Computer Vision and Pattern Recognition · arXiv:2610.13832 v1 · submitted 9 Oct 2026 · 85 references
SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
The paper says
“When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. But can Large Vision-Language Models (LVLMs) proactively think in the same way?”
more of the abstract
We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. To bridge this gap, we introduce SpatialASM, a scalable framework to internalize ASM into LVLMs through training-time spatial simulation. SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision: i) explicit ASM trajectories that teach LVLMs when and how to adaptively invoke view planning and imagination during spatial reasoning, and ii) automatically-constructed perspective-taking tasks that further strengthens viewpoint-conditioned scene imagination. On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities. The experimental results demonstrate that our approach improves average accuracy across the seven spatial benchmarks. And the synthesis of ASM trajectories provides a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision for eliciting active viewpoint exploration and mental scene simulation.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 12 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Solid, honest post-training recipe for spatial reasoning; the benchmark gains are believable, but the 'internalized ASM' mechanism is asserted more strongly than the trajectory filter actually verifies. This line is wrong
4 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 85 references say.
The paper says, at Sec. 3.2 (Eqs. 7-8), B.2.2
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The retention criterion in Eq. (7) never verifies the spatial correctness of r_plan, C_edit, or r_reason; it only checks that the teacher answers incorrectly from the original images, correctly after seeing the renders, and that every render has GeoP below the threshold.
more of objection 1
Because Eq. (8) replaces the rendered images with captions written by a different model (Qwen3.6-27B, B.2.2), a lossy or incorrect caption can carry a conclusion that only the removed image justified, and on 4-choice questions a wrong-to-right flip can occur by chance in about a quarter of cases where the renders are uninformative. The authors acknowledge that GeoP is a proxy (Sec. 3.1), but that acknowledgment does not address the missing check on the spatial content of the trajectories. The manuscript should add a direct audit of the retained trajectories (human annotation or an independent geometric checker on r_plan, C_edit, and r_reason) and an experiment showing that the benchmark gains depend on spatially correct ASM content rather than on 66k answer-correct reasoning samples.
The paper says, at Tab. 1, Sec. 4.2
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
All reported accuracies are single runs without error bars, confidence intervals, or repeated-seed statistics, so the central 'consistent improvement' claim is not yet statistically established. On Qwen3.5-9B, one of the seven benchmarks (3DSR) decreases by 0.9 points, and the average gains range from +1.2 to +3.5 points; without variance information it is unclear which per-benchmark differences are reliable.
more of objection 2
The authors should report multiple seeds, or at least bootstrap or paired evaluations, for the main table and for the statement that general-VQA performance stays within 0.6 points.
The paper says, at Sec. 4.2, Appendix D.3
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
Appendix D.3 states that SpatialASM-8B rarely emits the explicit plan-imagine-reason pattern in its outputs, yet the paper attributes the benchmark gains to internalized ASM. This underdetermines the mechanism: ten epochs of SFT on 66k answer-correct samples could improve spatial benchmarks by teaching a general reasoning style or answer-format imitation.
more of objection 3
The ablation in Tab. 4 lacks a control trained on matched CoT data without ASM content, so the specific contributions of view planning and view imagination are not isolated. Adding such a control, or showing that gains correlate with verified ASM behavior, would make the mechanism claim supported.
The paper says, at Sec. 3.1, B.1.3
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The GeoP filter is the only geometric check on rendered views, and the authors correctly describe GeoP as a proxy. However, B.1.3 explicitly encourages the NVG to imagine objects not present in the input images, provided they are plausible, which weakens the claim that hypothetical viewpoints are grounded in visual evidence.
more of objection 4
A rendered view that is plausible but not geometrically determined can pass GeoP while contributing spatially wrong captions to the SFT supervision. The paper should quantify how often retained trajectories contain captions describing objects or layouts not supported by the input views, and discuss the effect of such content on the distilled reasoning behavior.
5 minor objections
The paper says, at Sec. 4.2
Pith says: minor objection 1 of 5
The word 'degredation' should be 'degradation'.
The paper says, at B.2.1
Pith says: minor objection 2 of 5
The typo 'Mindtube-train' should be 'MindCube-train', and the sentence should make explicit that the 66k SFT subset contains no MindCube-train samples while the full 128k set does.
show 3 more
The paper says, at Tab. 2
Pith says: minor objection 3 of 5
The table layout is confusing: the Random-Plan row appears to be attached to the header block of the results, and the definitions of the two reported times (T and T_NVG) should be stated in the caption.
