Vision. Springer, 1–20

  43. [2025]

    SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards.arXiv preprint arXiv:2511.07403(2025)

    Pith/arXiv arXiv 2025

  44. [2026]

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

    SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 31029–31041