pith. sign in

arxiv: 2606.01247 · v1 · pith:UQOFNDQ4new · submitted 2026-05-31 · 💻 cs.CV

Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?

Liyang Li , Muzhi Zhu , Zhiyue Zhao , Hengyu Zhao , Ke Liu , Linhao Zhong , Hao Chen , Chunhua Shen This is my paper

Pith reviewed 2026-06-28 17:07 UTC · model grok-4.3

classification 💻 cs.CV
keywords target viewpoint reproductionactive explorationfoundation modelsembodied AITVRBenchvisual-action SFTmulti-turn GRPOspatial intelligence
0
0 comments X

The pith

Foundation models reach at most 12% success reproducing a target viewpoint through active 3D movement.

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

The paper introduces Target Viewpoint Reproduction, an active task in which an agent must move its body and head inside a 3D scene until its current view exactly matches a supplied target image. It reports that off-the-shelf open-source and closed-source models succeed on only 7.8% and 12.0% of held-out episodes in a new indoor simulator benchmark. A post-training pipeline that first applies visual-action supervised fine-tuning and then multi-turn group-relative policy optimization lifts a 9B model to 51.4% success. The work therefore tests whether passive image understanding in foundation models can be turned into reliable embodied spatial control.

Core claim

On the evaluation split of TVRBench, strongest open-source and closed-source models reach only 7.8% and 12.0% success at reproducing a target viewpoint. Visual-action SFT raises a 9B open-source model to 50.8% success while multi-turn GRPO from live simulator rollouts reaches 51.4% overall; CoT supervision and single-turn GRPO degrade closed-loop performance. The dominant failure modes are inability to maintain multi-turn visual history and sharp drops when viewpoint change requires body translation rather than in-place rotation.

What carries the argument

The TVR post-training framework that combines expert-trajectory visual-action SFT with on-policy multi-turn GRPO rollouts inside the simulator.

If this is right

  • Multi-turn visual history handling must be improved before reliable active viewpoint reproduction is possible.
  • Mapping observed spatial discrepancies to body translation actions remains a distinct bottleneck from rotation-only control.
  • CoT supervision and single-turn GRPO can reduce closed-loop success even when they improve open-loop metrics.
  • TVRBench functions as a reusable testbed for training and measuring active 3D perception in foundation models.

Where Pith is reading between the lines

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

  • If the simulator-to-real gap is small, the same post-training recipe could be applied directly to mobile robots for view-based navigation.
  • The observed translation bottleneck suggests that future models may need explicit 3D occupancy or depth prediction heads before action generation.
  • Low baseline success indicates that current vision-language models still treat 3D scenes largely as static image collections rather than controllable environments.

Load-bearing premise

Success rates measured inside the TVRBench indoor simulator reflect the spatial intelligence required for real-world embodied tasks.

What would settle it

Run the best post-trained model on a physical robot in a real indoor room and record the fraction of episodes in which the camera view matches the target image within a fixed tolerance after a bounded number of moves.

Figures

Figures reproduced from arXiv: 2606.01247 by Chunhua Shen, Hao Chen, Hengyu Zhao, Ke Liu, Linhao Zhong, Liyang Li, Muzhi Zhu, Zhiyue Zhao.