The paper says, at Tab. 1 and text throughout
Pith says: minor objection 4 of 5
Benchmark name capitalization is inconsistent (Mindcube vs MindCube, Viewspa vs ViewSpa, Embspa vs EmbSpa); please unify the names in the table and text.
The paper says, at C.3
Pith says: minor objection 5 of 5
The phrase 'similar to out SpatialASM-8B' contains a typo and should be 'similar to our SpatialASM-8B'.
What would settle it
Uniformly sample about 300 retained trajectories from SpatialASM-66k. For each, re-render I_edit,k = G(I, v_k) and have two human annotators plus a strong verifier model judge: (a) the planned viewpoints in r_plan are informative for q; (b) C_edit,k accurately describes the 2D layout of I_edit,k; (c) the spatial claims in r_reason follow from I and the captions alone; (d) â is entailed by (q, I, r_plan, {C_edit,k}) without the removed images. Report the tainted fraction, then train two SpatialASM-4B variants on the audited-clean versus audited-tainted subsets with matched sizes and compare their Table 1 deltas.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 12 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Short version: this is a solid, honest empirical paper and it deserves a serious referee. I believe the central claim — that the two-stage SFT-plus-GRPO recipe improves spatial reasoning by +1.2 to +3.5 average points across four Qwen backbones on seven external benchmarks while keeping general VQA roughly flat. The mechanism story ("internalized ASM") is plausible but underdetermined, and the paper partly concedes this itself.
What is actually new is the packaging. An allocentric novel-view generator that specifies a viewpoint as anchor image, pixel marker, and a single direction (no egocentric rollout); a model-based GeoP consistency metric built from feed-forward 3D reconstruction; caption-substituted plan-imagine-reason trajectories; and verifiable perspective-taking RL tasks. The evaluation is cleaner than most of this literature: benchmarks are external, no target accuracy is fitted, the 66k SFT set comes from a different source (HoliSpatial) than the main eval tasks, and the ablations show a real signal — GeoP filtering helps both stages, and SFT and RL are complementary. Related work is thorough and the citation pattern looks fine. They also state plainly that GeoP is a proxy (b99) and admit in D.3 that the trained model rarely emits the explicit ASM template, which is more candid than usual.
Soft spots, in order. First and largest: the Eq. (7) filter verifies the answer under the rendered images, but the SFT samples drop the images and keep only captions written by a different model (Qwen3.6-27B). Nothing re-checks that the question, original images, plan, and captions together entail the answer. So some retained trajectories plausibly contain fluent but wrong spatial description preceding a correct answer; on 4-choice questions a wrong-to-right flip can occur by chance in roughly a quarter of the uninformative-render cases, and the captions even describe objects the NVG was encouraged to invent (B.1.3). This does not sink the empirical claim — gains are consistent and the RL answers are genuinely verifiable — but the Sec. 4.3 sentence saying the trajectories "verify" ASM quality is stronger than the evidence supports. Second, no error bars or repeated runs anywhere; for effects of +1.2 to +3.5 points that is a moderate gap, standard for the field but worth requesting. One per-benchmark negative (Qwen3.5-9B on 3DSR) is minor. Third, only the code repo is released; the 66k and 58k data artifacts are not, so the synthesis pipeline is not yet independently checkable.
Who it is for: anyone building spatial reasoning data engines or post-training recipes for LVLMs. Send it to review. The right outcome is conditional acceptance with data release and a human or automated audit of trajectory spatial correctness as the main asks.
