pith. sign in

arxiv: 2606.13673 · v1 · pith:EXBDWUR7new · submitted 2026-06-11 · 💻 cs.CV · cs.AI

SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

Seokju Cho , Ryo Hachiuma , Abhishek Badki , Hang Su , Byung-Kwan Lee , Chan Hee Song , Sifei Liu , Subhashree Radhakrishnan
show 3 more authors
Seungryong Kim Yu-Chiang Frank Wang Min-Hung Chen
This is my paper

Pith reviewed 2026-06-27 06:47 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords spatial reasoningvision-language modelsagentic frameworkscode interfaces3D perceptiontool useiterative reasoning
0
0 comments X

The pith

A stateful code-based action interface enables vision-language agents to achieve higher accuracy on complex 3D and 4D spatial reasoning tasks.

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

The paper argues that the way agents invoke perception tools—the action interface—limits their spatial reasoning ability more than the underlying models do. It proposes that a stateful, step-by-step code execution environment allows agents to build analyses iteratively, reacting to intermediate results. This approach is tested on twenty benchmarks covering static and dynamic scenes in three and four dimensions. If correct, it suggests that interface design can unlock better performance from existing vision-language models without additional training.

Core claim

SpatialClaw is a framework that equips a vision-language model with a stateful Python kernel containing input data and perception primitives. The agent generates one executable code cell at a time, using outputs from previous steps to inform the next, which allows flexible composition of operations tailored to each task's demands in open-ended spatial reasoning.

What carries the argument

The stateful Python kernel pre-loaded with frames and primitives, serving as the action interface that supports iterative, conditioned code execution.

If this is right

  • The agent can revise its spatial analysis strategy based on intermediate visual and textual observations.
  • Performance gains appear across multiple vision-language model families without task-specific tuning.
  • The method applies uniformly to both static image and dynamic video spatial tasks.
  • Overall accuracy reaches 59.9 percent on a suite of twenty benchmarks.

Where Pith is reading between the lines

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

  • Similar stateful interfaces could improve agent performance in other domains requiring sequential decision-making, such as mathematical problem solving.
  • The emphasis on code flexibility might reduce reliance on hand-crafted tool schemas in agent design.
  • Future work could explore scaling the number of primitives available in the kernel to handle more intricate scenes.

Load-bearing premise

The performance advantage results from the stateful per-step code interface itself rather than from variations in prompts, primitive selection, or benchmark characteristics.

What would settle it

Running the same agent with single-pass code execution versus the stateful interface on identical benchmarks while keeping all other components fixed, and observing whether the accuracy difference disappears.

Figures

Figures reproduced from arXiv: 2606.13673 by Abhishek Badki, Byung-Kwan Lee, Chan Hee Song, Hang Su, Min-Hung Chen, Ryo Hachiuma, Seokju Cho, Seungryong Kim, Sifei Liu, Subhashree Radhakrishnan, Yu-Chiang Frank Wang.

Figure 1
Figure 1. Figure 1: SpatialClaw improves spatial reasoning across the board. Per-benchmark accuracy on 20 spatial reasoning benchmarks (Gemma 4-31B backbone), split into two panels by task category. Each axis is individu￾ally rescaled so SpatialClaw traces the constant-radius ring. Baselines are SpaceTools-Toolshed (Chen et al., 2026), pySpatial (Luo et al., 2026), and a no-tool backbone. Abstract Spatial reasoning, the abili… view at source ↗
Figure 2
Figure 2. Figure 2: SpatialClaw studies code as the action interface for spatial reasoning. Three action interfaces on the same question. (a) A single-pass program chooses a complete computation before seeing its intermediate outputs. (b) A structured tool interface exposes common operations through structured commands (e.g., JSON, XML). (c) SpatialClaw writes Python in a persistent kernel, renders intermediate evidence, and … view at source ↗
Figure 3
Figure 3. Figure 3: Agentic loop for iterative code execution. SpatialClaw wraps a persistent kernel in a five-stage loop. A planner receives the question and tool documentation but not the images, and produces an analysis plan. The main agent generates a Python cell executed in the persistent kernel. Feedback comprising stdout, variable summaries, and images registered via show() is appended to the model context. The loop co… view at source ↗
Figure 4
Figure 4. Figure 4: Pairwise win/loss margin of SpatialClaw over baselines across 13 meta-categories. SpatialClaw outperforms both (a) Structured tool-call and (b) Single-pass Code in 11/13 categories. The largest gains concentrate in categories that demand multi-step geometric composition [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Composition adapts to the question type. Primitive usage frequency across meta-categories. To understand how the agent composes tools, we analyze the distribution of primitives (i.e., numpy and scipy operations) invoked across meta-categories ( [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attribution of SpatialClaw’s wins over struc￾tured tool-call via LLM-as-judge. Over half of the gains are driven by code composition, 19.5% by control flow, and 28.3% are interface-neutral wins on perceptual tasks unaffected by the action interface. To identify the main driver of SpatialClaw’s gains over structured tool-call baseline, we examine the instances where SpatialClaw answers correctly but structu… view at source ↗
Figure 7
Figure 7. Figure 7: Failure-mode breakdown of incorrect agent sessions. Each session is classified by an LLM-as-Judge (Gemini-3.1-Pro (Team et al., 2023)) into one of 11 fine￾grained failure categories. In [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
read the original abstract

