pith. sign in

arxiv: 2607.00491 · v1 · pith:ADUVKJQLnew · submitted 2026-07-01 · 💻 cs.CV · cs.AI· cs.CL

MindEdit-Bench: Benchmarking Object-Level Counterfactual Spatial Reasoning in VLMs from In-the-Wild Photos

Leyuan Yu , Xiao Tang , Minghao Liu , Xinyuan Li , Xiaokai Bai , Sheng Zhou , Qunshu Lin , Weihao Xuan
show 1 more author
Naoto Yokoya
This is my paper

Pith reviewed 2026-07-02 14:58 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.CL
keywords vision-language modelsspatial reasoningcounterfactual reasoningbenchmarkobject editingindoor scenesperspective transformationvisibility reasoning
0
0 comments X

The pith

Vision-language models reach only 8-31% accuracy on tasks requiring them to predict the effects of moving or rotating objects in real indoor photos, while humans reach 81-97%.

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

The paper presents MindEdit-Bench, a set of six tasks that test whether VLMs can perform object-level counterfactual spatial reasoning on photos taken in everyday rooms. Four tasks check perception of existing structure and viewpoint changes; two new tasks ask models to imagine the results of editing an object's position or its visibility from another angle, answers that cannot be read off the input images. The benchmark is built automatically from three-photo triplets of 120 private scenes, with 8-24 structured choices per question to expose specific spatial mistakes. Evaluation of 15 models shows consistently low accuracy and a 53-point pooled gap versus human majority votes, with at least 39 points on every task.

Core claim

MindEdit-Bench shows that current VLMs lack reliable object-level counterfactual spatial reasoning: on L4 spatial-editing and L5 cross-view visibility-editing questions the correct answers are absent from all input images, yet models achieve only single-digit to low-thirty percent accuracy while human majority votes reach 81-97 percent, with the structured answer format exposing uneven failures such as weaker inference along the camera depth axis.

What carries the argument

MindEdit-Bench, built from three-photo smartphone triplets via automatic 3D scene-graph extraction, supplies the six tasks and 8-24 structured answer choices that let the evaluation isolate spatial versus fallback errors on counterfactual object edits.

If this is right

  • VLMs exhibit non-uniform spatial weaknesses, including poorer performance on depth-axis inferences and fallback errors on visibility-editing cases.
  • The 53-point human-best-VLM gap holds across all six tasks and persists after controlling for public-data overlap by using private indoor scenes.
  • Structured multiple-choice format with 8-24 options enables per-question diagnosis of whether models fail on perspective transformation, object permanence, or simple pattern matching.
  • Four perception tasks already show large gaps, indicating that even non-counterfactual spatial reasoning remains unreliable before counterfactual demands are added.

Where Pith is reading between the lines

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

  • Training regimes that rely only on observational image-text pairs may be insufficient to produce robust counterfactual object reasoning.
  • Robotics or augmented-reality systems that must simulate object rearrangements would need additional 3D simulation or explicit spatial modules beyond current VLMs.
  • Extending the benchmark to outdoor scenes or video sequences could test whether the observed gaps generalize beyond controlled indoor triplets.

Load-bearing premise

The automatic 3D scene-graph pipeline produces accurate ground-truth labels and counterfactual answers for the editing tasks without systematic artifacts that would inflate the human-VLM gap.

What would settle it

A manual re-annotation of the L4 and L5 questions by multiple humans that yields VLM accuracies above 60 percent on the same items would falsify the reported capability gap.

Figures

Figures reproduced from arXiv: 2607.00491 by Leyuan Yu, Minghao Liu, Naoto Yokoya, Qunshu Lin, Sheng Zhou, Weihao Xuan, Xiaokai Bai, Xiao Tang, Xinyuan Li.