The strongest objection
The mechanism claim — that SpatialASM internalizes correct view planning and view imagination — depends on the quality of the trajectories in Eq. (8), and Eq. (7) does not check their spatial content. A trajectory is admitted if the teacher is wrong from I alone, correct after seeing the rendered views, and every render has GeoP below the threshold; nothing verifies r_plan, C_edit, or r_reason. Two protocol details make this load-bearing. First, the condition â = a is verified under the image-augmented input of Eq. (6), but Eq. (8) removes the images and substitutes captions written by a different model (Qwen3.6-27B); no step re-checks that (q, I, r_plan, {C_edit}) entails â, so a lossy or wrong caption can carry a conclusion that only the removed image justified. Second, on 4-choice questions a wrong-to-right flip can occur by chance in roughly a quarter of cases where the renders are uninformative, so some retained trajectories may contain post-hoc rationalized reasoning. The authors admit GeoP is a proxy (b99) and choose a lenient threshold of 2.5 for SFT (B.2.2); the NVG is even encouraged to synthesize objects absent from the inputs (B.1.3), so caption content is less constrained than the filter suggests. The paper's own D.3 observation that SpatialASM-8B rarely emits the explicit plan-imagine-reason pattern leaves the mechanism underdetermined: ten epochs of SFT on 66k answer-correct samples could improve benchmarks without any correct ASM content. The RL tasks have verifiable answers and partially mitigate this, but they supervise view-pose association, not the SFT trajectory content.
I read the paper as making an empirical claim (Table 1: average spatial gains of +1.2 to +3.5 across four backbones, general-VQA averages within 0.6) and a mechanism claim (the gains come from internalizing correct view planning and view imagination through synthesized ASM trajectories). The empirical claim is plausible: gains are consistent across backbones, benchmarks are external, and the ablations show filtering and the RL stage each contribute. The mechanism claim, however, inherits an unverified condition: the free-form spatial content of the trajectories must be correct. This is exactly the reader's weakest assumption, and I confirm it is the most load-bearing condition. If r_plan, C_edit, or r_reason are plausible but wrong, SFT teaches unreliable mentalization and the gains become attributable to reasoning-style imitation or generic CoT benefits rather than to ASM. The paper's own details sharpen the concern. Eq. (7) verifies correctness only for the image-augmented input of Eq. (6); Eq. (8) distills caption-substituted trajectories without re-verification, and the captioner differs from the teacher. The SFT threshold is admitted to be lenient (B.2.2), GeoP is admitted to be a model-based proxy (b99), and the NVG is permitted to imagine unobserved objects (B.1.3), so caption content can drift from the verified evidence. The wrong-to-right flip condition is also subject to chance correction on four-choice items. The D.3 observation that the trained model rarely emits the ASM pattern further undercuts a strong mechanism reading, and no answer-only or corrupted-reasoning SFT control is reported. Independent support exists: external benchmarks, four backbones, per-stage ablations, GeoP-filtering comparisons, and the careful exclusion of MindCube-train from SFT data (B.2.1). Nothing suggests fabrication or circularity. The open question is whether the benchmark gains measure internalized ASM or a weaker effect; the proposed audit, with a differential training follow-through, settles which claim the paper has actually established. Because the reader's CONDITIONAL verdict already centers on exactly this assumption, my stress-test does not move the verdict.
The referee's report
If the central mechanism claim holds, SpatialASM is a genuinely useful contribution: it offers a scalable route to spatial chain-of-thought supervision that does not require 3D annotations, it operates through standard post-training without architectural changes, and it is evaluated on external benchmarks with four backbones and with code released. The allocentric viewpoint interface and the GeoP-based filtering idea are also of independent interest. The paper is honest that GeoP is a proxy rather than a ground-truth metric. However, the empirical claim of consistent improvement is not yet statistically supported, and the mechanism claim that the gains come from internalizing correct view planning and view imagination is underdetermined by the current evidence. These are load-bearing issues for the paper's central narrative, though both are addressable within the manuscript's scope.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that current large vision-language models (LVLMs) lack a capability humans routinely use in spatial tasks: proactively deciding which unobserved viewpoints would supply useful evidence, then mentally reconstructing the scene from those viewpoints. SpatialASM is a post-training framework that internalizes this plan-imagine-reason cycle using a novel-view generator that renders scene observations from question-conditioned viewpoints at training time only. The rendered images are temporary scaffolding: in the final supervision they are replaced by text captions, so the model must learn to imagine the missing observations from the original images alone. Across four Qwen3-VL and Qwen3.5 backbones, average accuracy over seven spatial benchmarks rises by +1.2 to +3.5 points while general-VQA performance shifts by no more than 0.6 points. If the claim holds, the framework offers a scalable way to convert abundant answer-only spatial QA data into chain-of-thought supervision, which is currently scarce.