Spatial reasoning, the ability to determine where objects are, how they relate, and how they move in 3D, remains a fundamental challenge for vision-language models (VLMs). Tool-augmented agents attempt to address this by augmenting VLMs with specialist perception modules, yet their effectiveness is bounded by the action interface through which those tools are invoked. In this work, we study how the design of this interface shapes the agent's capacity for open-ended spatial reasoning. Existing spatial agents either employ single-pass code execution, which commits to a full analysis strategy before any intermediate result is observed, or rely on a structured tool-call interface that often offers less flexibility for freely composing operations or tailoring the analysis to each task. Both designs offer limited flexibility for open-ended, complex 3D/4D spatial reasoning. We therefore propose SpatialClaw, a training-free framework for spatial reasoning that adopts code as the action interface. SpatialClaw maintains a stateful Python kernel pre-loaded with input frames and a suite of perception and geometry primitives, letting a VLM-backed agent write one executable cell per step conditioned on all prior outputs, enabling the agent to flexibly compose and manipulate perception results and adapt its analysis to both intermediate text and visual observations and the demands of each problem. Evaluated across 20 spatial reasoning benchmarks spanning a broad range of static and dynamic 3D/4D spatial reasoning tasks, SpatialClaw achieves 59.9% average accuracy, outperforming the recent spatial agent by +11.2 points, with consistent gains across six VLM backbones from two model families without any benchmark- or model-specific adaptation.

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 proposes SpatialClaw, a training-free framework for agentic spatial reasoning that uses code as the action interface: a stateful Python kernel pre-loaded with input frames and perception/geometry primitives, allowing a VLM to emit one executable cell per step conditioned on prior outputs. This is contrasted with single-pass code execution and structured tool-call interfaces. The central empirical claim is that SpatialClaw attains 59.9% average accuracy across 20 static and dynamic 3D/4D spatial reasoning benchmarks, outperforming a recent spatial agent baseline by +11.2 points with consistent gains across six VLM backbones from two families and no benchmark- or model-specific adaptation.

Significance. If the reported gains can be isolated to the stateful per-step code interface after appropriate controls, the result would provide concrete evidence that interface flexibility matters for open-ended spatial reasoning and could guide the design of future VLM agents. The work is notable for its breadth (20 benchmarks, multiple backbones) and training-free nature.

major comments (1)
  1. [Experiments / Results] The central attribution of the +11.2 point gain to the stateful code interface (as opposed to differences in primitives, prompts, or observation handling) is load-bearing for the headline result, yet the manuscript provides no information on whether the compared recent spatial agent baseline employs identical perception/geometry primitives, identical VLM prompt templates, or the same handling of intermediate visual/text outputs. Without these controls or ablations, the causal claim cannot be verified from the reported numbers alone.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment regarding experimental controls and attribution of gains below.

read point-by-point responses

  1. Referee: [Experiments / Results] The central attribution of the +11.2 point gain to the stateful code interface (as opposed to differences in primitives, prompts, or observation handling) is load-bearing for the headline result, yet the manuscript provides no information on whether the compared recent spatial agent baseline employs identical perception/geometry primitives, identical VLM prompt templates, or the same handling of intermediate visual/text outputs. Without these controls or ablations, the causal claim cannot be verified from the reported numbers alone.