Figure 1
Figure 1. Figure 1: MindEdit-Bench overview. MindEdit-Bench probes object-level counterfactual spatial reasoning in VLMs, built from three-photo smartphone triplets sampled from privately captured indoor scenes. (a) An automatic feed-forward pipeline reconstructs an object-centric 3D scene graph. (b) An L4 spatial-editing question whose correct answer is absent from all input photos. (c) The six tasks (L1–L5) span three cogni… view at source ↗
Figure 2
Figure 2. Figure 2: 3D Scene-Graph Extraction Pipeline. The two-stage feed-forward pipeline takes three smartphone photos and produces an object-centric 3D scene graph that feeds the question generator (§3). Stage 1: per-view 2D perception. DA3 produces a metric depth map and camera pose for each photo; SAM3, prompted by an LLM-generated object vocabulary (e.g., sofa, table, lamp), produces per-view object masks. Stage 2: cro… view at source ↗
Figure 3
Figure 3. Figure 3: Per-model accuracy scoreboard across six task types. Rows are 15 VLMs ranked by overall accuracy; [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Task-level human–SOTA gaps. Coloured segments show the gap from the best model to human majority vote; random chance and the 15-model mean are shown as context. α = 0.82 (nominal, overall) indicates substantial inter-annotator agreement; we use majority vote as the human label. Answer permutation. To verify that model an￾swers are not driven by letter position, we shuffle the option-letter ordering and re-… view at source ↗
Figure 5
Figure 5. Figure 5: Representative examples for the four perception and perspective-transformation tasks (L1–L3b). Each [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative examples for the two counterfactual editing tasks (L4 and L5). Same panel layout as [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Answer-permutation control. Each panel shows task-level accuracy for three representative VLMs under the main run (rose stars and dashed guides) and three option-letter shuffles (blue points connected by range lines). Reordering option letters changes overall accuracy by at most 1.4 pp, while the largest task-level fluctuation is 6.9 pp. RoboBrain2.5-8B selects this option 0% of the time, reflecting an opp… view at source ↗
read the original abstract

Benchmarks for vision-language models (VLMs) mostly test observational spatial reasoning: models describe relations already visible in the input. Existing what-if tasks typically vary the observer while keeping the scene fixed. Can VLMs instead predict the consequences of hypothetically moving or rotating an object? We introduce MindEdit-Bench, a benchmark of six spatial reasoning tasks built from three-photo smartphone triplets of newly captured indoor scenes via an automatic in-the-wild 3D scene-graph extraction pipeline. Four tasks probe perception and perspective transformation over observed structure; two new tasks, L4 (spatial editing) and L5 (cross-view visibility editing), probe object-level counterfactual reasoning, where correct answers are absent from all input images. Each question provides 8-24 structured answer choices, enabling answer-letter-level diagnosis of spatial and fallback errors. The benchmark covers 120 private indoor scenes not drawn from public datasets, reducing public-data pretraining-overlap risk. Across 15 VLMs on 1,003 human-verified questions, task-wise mean VLM accuracy is only 8%-31%, versus 81%-97% human majority-vote accuracy. The pooled human--best-VLM gap is 53 pp, with at least 39 pp on every task. The structured answer space further reveals non-uniform failures, including weaker camera-depth-axis inference and fallback behavior on difficult visibility-editing cases.

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 / 2 minor

Summary. The paper introduces MindEdit-Bench, a benchmark of six spatial reasoning tasks constructed from three-photo smartphone triplets of 120 private indoor scenes using an automatic in-the-wild 3D scene-graph extraction pipeline. Four tasks address perception and perspective transformation on observed structure; L4 (spatial editing) and L5 (cross-view visibility editing) require object-level counterfactual reasoning where correct answers are absent from all input images. Each question offers 8-24 structured answer choices. On 1,003 human-verified questions, 15 VLMs achieve task-wise mean accuracies of 8%-31% versus 81%-97% for human majority votes, yielding a pooled gap of 53 pp (at least 39 pp per task). The structured answers enable diagnosis of specific spatial and fallback errors.