What carries the argument. The argument runs on three mechanisms. The first is the allocentric novel-view generator (NVG) $G$, a fine-tuned image-editing model that renders the observation from a requested viewpoint in a single step; unlike egocentric simulators that reach a target through sequences of camera actions, $G$ is conditioned directly on a viewpoint $v = (i, x, y, d)$, so the planner only states where and in which direction to look. The second is the Geometry Perturbation (GeoP) metric, which feeds a rendered view into a feed-forward 3D reconstruction model and measures the Chamfer distance between the reconstructed point clouds with and without that view, as a proxy for geometric consistency. The third is caption substitution: every rendered view in a retained trajectory is replaced by a text caption of its 2D layout, so that at inference the model reconstructs imagined observations internally from the original images alone. Together these convert answer-only spatial QA data into trajectories that supervise the full plan-imagine-reason procedure, plus RL tasks that reinforce viewpoint-to-observation association.
Core claim. The paper's central claim is that Active Spatial Mentalization (ASM) — the joint ability to plan task-relevant viewpoints and imagine the corresponding observations — is a key bottleneck in LVLM spatial reasoning, and that it can be instilled by training-time simulation rather than inference-time tool use. SpatialASM fine-tunes an image-editing model to render any scene from an allocentric viewpoint $v = (i, x, y, d)$, specified in a single query as a pixel location and viewing direction on an anchor image. A Geometry Perturbation (GeoP) metric, the Chamfer distance between 3D reconstructions with and without a rendered view, filters out geometrically inconsistent renders. The retained renders are converted into two supervision forms: explicit ASM trajectories for supervised fine-tuning, in which rendered views are replaced by textual captions so the model learns to imagine rather than consume, and automatically verifiable perspective-taking multiple-choice tasks for reinforcement learning. The paper reports average gains of +1.2 to +3.5 points across four backbones on seven spatial benchmarks, with the largest gains on tasks requiring viewpoint transformation and mental modeling, and general-VQA accuracy within 0.6 points of the baselines.
The premise it rests on. The filtering rule that admits a trajectory into training data checks only that the teacher model answers incorrectly from the original images, correctly after seeing the rendered views, and that every rendered view passes the GeoP threshold — it never verifies that the view-planning reasoning, the imagined captions, or the downstream reasoning are spatially correct, so plausible-but-wrong spatial descriptions that still reach the right answer would be reinforced as if they were valid mentalization.
If the paper is right (Pith's inferences, not the authors')
Gains of +1.2 to +3.5 average points on seven spatial benchmarks hold across four backbones from two model families, with the largest per-task gains (+8.7 on Spatial457-6d, +5.2 on MindCube) concentrated on tasks requiring viewpoint transformation and mental scene modeling.
General-VQA performance stays within 0.6 points of the baselines on all four backbones, indicating that the spatial supervision does not trade away broad visual capability.
Because rendered views never appear in the SFT data and the NVG is not called at inference, the model performs mentalization from the original images alone, so internalized spatial reasoning carries no inference-time simulation overhead.
Internalized ASM and external novel-view rendering are complementary: on MindCube subtasks, combining the tuned model with an external generator yields the highest accuracy, higher than either approach alone.
GeoP-based filtering improves both the SFT and RL stages individually, confirming that removing geometrically distorted renders is a necessary condition for the supervision to transfer.