    Authors: We agree that the attribution of gains to the stateful code interface requires explicit controls to rule out confounding factors from primitives, prompts, or observation handling. The baseline comparison follows the original implementation and reported numbers from the cited spatial agent paper, with our reimplementation matching their described perception/geometry primitives and prompt structure as closely as possible. To address this directly, the revised manuscript will add a new subsection detailing the exact primitives, VLM prompt templates, and intermediate output handling used in both SpatialClaw and the baseline. We will also include a controlled ablation that applies our stateful per-step code interface to the baseline's primitives and prompts on a subset of benchmarks to better isolate the interface contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison with no mathematical derivation or self-referential reduction

full rationale

The paper reports experimental results from a training-free agent framework evaluated on 20 benchmarks, claiming accuracy gains over a baseline spatial agent. No equations, fitted parameters, or derivation steps are present in the provided text. Performance attribution rests on direct comparison rather than any self-definitional, fitted-input, or self-citation load-bearing structure. The central claim does not reduce to its inputs by construction; external benchmark results serve as independent evidence. This matches the default expectation for non-circular empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract introduces no free parameters, mathematical axioms, or new invented entities; the framework composes existing VLM backbones and standard perception primitives.

pith-pipeline@v0.9.1-grok · 5876 in / 1151 out tokens · 28566 ms · 2026-06-27T06:47:37.314673+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients

    cs.CL 2026-06 unverdicted novelty 7.0

    ZPPO improves distillation to small vision-language models by using binary and negative candidate prompts plus a replay buffer for hard questions, outperforming standard distillation and GRPO on a 31-benchmark suite w...

Reference graph

Works this paper leans on

58 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    train on the test set

    Ellis Brown, Jihan Yang, Shusheng Yang, Rob Fergus, and Saining Xie. Benchmark designers should “train on the test set” to expose exploitable non-visual shortcuts.ArXiv Preprint, 2025. 7, 8, 13, 14

    2025

  2. [2]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision (ECCV), 2020. 2

    2020

  3. [3]

    Sam 3: Segment anything with concepts

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts. InInternational Conference on Learning Representations (ICLR), 2026. 2, 3, 5, 16, 22

    2026

  4. [4]

    Spa- tialvlm: Endowing vision-language models with spatial reasoning capabilities

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spa- tialvlm: Endowing vision-language models with spatial reasoning capabilities. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 13

    2024

  5. [5]

    Eagle 2.5: Boosting long-context post-training for frontier vision-language models

    Guo Chen, Zhiqi Li, Shihao Wang, Jindong Jiang, Yicheng Liu, Lidong Lu, De-An Huang, Wonmin Byeon, Matthieu Le, Max Ehrlich, Tong Lu, Limin Wang, Bryan Catanzaro, Jan Kautz, Andrew Tao, Zhiding Yu, and Guilin Liu. Eagle 2.5: Boosting long-context post-training for frontier vision-language models. In Advances in Neural Information Processing Systems (NeurI...

    2025

  6. [6]

    Spacetools: Tool-augmented spatial reasoning via double interactive rl

    Siyi Chen, Mikaela Angelina Uy, Chan Hee Song, Faisal Ladhak, Adithyavairavan Murali, Qing Qu, Stan Birchfield, Valts Blukis, and Jonathan Tremblay. Spacetools: Tool-augmented spatial reasoning via double interactive rl. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 1, 2, 3, 4, 8, 9, 13

    2026

  7. [7]

    Spatialrgpt: Grounded spatial reasoning in vision-language models

    An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision-language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 2, 13

    2024

  8. [8]

    4dp-qa: Scalable qa for 4d perception in vision language models

    Seokju Cho, Abhishek Badki, Hang Su, Jindong Jiang, Ziyao Zeng, Seungryong Kim, Sifei Liu, and Orazio Gallo. 4dp-qa: Scalable qa for 4d perception in vision language models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 2

    2026

  9. [9]

    Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction

    Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 7, 8, 9, 13, 14

    2026

  10. [10]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395,

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6):381–395,

  11. [11]

    GPTQ: Accurate post-training compres- sion for generative pretrained transformers.ArXiv Preprint, 2022

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ: Accurate post-training compres- sion for generative pretrained transformers.ArXiv Preprint, 2022. 19

    2022

  12. [12]