Significance. If the 3D pipeline produces reliable counterfactual ground truth, the benchmark would establish a clear, large gap between human and VLM performance on object-level counterfactual spatial reasoning from in-the-wild photos. Strengths include the use of private scenes to limit pretraining overlap, the multi-choice format for fine-grained error analysis, and the focus on editing tasks absent from prior observational benchmarks. Such a result would be relevant for robotics, AR, and planning applications where models must simulate hypothetical object manipulations.

major comments (1)
  1. [Methods (automatic in-the-wild 3D scene-graph extraction pipeline description)] The central claim that VLMs exhibit 39-53 pp gaps on L4/L5 counterfactual tasks rests on the accuracy of the automatic 3D scene-graph extraction pipeline for inferring object poses, depths, cross-view visibilities, and generating ground-truth answers absent from input photos. No quantitative validation (reconstruction error, human agreement on generated counterfactual labels, or error analysis restricted to L4/L5) is reported, so it is impossible to rule out systematic artifacts from camera calibration, occlusion handling, or 3D lifting that would inflate the reported human-VLM gap.
minor comments (2)
  1. [Abstract] The abstract reports 'task-wise mean VLM accuracy is only 8%-31%' without referencing the specific table or figure that breaks this down per task and per model; adding that pointer would improve traceability.
  2. [Abstract] The human verification process for the 1,003 questions is mentioned but the number of annotators per question and the exact majority-vote threshold are not stated in the abstract; these details belong in the main text or a methods subsection.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the importance of validating the 3D scene-graph extraction pipeline. We address the major comment below and commit to revisions that strengthen the manuscript's claims.

read point-by-point responses

  1. Referee: The central claim that VLMs exhibit 39-53 pp gaps on L4/L5 counterfactual tasks rests on the accuracy of the automatic 3D scene-graph extraction pipeline for inferring object poses, depths, cross-view visibilities, and generating ground-truth answers absent from input photos. No quantitative validation (reconstruction error, human agreement on generated counterfactual labels, or error analysis restricted to L4/L5) is reported, so it is impossible to rule out systematic artifacts from camera calibration, occlusion handling, or 3D lifting that would inflate the reported human-VLM gap.

    Authors: We agree this is a substantive limitation. The manuscript does not report quantitative metrics such as reconstruction error, human agreement rates on the generated counterfactual labels for L4/L5, or a dedicated error analysis of the pipeline outputs. While all 1,003 questions received human verification to confirm the final question-answer pairs, this does not directly quantify pipeline accuracy on object poses, depths, or cross-view visibilities. To address the concern, we will add a new subsection in the Methods and an appendix with (1) human agreement statistics on a random sample of 200 L4/L5 ground-truth labels, (2) qualitative error analysis of pipeline failures on occluded or depth-ambiguous cases, and (3) discussion of how such errors could affect the reported gaps. These additions will allow readers to assess whether systematic artifacts are present. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark with direct accuracy measurements

full rationale

The paper introduces MindEdit-Bench as an empirical evaluation of VLMs on spatial reasoning tasks using human-verified questions from in-the-wild scenes. No derivations, equations, fitted parameters, or predictions are claimed that reduce to the authors' own inputs or prior self-citations. Task accuracies are computed directly from model outputs against ground-truth labels and human majority votes, with no self-definitional loops, fitted-input predictions, or load-bearing uniqueness theorems. The automatic 3D pipeline is an implementation detail whose accuracy is an external validity concern, not a circular reduction within the reported results.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an empirical benchmark paper. No mathematical derivations, fitted constants, or postulated physical entities are introduced. The central claims rest on the correctness of the automatic 3D scene-graph pipeline and the representativeness of the 120 private scenes, but these are not formalized as axioms or free parameters in the provided abstract.

pith-pipeline@v0.9.1-grok · 5816 in / 1239 out tokens · 24265 ms · 2026-07-02T14:58:44.016273+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