Figure 1
Figure 1. Figure 1: Target Viewpoint Reproduction (TVR). Can a foundation model actively reproduce a target viewpoint in 3D, closing the perception–reasoning–action loop through body translation and head rotation? Abstract Humans can reproduce the viewpoint speci￾fied by a target image through active head and body motion, yet spatial intelligence in founda￾tion models has largely been studied as pas￾sive understanding of pre-… view at source ↗
Figure 2
Figure 2. Figure 2: TVRBench Task Structure A 2×2 task design crossing scene scale and target-view visual richness. Each category shows one representative task: an orthographic top-down with start (yellow) and target (red) poses, and first-person views at both poses. We label the four categories Single-easy, Single-hard, Multi-easy, Multi-hard. shift the camera horizon by 30◦ . Stop signals task completion and ends the episod… view at source ↗
Figure 3
Figure 3. Figure 3: Why an untrained 9B fails at TVR. Top: Qwen3.5-9B visits only 3.5 distinct grid positions per episode and revisits 83% of poses, producing two stable failure modes—walks in circles (left) and looks in loops (right). Bottom-left: action selection distribution (rotation 50.8%, body translation 26.1%, Stop 0.1%). Bottom-middle: enabling chain-of-thought multiplies tokens per response by ∼10× without changing … view at source ↗
Figure 4
Figure 4. Figure 4: Post-training pipelines on TVRBench. SFT: supervised fine-tuning on rule-based expert trajectories (optionally with CoT). Single-turn RL: GRPO on fixed (It, I⋆ , a∗ t ) prompts. Multi-turn RL: GRPO on on-policy rollouts in TVRBench with dense per-step plus terminal reward. 5 Can Post-Training Improve Active Viewpoint Control? Setup. We use Qwen3.5-9B as the backbone for post-training. For supervised fine-t… view at source ↗
Figure 5
Figure 5. Figure 5: Instructions appended to the per-step SFT user message at step [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: SFT sample format under action-only memory. Each trajectory step becomes one independent single-turn sample with a textual recent-action history; no past observations remain in context. The <think>. . . </think> prefix appears only in CoT variants. “Valid actions at this step:” lists the actions the simulator allows at the current pose. The 1,600 SFT trajectories expand to ≈ 20,700 such per-step samples. p… view at source ↗
Figure 7
Figure 7. Figure 7: SFT sample format under visual-action memory. The entire trajectory is packed into a single multi-turn sample; all past observations remain in context at every step (the SYSTEM prompt is identical to [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: A failure case from the untrained Qwen3.5-9B. With action-only memory, the agent advances twice in its first four steps and then issues 35 consecutive Rotate actions at the same position until the 40-step budget runs out. The action history alone cannot tell the policy it has already tried—and rejected—each yaw, so the same micro-decision repeats indefinitely. For space, the panels show only the first 28 o… view at source ↗
Figure 9
Figure 9. Figure 9: A failure case from the untrained Qwen3.5-9B. On a single-room iTHOR scene the agent shuttles between a handful of cells—issuing 12 Move actions among only four distinct positions—without ever closing the gap to the target. The action history alone cannot register that these cells have already been visited, so the same short walking loop repeats. For space, the panels show only the first 28 of 30 steps [P… view at source ↗
Figure 10
Figure 10. Figure 10: A single-room success from VA-SFT + Multi-turn GRPO. On a single-room iTHOR scene, the policy translates and rotates to align with the target view within a handful of steps and terminates with Stop at the correct pose. Visual-action memory lets each step condition on the actual observation history, so the model no longer revisits previously tried yaws. A Multi-room success from VA-SFT + Multi-turn GRPO ta… view at source ↗
Figure 11
Figure 11. Figure 11: A multi-room success from VA-SFT + Multi-turn GRPO. On a multi-room ProcTHOR scene, the policy traverses the layout across rooms and aligns with the target view before issuing Stop. Multi-room tasks are where the SFT initialisation alone is weakest ( [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
read the original abstract

Humans can reproduce the viewpoint specified by a target image through active head and body motion, yet spatial intelligence in foundation models has largely been studied as passive understanding of pre-collected observations. We introduce Target Viewpoint Reproduction (TVR) -- an active task where an agent adjusts its viewpoint in a 3D environment until its observation matches a given target image -- and TVRBench, an indoor-simulation benchmark spanning scene scale and target-view visual richness. TVR is far from solved: on the evaluation split, the strongest open-source and closed-source models reach only 7.8% and 12.0% success. Fine-grained analysis identifies two consistent bottlenecks: off-the-shelf models struggle with multi-turn visual history, and performance drops sharply when viewpoint reproduction requires body translation rather than in-place rotation, exposing a gap in mapping spatial discrepancies to embodied movement. To study reducing this gap, we build a unified TVR post-training framework covering expert-trajectory SFT, rationale-supervised CoT-SFT, offline Single-turn GRPO, and on-policy Multi-turn GRPO from live simulator rollouts. Visual-action SFT supplies the main gain, raising a 9B open-source model to 50.8% success; Multi-turn GRPO provides targeted multi-room refinement and reaches 51.4% overall, while CoT supervision and Single-turn GRPO degrade closed-loop performance. These results establish TVRBench as a testbed for measuring and training foundation models that actively perceive and act in 3D environments. Our code, data, and models are available at https://github.com/aim-uofa/TVRBench.

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 the Target Viewpoint Reproduction (TVR) task, in which an agent must actively adjust its viewpoint in a 3D indoor environment until its observation matches a provided target image, along with the TVRBench simulator benchmark spanning scene scale and target-view richness. Zero-shot evaluation shows low success (7.8% strongest open-source, 12.0% closed-source). A unified post-training framework is presented that includes expert-trajectory SFT, rationale-supervised CoT-SFT, offline Single-turn GRPO, and on-policy Multi-turn GRPO; visual-action SFT raises a 9B model to 50.8% success while Multi-turn GRPO reaches 51.4% overall. Code, data, and models are released.