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 7, 8, 13, 14

    2025

  13. [13]

    Blink: Multimodal large language models can see but not perceive

    Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision (ECCV), 2024. 7, 8, 9, 13, 14 26 SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

    2024

  14. [14]

    Gemma 4.https://deepmind.google/models/gemma/gemma-4/, 2026

    Google DeepMind. Gemma 4.https://deepmind.google/models/gemma/gemma-4/, 2026. Accessed: 2026-04-14. 2, 3, 4, 7, 8, 9, 19

    2026

  15. [15]

    Visual programming: Compositional visual reasoning without training

    Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 13

    2023

  16. [16]

    Tiger: Tool-integrated geometric reasoning in vision-language models for robotics

    Yi Han, Enshen Zhou, Shanyu Rong, Jingkun An, Pengwei Wang, Zhongyuan Wang, Cheng Chi, Lu Sheng, and Shanghang Zhang. Tiger: Tool-integrated geometric reasoning in vision-language models for robotics. ArXiv Preprint, 2025. 2

    2025

  17. [17]

    and Millman, K

    Charles R. Harris, K. Jarrod Millman, Stéfan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fernández del Río, Mark Wiebe, Pearu Peterson, Pierre Gérard-Marchant, Kevin Shepp...

    work page doi:10.1038/s41586-020-2649-2 2020

  18. [18]

    G2vlm: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning

    Wenbo Hu, Jingli Lin, Yilin Long, Yunlong Ran, Lihan Jiang, Yifan Wang, Chenming Zhu, Runsen Xu, Tai Wang, and Jiangmiao Pang. G2vlm: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 13

    2026

  19. [19]

    J. D. Hunter. Matplotlib: A 2d graphics environment.Computing in Science & Engineering, 9(3):90–95,

  20. [20]

    doi: 10.1109/MCSE.2007.55. 4

    work page doi:10.1109/mcse.2007.55 2007

  21. [21]

    Omnispatial: Towards comprehensive spatial reasoning benchmark for vision language models

    Mengdi Jia, Zekun Qi, Shaochen Zhang, Wenyao Zhang, Xinqiang Yu, Jiawei He, He Wang, and Li Yi. Omnispatial: Towards comprehensive spatial reasoning benchmark for vision language models. In International Conference on Learning Representations (ICLR), 2026. 7, 8, 9, 13, 14

    2026

  22. [22]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles,

  23. [23]

    Viewspatial-bench: Evaluating multi-perspective spatial localization in vision-language models.ArXiv Preprint, 2025

    Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Viewspatial-bench: Evaluating multi-perspective spatial localization in vision-language models.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  24. [24]

    PerceptionComp: A video benchmark for complex perception-centric reasoning.ArXiv Preprint, 2026

    Shaoxuan Li, Zhixuan Zhao, Hanze Deng, Zirun Ma, Shulin Tian, Zuyan Liu, Yushi Hu, Haoning Wu, Yuhao Dong, Benlin Liu, Ziwei Liu, and Ranjay Krishna. PerceptionComp: A video benchmark for complex perception-centric reasoning.ArXiv Preprint, 2026. 7, 8, 13, 14

    2026

  25. [25]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. InInternational Conference on Learning Representations (ICLR), 2026. 3, 5, 16, 21

    2026

  26. [26]

    Mmsi-video-bench: A holistic benchmark for video-based spatial intelligence.ArXiv Preprint, 2025

    Jingli Lin, Runsen Xu, Shaohao Zhu, Sihan Yang, Peizhou Cao, Yunlong Ran, Miao Hu, Chenming Zhu, Yiman Xie, Yilin Long, Wenbo Hu, Dahua Lin, Tai Wang, and Jiangmiao Pang. Mmsi-video-bench: A holistic benchmark for video-based spatial intelligence.ArXiv Preprint, 2025. 7, 8, 13, 14

    2025

  27. [27]

    Octotools: An agentic framework with extensible tools for complex reasoning

    Pan Lu, Bowen Chen, Sheng Liu, Rahul Thapa, Joseph Boen, and James Zou. Octotools: An agentic framework with extensible tools for complex reasoning. InProceedings of the Association for Computational Linguistics (ACL), 2026. 2, 13 27 SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