48 extracted references · 17 canonical work pages · 8 internal anchors

  1. [1]

    Chen, Boyuan and Xu, Zhuo and Kirmani, Sean and Ichter, Brian and Sadigh, Dorsa and Guibas, Leonidas and Xia, Fei , booktitle =

  2. [2]

    Cheng, An-Chieh and Yin, Hongxu and Fu, Yang and Guo, Qiushan and Yang, Ruihan and Kautz, Jan and Wang, Xiaolong and Liu, Sifei , booktitle =

  3. [3]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  4. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  5. [10]

    Ma, Xiaojian and Yong, Silong and Zheng, Zilong and Li, Qing and Liang, Yitao and Zhu, Song-Chun and Huang, Siyuan , booktitle =

  6. [14]

    Science , volume =

    Mental Rotation of Three-Dimensional Objects , author =. Science , volume =

  7. [15]

    1983 , publisher =

    Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness , author =. 1983 , publisher =

    1983

  8. [16]

    European Conference on Spatial Information Theory (COSIT) , pages =

    Cognitive Maps, Cognitive Collages, and Spatial Mental Models , author =. European Conference on Spatial Information Theory (COSIT) , pages =

  9. [17]

    Trends in Cognitive Sciences , volume =

    Mechanical Reasoning by Mental Simulation , author =. Trends in Cognitive Sciences , volume =

  10. [18]

    Proceedings of the National Academy of Sciences (PNAS) , volume =

    Simulation as an Engine of Physical Scene Understanding , author =. Proceedings of the National Academy of Sciences (PNAS) , volume =

  11. [19]

    Trends in Cognitive Sciences , volume =

    Counterfactual Simulation in Causal Cognition , author =. Trends in Cognitive Sciences , volume =. 2024 , doi =

    2024

  12. [20]

    and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie

    Dai, Angela and Chang, Angel X. and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  13. [23]

    2026 , url =

    System Card:. 2026 , url =

    2026

  14. [30]

    Anthropic . 2026. https://www.anthropic.com/system-cards/ System card: Claude Opus 4.6 . Technical report, Anthropic

    2026

  15. [31]

    Battaglia, Jessica B

    Peter W. Battaglia, Jessica B. Hamrick, and Joshua B. Tenenbaum. 2013. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences (PNAS), 110(45):18327--18332

    2013

  16. [32]

    ByteDance Seed . 2026. https://seed.bytedance.com/en/seed2 Seed2.0 model card

    2026

  17. [33]

    Zhongang Cai, Ruisi Wang, Chenyang Gu, Fanyi Pu, Junxiang Xu, Yubo Wang, Wanqi Yin, Zhitao Yang, Chen Wei, Qingping Sun, Tongxi Zhou, Jiaqi Li, Hui En Pang, Oscar Qian, Yukun Wei, Zhiqian Lin, Xuanke Shi, Kewang Deng, Xiaoyang Han, and 10 others. 2025 a . https://doi.org/10.48550/arXiv.2511.13719 Scaling spatial intelligence with multimodal foundation mod...

    work page doi:10.48550/arxiv.2511.13719 2025

  18. [34]

    Zhongang Cai, Yubo Wang, Qingping Sun, Ruisi Wang, Chenyang Gu, Wanqi Yin, Zhiqian Lin, Zhitao Yang, Chen Wei, Oscar Qian, Hui En Pang, Xuanke Shi, Kewang Deng, Xiaoyang Han, Zukai Chen, Jiaqi Li, Xiangyu Fan, Hanming Deng, Lewei Lu, and 5 others. 2025 b . https://doi.org/10.48550/arXiv.2508.13142 Holistic evaluation of multimodal LLM s on spatial intelli...

    work page doi:10.48550/arxiv.2508.13142 2025

  19. [35]

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R \"a dle, and 19 others. 2025. https://doi.org/10.48550/arXiv.2511.16719 SAM 3: Segment anythi...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.16719 2025