Significance. If the performance deltas and failure-mode analysis hold, the work supplies a reproducible testbed and training recipe for active spatial reasoning in foundation models, a capability that has received less attention than passive image understanding. The explicit release of code, data, and models is a concrete strength that supports follow-on research. The reported bottlenecks (multi-turn history handling and translation versus rotation) supply falsifiable targets for subsequent method development.

major comments (1)
  1. [Benchmark construction and evaluation split] Benchmark construction and evaluation split: the manuscript does not supply sufficient detail on how the evaluation scenes and target images were selected to guarantee no overlap with any training trajectories or to balance scene scale and visual richness; without these controls the reported 7.8% and 12.0% zero-shot figures cannot be confidently interpreted as general measures of spatial intelligence rather than simulator-specific artifacts.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation of minor revision. We address the single major comment below.

read point-by-point responses

  1. Referee: Benchmark construction and evaluation split: the manuscript does not supply sufficient detail on how the evaluation scenes and target images were selected to guarantee no overlap with any training trajectories or to balance scene scale and visual richness; without these controls the reported 7.8% and 12.0% zero-shot figures cannot be confidently interpreted as general measures of spatial intelligence rather than simulator-specific artifacts.

    Authors: We agree that the current manuscript provides insufficient detail on these aspects of benchmark construction. In the revised version we will expand the TVRBench section with a dedicated subsection that (1) describes the scene selection criteria used to span scale and target-view richness, (2) specifies how target images were sampled within each scene, and (3) documents the procedure that ensures the evaluation split is disjoint from all trajectories used for expert-trajectory SFT and GRPO training. These additions will allow readers to assess whether the reported zero-shot numbers reflect general spatial capabilities. revision: yes

Circularity Check

0 steps flagged

No circularity; results are direct empirical measurements

full rationale

The paper introduces a new task (TVR) and benchmark (TVRBench), then reports success rates from model evaluations and post-training experiments inside the simulator. These are measured outcomes on held-out splits, not quantities derived from the paper's own equations, fitted parameters renamed as predictions, or self-citation chains. No load-bearing derivation steps exist that reduce to inputs by construction; the central claims are scoped to performance deltas within the described simulator setup.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claims rest on the domain assumption that the indoor simulator faithfully captures the visual and movement dynamics needed to test spatial intelligence, plus the modeling choice that exact image matching is the right success criterion.

axioms (1)
  • domain assumption The indoor simulation environment provides accurate visual observations and movement controls for the agent.
    Benchmark validity depends on the simulator being a reasonable proxy for real 3D spatial tasks.

pith-pipeline@v0.9.1-grok · 5851 in / 1398 out tokens · 33201 ms · 2026-06-28T17:07:04.290730+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

66 extracted references · 19 canonical work pages · 10 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Is a picture worth a thousand words? delving into spatial reasoning for vision language models , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spatialvlm: Endowing vision-language models with spatial reasoning capabilities , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Spatialrgpt: Grounded spatial reasoning in vision-language models , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Transactions of the Association for Computational Linguistics , volume=

    Visual spatial reasoning , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher=

    2023

  5. [5]

    and Girshick, Ross , title =

    Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens and Fei-Fei, Li and Lawrence Zitnick, C. and Girshick, Ross , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  6. [6]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Seeground: See and ground for zero-shot open-vocabulary 3d visual grounding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  7. [7]

    Structural Priors for Vision Workshop at ICCV'25 , year=

    Spatial mental modeling from limited views , author=. Structural Priors for Vision Workshop at ICCV'25 , year=

  8. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    3d concept learning and reasoning from multi-view images , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  9. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Seeing from another perspective: Evaluating multi-view understanding in mllms , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  10. [12]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Thinking in space: How multimodal large language models see, remember, and recall spaces , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  11. [13]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Vlm4d: Towards spatiotemporal awareness in vision language models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  12. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Navgpt: Explicit reasoning in vision-and-language navigation with large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  13. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Openeqa: Embodied question answering in the era of foundation models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [18]

    2017 IEEE international conference on robotics and automation (ICRA) , pages=

    Target-driven visual navigation in indoor scenes using deep reinforcement learning , author=. 2017 IEEE international conference on robotics and automation (ICRA) , pages=. 2017 , organization=

    2017

  15. [19]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  16. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Move to understand a 3d scene: Bridging visual grounding and exploration for efficient and versatile embodied navigation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  17. [24]