    2026

  28. [28]

    pyspatial: Generating 3d visual programs for zero-shot spatial reasoning

    Zhanpeng Luo, Ce Zhang, Silong Yong, Cunxi Dai, Qianwei Wang, Haoxi Ran, Guanya Shi, Katia Sycara, and Yaqi Xie. pyspatial: Generating 3d visual programs for zero-shot spatial reasoning. InInternational Conference on Learning Representations (ICLR), 2026. 1, 2, 4, 8, 9, 13

    2026

  29. [29]

    Visual agentic ai for spatial reasoning with a dynamic api

    Damiano Marsili, Rohun Agrawal, Yisong Yue, and Georgia Gkioxari. Visual agentic ai for spatial reasoning with a dynamic api. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2, 4, 7, 8, 9, 13, 14

    2025

  30. [30]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URLhttps://qwen.ai/blog ?id=qwen3.5. 2, 3, 4, 7, 8, 19

    2026

  31. [31]

    Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026

    Qwen Team. Qwen3.6-27B: Flagship-level coding in a 27B dense model, April 2026. URLhttps: //qwen.ai/blog?id=qwen3.6-27b. 3, 7, 8, 19

    2026

  32. [32]

    Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026

    Qwen Team. Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026. URLhttps: //qwen.ai/blog?id=qwen3.6-35b-a3b. 3, 7, 19

    2026

  33. [33]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. In International Conference on Learning Representations (ICLR), 2025. 2

    2025

  34. [34]

    Riemind: Geometry-grounded spatial agent for scene understanding.ArXiv Preprint, 2026

    FernandoRopero, ErkinTurkoz, DanielMatos, JunqingDu, AntonioRuiz, YanfengZhang, LuLiu, Mingwei Sun, and Yongliang Wang. Riemind: Geometry-grounded spatial agent for scene understanding.ArXiv Preprint, 2026. 3, 4, 13

    2026

  35. [35]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. 2, 13

    2023

  36. [36]

    RoboSpatial: Teaching spatial understanding to 2D and 3D vision-language models for robotics

    Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. RoboSpatial: Teaching spatial understanding to 2D and 3D vision-language models for robotics. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 13

    2025

  37. [37]

    Vipergpt: Visual inference via python execution for reasoning

    Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. InIEEE International Conference on Computer Vision (ICCV), 2023. 2, 13

    2023

  38. [38]

    Gemini: afamilyofhighlycapablemultimodal models.ArXiv Preprint, 2023

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk,AndrewMDai,AnjaHauth,KatieMillican,etal. Gemini: afamilyofhighlycapablemultimodal models.ArXiv Preprint, 2023. 15

    2023

  39. [39]

    Gemini robotics: Bringing ai into the physical world.ArXiv Preprint, 2025

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  40. [40]

    Video-mme-v2: Evaluating true understanding and reasoning in video mllms.ArXiv Preprint, 2026

    Video-MME Team. Video-mme-v2: Evaluating true understanding and reasoning in video mllms.ArXiv Preprint, 2026. URLhttps://github.com/Video-MME/Video-MME-v2. 7, 8, 13, 14

    2026

  41. [41]

    E., et al

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stéfan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, İlhan Polat, Yu Feng, Eric W. Mo...

    work page doi:10.1038/s41592-019-0686-2 2020

  42. [42]

    Spatial mental modeling from limited views.ArXiv Preprint, 2025

    Qineng Wang, Baiqiao Yin, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, Saining Xie, Manling Li, Jiajun Wu, and Li Fei-Fei. Spatial mental modeling from limited views.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  43. [43]

    Executable code actions elicit better llm agents

    Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better llm agents. InInternational Conference on Machine Learning (ICML), 2024. 13

    2024

  44. [44]

    𝜋3: Permutation-equivariantvisualgeometrylearning

    Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, ChunhuaShen, andTongHe. 𝜋3: Permutation-equivariantvisualgeometrylearning. InInternational Conference on Learning Representations (ICLR), 2026. 2

    2026

  45. [45]

    From indoor to open world: Revealing the spatial reasoning gap in mllms.ArXiv Preprint, 2025

    Mingrui Wu, Zhaozhi Wang, Fangjinhua Wang, Jiaolong Yang, Marc Pollefeys, and Tong Zhang. From indoor to open world: Revealing the spatial reasoning gap in mllms.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  46. [46]