  20. [36]

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. 2024. SpatialVLM : Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  21. [37]

    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. In Advances in Neural Information Processing Systems (NeurIPS)

    2024

  22. [38]

    Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nie ner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nie ner. 2017. ScanNet : Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  23. [39]

    Yuanyuan Gao, Hao Li, Yifei Liu, Xinhao Ji, Yuning Gong, Yuanjun Liao, Fangfu Liu, Manyuan Zhang, Yuchen Yang, Dan Xu, Xue Yang, Huaxi Huang, Hongjie Zhang, Ziwei Liu, Xiao Sun, Dingwen Zhang, and Zhihang Zhong. 2026. https://doi.org/10.48550/arXiv.2603.07660 Holi-Spatial : Evolving video streams into holistic 3d spatial intelligence . arXiv preprint arXi...

    work page doi:10.48550/arxiv.2603.07660 2026

  24. [40]

    Tobias Gerstenberg. 2024. https://doi.org/10.1016/j.tics.2024.04.012 Counterfactual simulation in causal cognition . Trends in Cognitive Sciences, 28(10):924--936

    work page doi:10.1016/j.tics.2024.04.012 2024

  25. [41]

    Google DeepMind . 2026. https://deepmind.google/models/model-cards/gemini-3-1-pro Gemini 3.1 Pro model card . Technical report, Google DeepMind

    2026

  26. [42]

    Mary Hegarty. 2004. Mechanical reasoning by mental simulation. Trends in Cognitive Sciences, 8(6):280--285

    2004

  27. [43]

    Johnson-Laird

    Philip N. Johnson-Laird. 1983. Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness. Harvard University Press

    1983

  28. [44]

    Kimi Team . 2026. https://doi.org/10.48550/arXiv.2602.02276 Kimi K2.5 : Visual agentic intelligence . arXiv preprint arXiv:2602.02276

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.02276 2026

  29. [45]

    Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. 2025. https://doi.org/10.48550/arXiv.2505.21500 ViewSpatial-Bench : Evaluating multi-perspective spatial localization in vision-language models . arXiv preprint arXiv:2505.21500

    work page doi:10.48550/arxiv.2505.21500 2025

  30. [46]

    Depth Anything 3: Recovering the Visual Space from Any Views

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. 2025. https://doi.org/10.48550/arXiv.2511.10647 Depth anything 3: Recovering the visual space from any views . arXiv preprint arXiv:2511.10647

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.10647 2025

  31. [47]

    Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. 2023. SQA3D : Situated question answering in 3d scenes. In International Conference on Learning Representations (ICLR)

    2023

  32. [48]

    OpenAI . 2026. https://openai.com/index/gpt-5-4-thinking-system-card/ GPT-5.4 thinking system card

    2026

  33. [49]

    Qwen Team . 2026 a . https://huggingface.co/Qwen/Qwen3.5-35B-A3B Qwen3.5-35B-A3B model card

    2026

  34. [50]

    Qwen Team . 2026 b . https://qwen.ai/blog?id=qwen3.6 Qwen3.6-Plus : Towards real world agents

    2026

  35. [51]

    Qwen Team, Alibaba . 2025. https://doi.org/10.48550/arXiv.2511.21631 Qwen3-VL technical report . arXiv preprint arXiv:2511.21631

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631 2025

  36. [52]

    Shepard and Jacqueline Metzler

    Roger N. Shepard and Jacqueline Metzler. 1971. Mental rotation of three-dimensional objects. Science, 171(3972):701--703

    1971

  37. [53]

    Huajie Tan, Enshen Zhou, Zhiyu Li, Yijie Xu, Yuheng Ji, Xiansheng Chen, Cheng Chi, Pengwei Wang, Huizhu Jia, Yulong Ao, Mingyu Cao, Sixiang Chen, Zhe Li, Mengzhen Liu, Zixiao Wang, Shanyu Rong, Yaoxu Lyu, Zhongxia Zhao, Peterson Co, and 16 others. 2026. https://doi.org/10.48550/arXiv.2601.14352 RoboBrain 2.5 : Depth in sight, time in mind . arXiv preprint...