    The Fourteenth International Conference on Learning Representations , year=

    Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration? , author=. The Fourteenth International Conference on Learning Representations , year=

  18. [26]

    2026 , eprint=

    ESI-Bench: Towards Embodied Spatial Intelligence that Closes the Perception-Action Loop , author=. 2026 , eprint=

    2026

  19. [30]

    Advances in neural information processing systems , volume=

    Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence , author=. Advances in neural information processing systems , volume=

  20. [35]

    NeurIPS , year=

    Matt Deitke and Eli VanderBilt and Alvaro Herrasti and Luca Weihs and Jordi Salvador and Kiana Ehsani and Winson Han and Eric Kolve and Ali Farhadi and Aniruddha Kembhavi and Roozbeh Mottaghi , title=. NeurIPS , year=

  21. [36]

    arXiv , year=

    Eric Kolve and Roozbeh Mottaghi and Winson Han and Eli VanderBilt and Luca Weihs and Alvaro Herrasti and Daniel Gordon and Yuke Zhu and Abhinav Gupta and Ali Farhadi , title=. arXiv , year=

  22. [40]

    2026 , howpublished=

    MiMo-V2.5-Pro , author=. 2026 , howpublished=

    2026

  23. [41]

    Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S \"u nderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. 2018. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3674--3683

    2018

  24. [42]

    Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. 2020. Objectnav revisited: On evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171

    work page arXiv 2020

  25. [43]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, and 1 others. 2024. _0 : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [44]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455--14465

    2024

  27. [45]

    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 Systems, 37:135062--135093

    2024

  28. [46]

    Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, and Roozbeh Mottaghi. 2022. ProcTHOR: Large-Scale Embodied AI Using Procedural Generation . In NeurIPS. Outstanding Paper Award

    2022

  29. [47]

    Google DeepMind . 2026. https://deepmind.google/models/model-cards/gemini-3-1-pro/ Gemini 3.1 Pro . Model Card

    2026

  30. [48]

    Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 2023. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9202--9212

    2023

  31. [49]

    Yining Hong, Jiageng Liu, Han Yin, Manling Li, Leonidas Guibas, Li Fei-Fei, Jiajun Wu, and Yejin Choi. 2026. https://arxiv.org/abs/2605.18746 Esi-bench: Towards embodied spatial intelligence that closes the perception-action loop . Preprint, arXiv:2605.18746

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  32. [50]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  33. [51]

    Lawrence Zitnick, and Ross Girshick

    Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  34. [52]

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, and 1 others. 2024. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [53]

    Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. 2017. AI2-THOR: An Interactive 3D Environment for Visual AI . arXiv

    2017

  36. [54]

    Juil Koo, Daehyeon Choi, Sangwoo Youn, Phillip Y Lee, and Minhyuk Sung. 2025. Toward ambulatory vision: Learning visually-grounded active view selection. arXiv preprint arXiv:2512.13250

    work page arXiv 2025

  37. [55]

    Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. 2022. Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876

    work page arXiv 2022

  38. [56]

    Rong Li, Shijie Li, Lingdong Kong, Xulei Yang, and Junwei Liang. 2025. Seeground: See and ground for zero-shot open-vocabulary 3d visual grounding. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3707--3717

    2025

  39. [57]

    Zhenyi Liao, Qingsong Xie, Yanhao Zhang, Zijian Kong, Haonan Lu, Zhenyu Yang, and Zhijie Deng. 2025. Improved visual-spatial reasoning via r1-zero-like training. arXiv preprint arXiv:2504.00883

    work page arXiv 2025

  40. [58]

    Fangyu Liu, Guy Emerson, and Nigel Collier. 2023. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635--651

    2023

  41. [59]

    Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. 2022. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474

    work page arXiv 2022

  42. [60]

    Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, and 1 others. 2024. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16488--16498

    2024

  43. [61]

    Qwen Team . 2026 a . https://qwen.ai/blog?id=qwen3.5 Qwen3.5 : Towards native multimodal agents

    2026

  44. [62]

    Qwen Team . 2026 b . https://qwen.ai/blog?id=qwen3.6-27b Qwen3.6-27B : Flagship-level coding in a 27B dense model

    2026

  45. [63]

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

    2026

  46. [64]

    Arijit Ray, Jiafei Duan, Ellis Brown, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, and 1 others. 2024. Sat: Dynamic spatial aptitude training for multimodal language models. arXiv preprint arXiv:2412.07755

    work page arXiv 2024

  47. [65]