    Spatialtree: How spatial abilities branch out in mllms.ArXiv Preprint, 2025

    Yuxi Xiao, Longfei Li, Shen Yan, Xinhang Liu, Sida Peng, Yunchao Wei, Xiaowei Zhou, and Bingyi Kang. Spatialtree: How spatial abilities branch out in mllms.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  47. [47]

    Spatialbench: Benchmarking multimodal large language models for spatial cognition.ArXiv Preprint, 2025

    Peiran Xu, Sudong Wang, Yao Zhu, Jianing Li, and Yunjian Zhang. Spatialbench: Benchmarking multimodal large language models for spatial cognition.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  48. [48]

    Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces

    Jihan Yang, Shusheng Yang, Anjali Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2, 4, 14

    2025

  49. [49]

    Mmsi-bench: A benchmark for multi-image spatial intelligence

    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. Mmsi-bench: A benchmark for multi-image spatial intelligence. InICLR, 2025. 7, 8, 9, 13, 14

    2025

  50. [50]

    Geometrically-constrained agent for spatial reasoning.ArXiv Preprint, 2025

    Chen Zeren, Lu Xiaoya, Zheng Zhijie, Li Pengrui, He Lehan, Zhou Yijin, Shao Jing, Zhuang Bohan, and Sheng Lu. Geometrically-constrained agent for spatial reasoning.ArXiv Preprint, 2025. 3, 4, 13

    2025

  51. [51]

    From flatland to space: Teaching vision-language models to perceive and reason in 3d

    Jiahui Zhang, Yurui Chen, Yanpeng Zhou, Yueming Xu, Ze Huang, Jilin Mei, Junhui Chen, Yujie Yuan, Xinyue Cai, Guowei Huang, Xingyue Quan, Hang Xu, and Li Zhang. From flatland to space: Teaching vision-language models to perceive and reason in 3d. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. 7, 8, 9, 13, 14

    2025

  52. [52]

    Make geometry matter for spatial reasoning.ArXiv Preprint, 2026

    Shihua Zhang, Qiuhong Shen, Shizun Wang, Tianbo Pan, and Xinchao Wang. Make geometry matter for spatial reasoning.ArXiv Preprint, 2026. 13

    2026

  53. [53]

    Think3d: Thinking with space for spatial reasoning.ArXiv Preprint,

    Zaibin Zhang, Yuhan Wu, Lianjie Jia, Yifan Wang, Zhongbo Zhang, Yijiang Li, Binghao Ran, Fuxi Zhang, Zhuohan Sun, Zhenfei Yin, et al. Think3d: Thinking with space for spatial reasoning.ArXiv Preprint,

  54. [54]

    Dsi-bench: A benchmark for dynamic spatial intelligence.ArXiv Preprint, 2025

    Ziang Zhang, Zehan Wang, Guanghao Zhang, Weilong Dai, Yan Xia, Ziang Yan, Minjie Hong, and Zhou Zhao. Dsi-bench: A benchmark for dynamic spatial intelligence.ArXiv Preprint, 2025. 7, 8, 9, 13, 14

    2025

  55. [55]

    Pyvision: Agentic vision with dynamic tooling.ArXiv Preprint, 2025

    Shitian Zhao, Haoquan Zhang, Shaoheng Lin, Ming Li, Qilong Wu, Kaipeng Zhang, and Chen Wei. Pyvision: Agentic vision with dynamic tooling.ArXiv Preprint, 2025. 13

    2025

  56. [56]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 15

    2023

  57. [57]

    Pai-bench: Acomprehensive benchmark for physical ai.ArXiv Preprint, 2025

    FengzheZhou, JiannanHuang, JialuoLi, DevaRamanan, andHumphreyShi. Pai-bench: Acomprehensive benchmark for physical ai.ArXiv Preprint, 2025. 7, 8, 9, 13, 14 29 SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

    2025

  58. [58]

    Cvbench: Benchmarking cross-video synergies for complex multimodal reasoning.ArXiv Preprint, 2025

    Nannan Zhu, Yonghao Dong, Teng Wang, Xueqian Li, Shengjun Deng, Yijia Wang, Zheng Hong, Tiantian Geng, Guo Niu, Hanyan Huang, et al. Cvbench: Benchmarking cross-video synergies for complex multimodal reasoning.ArXiv Preprint, 2025. 7, 8, 9, 13, 14 30

    2025