    work page doi:10.48550/arxiv.2601.14352 2026

  38. [54]

    Barbara Tversky. 1993. Cognitive maps, cognitive collages, and spatial mental models. In European Conference on Spatial Information Theory (COSIT), pages 14--24

    1993

  39. [55]

    Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. 2024. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In Advances in Neural Information Processing Systems (NeurIPS)

    2024

  40. [56]

    Qineng Wang, Baiqiao Yin, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, Saining Xie, Jiajun Wu, Li Fei-Fei, and Manling Li. 2025. https://doi.org/10.48550/arXiv.2506.21458 MindCube : Spatial mental modeling from limited views . arXiv preprint arXiv:2506.21458

    work page doi:10.48550/arxiv.2506.21458 2025

  41. [57]

    Haoning Wu, Xiao Huang, Yaohui Chen, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2025. https://doi.org/10.48550/arXiv.2505.17012 SpatialScore : Towards comprehensive evaluation for spatial intelligence . arXiv preprint arXiv:2505.17012

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.17012 2025

  42. [58]

    Yicheng Xiao, Wenhu Zhang, Lin Song, Yukang Chen, Wenbo Li, Nan Jiang, Tianhe Ren, Haokun Lin, Wei Huang, Haoyang Huang, Xiu Li, Nan Duan, and Xiaojuan Qi. 2026. https://doi.org/10.48550/arXiv.2604.04911 SpatialEdit : Benchmarking fine-grained image spatial editing . arXiv preprint arXiv:2604.04911

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.04911 2026

  43. [59]

    Yuxi Xiao, Longfei Li, Shen Yan, Xinhang Liu, Sida Peng, Yunchao Wei, Xiaowei Zhou, and Bingyi Kang. 2025. https://doi.org/10.48550/arXiv.2512.20617 SpatialTree : How spatial abilities branch out in MLLM s . arXiv preprint arXiv:2512.20617

    work page doi:10.48550/arxiv.2512.20617 2025

  44. [60]

    Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie

    Jihan Yang, Shusheng Yang, Anjali W. Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. 2025 a . https://openaccess.thecvf.com/content/CVPR2025/html/Yang_Thinking_in_Space_How_Multimodal_Large_Language_Models_See_Remember_CVPR_2025_paper.html Thinking in space: How multimodal large language models see, remember, and recall spaces . In Proceedings of the IEEE/...

    2025

  45. [61]

    Rui Yang, Ziyu Zhu, Yanwei Li, Jingjia Huang, Shen Yan, Siyuan Zhou, Zhe Liu, Xiangtai Li, Shuangye Li, Wenqian Wang, Yi Lin, and Hengshuang Zhao. 2025 b . https://doi.org/10.48550/arXiv.2511.05491 Visual spatial tuning . arXiv preprint arXiv:2511.05491

    work page doi:10.48550/arxiv.2511.05491 2025

  46. [62]

    Sihan Yang, Runsen Xu, Yiman Xie, Sizhe Yang, Mo Li, Jingli Lin, Chenming Zhu, Xiaochen Chen, Haodong Duan, Xiangyu Yue, Dahua Lin, Tai Wang, and Jiangmiao Pang. 2025 c . https://doi.org/10.48550/arXiv.2505.23764 MMSI-Bench : A benchmark for multi-image spatial intelligence . arXiv preprint arXiv:2505.23764

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.23764 2025

  47. [63]

    Z.ai . 2025. https://huggingface.co/zai-org/GLM-4.6V-Flash GLM-4.6V-Flash model card

    2025

  48. [64]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, and 32 others. 2025. https://doi.org/10.48550/arXiv.2504.10479 InternVL3 : Exploring advanced training and test-time recipes f...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.10479 2025