    Koya Sakamoto, Taiki Miyanishi, Daichi Azuma, Shuhei Kurita, Shu Morikuni, Naoya Chiba, Motoaki Kawanabe, Yusuke Iwasawa, and Yutaka Matsuo. 2026. E3vs-bench: A benchmark for viewpoint-dependent active perception in 3d gaussian splatting scenes. arXiv preprint arXiv:2604.17969

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  48. [66]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [67]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, and 1 others. 2025. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  50. [68]

    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. Advances in Neural Information Processing Systems, 37:75392--75421

    2024

  51. [69]

    Diankun Wu, Fangfu Liu, Yi-Hsin Hung, and Yueqi Duan. 2026. Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence. Advances in neural information processing systems, 38:13569--13597

    2026

  52. [70]

    Xiaomi MiMo Team . 2026. Mimo-v2.5-pro. https://huggingface.co/collections/XiaomiMiMo/mimo-v25

    2026

  53. [71]

    Runsen Xu, Weiyao Wang, Hao Tang, Xingyu Chen, Xiaodong Wang, Fu-Jen Chu, Dahua Lin, Matt Feiszli, and Kevin J Liang. 2025. Multi-spatialmllm: Multi-frame spatial understanding with multi-modal large language models. arXiv preprint arXiv:2505.17015

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  54. [72]

    Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. 2025 a . Thinking in space: How multimodal large language models see, remember, and recall spaces. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 10632--10643

    2025

  55. [73]

    Sihan Yang, Runsen Xu, Yiman Xie, Sizhe Yang, Mo Li, Jingli Lin, Chenming Zhu, Xiaochen Chen, Haodong Duan, Xiangyu Yue, and 1 others. 2025 b . Mmsi-bench: A benchmark for multi-image spatial intelligence. arXiv preprint arXiv:2505.23764

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  56. [74]

    Chun-Hsiao Yeh, Chenyu Wang, Shengbang Tong, Ta-Ying Cheng, Ruoyu Wang, Tianzhe Chu, Yuexiang Zhai, Yubei Chen, Shenghua Gao, and Yi Ma. 2026. Seeing from another perspective: Evaluating multi-view understanding in mllms. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 12000--12008

    2026

  57. [75]

    Baiqiao Yin, Qineng Wang, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, and 1 others. 2025. Spatial mental modeling from limited views. In Structural Priors for Vision Workshop at ICCV'25

    2025

  58. [76]

    Heyang Yu, Yinan Han, Xiangyu Zhang, Baiqiao Yin, Bowen Chang, Xiangyu Han, Xinhao Liu, Jing Zhang, Marco Pavone, Chen Feng, and 1 others. 2025. Thinking in 360 \ deg \ : Humanoid visual search in the wild. arXiv preprint arXiv:2511.20351

    work page arXiv 2025

  59. [77]

    Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. 2024. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv preprint arXiv:2406.10721

    work page arXiv 2024

  60. [78]

    Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. 2024. Poliformer: Scaling on-policy rl with transformers results in masterful navigators. arXiv preprint arXiv:2406.20083

    work page arXiv 2024

  61. [79]

    Pingyue Zhang, Zihan Huang, Yue Wang, Jieyu Zhang, Letian Xue, Zihan Wang, Qineng Wang, Keshigeyan Chandrasegaran, Ruohan Zhang, Yejin Choi, and 1 others. 2026. Theory of space: Can foundation models construct spatial beliefs through active exploration? In The Fourteenth International Conference on Learning Representations

    2026

  62. [80]

    Gengze Zhou, Yicong Hong, and Qi Wu. 2024. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7641--7649

    2024

  63. [81]

    Shijie Zhou, Alexander Vilesov, Xuehai He, Ziyu Wan, Shuwang Zhang, Aditya Nagachandra, Di Chang, Dongdong Chen, Xin Eric Wang, and Achuta Kadambi. 2025. Vlm4d: Towards spatiotemporal awareness in vision language models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8600--8612

    2025

  64. [82]

    Muzhi Zhu, Hao Zhong, Canyu Zhao, Zongze Du, Zheng Huang, Mingyu Liu, Hao Chen, Cheng Zou, Jingdong Chen, Ming Yang, and 1 others. 2025 a . Active-o3: Empowering multimodal large language models with active perception via grpo. arXiv preprint arXiv:2505.21457

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  65. [83]

    Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. 2017. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3357--3364. ieee

    2017

  66. [84]

    Ziyu Zhu, Xilin Wang, Yixuan Li, Zhuofan Zhang, Xiaojian Ma, Yixin Chen, Baoxiong Jia, Wei Liang, Qian Yu, Zhidong Deng, and 1 others. 2025 b . 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, pages 8120--8132

    2025