The paper's own external-ASM experiments show that geometric consistency alone does not predict usefulness: the generator with the lowest GeoP does not yield the highest reasoning accuracy. An untested corollary is that the allocentric viewpoint format — marking a location on the anchor image — may help the LVLM align the generated view with the input scene, a property one could isolate by ablatin
A risk the paper does not fully resolve is that trajectory filtering admits chains whose intermediate captions could be spatially wrong while still reaching the correct answer, so the benchmark gains might partly be imitation of a reasoning style rather than improved spatial reconstruction; one could probe this by measuring how often the tuned model's imagined captions are spatially correct on hel
Because trajectory synthesis requires no 3D annotations, only posed image pairs, the pipeline extends naturally to other spatial QA sources; a testable prediction is that gains should scale with the diversity of scenes in the NVG training set, since GeoP filtering can only admit views the generator can actually render consistently.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. But can Large Vision-Language Models (LVLMs) proactively think in the same way? We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. To bridge this gap, we introduce SpatialASM, a scalable framework to internalize ASM into LVLMs through training-time spatial simulation. SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision: i) explicit ASM trajectories that teach LVLMs when and how to adaptively invoke view planning and imagination during spatial reasoning, and ii) automatically-constructed perspective-taking tasks that further strengthens viewpoint-conditioned scene imagination. On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities. The experimental results demonstrate that our approach improves average accuracy across the seven spatial benchmarks. And the synthesis of ASM trajectories provides a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision for eliciting active viewpoint exploration and mental scene simulation.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-12
The authors' main claim, as the classifier picked it out 0.69
“On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.15 · human participants 0.08 · controlled comparison 0.43 · randomized assignment 0.02 · states quantitative result 0.60 · compares with prior results 0.52 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints. | background 1.00 | 0.08 | 0.03 | asserted 1.00 |
| But can Large Vision-Language Models (LVLMs) proactively think in the same way? | gap 0.90 | 0.04 | 0.02 | unclear 0.55 |
| We term this ability Active Spatial Mentalization (ASM), an intrinsic component of visual intelligence that remains substantially underdeveloped in current LVLMs. | gap 0.93 | 0.19 | 0.08 | asserted 1.00 |
| To bridge this gap, we introduce SpatialASM, a scalable framework to internalize ASM into LVLMs through training-time spatial simulation. | contribution 0.95 | 0.86 | 0.09 | asserted 0.99 |
| SpatialASM employs an allocentric novel-view generator to ground hypothetical viewpoints in visual evidence and produces two complementary forms of supervision: i) explicit ASM trajectories that teach LVLMs when and how to adaptively invoke view planning and imagination during spatial reasoning, and ii) automatically-constructed perspective-taking tasks that further strengthens viewpoint-conditioned scene imagination. | method 0.99 | 0.88 | 0.06 | asserted 1.00 |
| On seven benchmarks spanning relatively perception-oriented spatial understanding to more challenging tasks requiring multi-image, cross-view and viewpoint-transformation reasoning, SpatialASM yields performance gain on LVLMs of four LVLM backbones in spatial reasoning while retaining their original general vision capabilities. | result 1.00 | 0.96 | 0.62 | asserted 1.00 |
| The experimental results demonstrate that our approach improves average accuracy across the seven spatial benchmarks. | result 1.00 | 0.96 | 0.37 | asserted 1.00 |
| And the synthesis of ASM trajectories provides a scalable pathway to alleviate the scarcity of spatial Chain-of-Thought (CoT) supervision for eliciting active viewpoint exploration and mental scene simulation. | interpretation 0.95 | 0.78 | 0.19 | asserted 0.67 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-12: 1 request, 6,353 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
References (85)
- [1] Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Adeline Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yogesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800, 2026. 2 on PithPith's reviewarXiv
- [2] Lei Bai, Jiaqi Cao, Chiyu Chen, Guanzhou Chen, Kai Chen, Guangran Cheng, Erfei Cui, Xuanlang Dai, Shengyuan Ding, Shangheng Du, et al. Intern-s2-preview: Scientific agentic foundation model.arXiv preprint arXiv:2608.13505, 2026. 1 on PitharXiv
- [3] Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025. 1, 4, C.3, C.5 on PithPith's reviewarXiv
- [4] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data.arXiv preprint arXiv:2111.08897, 2021. 2 on PithPith's reviewarXiv
- [5] Mahtab Bigverdi, Linjie Li, Weikai Huang, Yiming Liu, Jaemin Cho, Tuhin Kundu, Chris Dongjoo Kim, Zelun Luo, Jieyu Zhang, Linda Shapiro, et al. Imaginative perception tokens enhance spatial reasoning in multimodal language models. InEuropean Conference on Computer Vision, page... on Pith
- [6] Black Forest Labs. FLUX.2: Frontier visual intelligence.https://bfl.ai/blog/flux-2, 2025. 3.1, C.1, D.1 on Pith
- [7] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. InForty-first international conference on machine learning, 2024. 2 on Pith
- [8] Wenxiao Cai, Iaroslav Ponomarenko, Jianhao Yuan, Xiaoqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 9490–9498. IEEE, 2025. A.2 on Pith
- [9] Zhongang Cai, Ruisi Wang, Chenyang Gu, Fanyi Pu, Junxiang Xu, Yubo Wang, Wanqi Yin, Zhitao Yang, Chen Wei, Tongxi Zhou, et al. Scaling spatial intelligence with multimodal foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ... on Pith
- [10] Meng Cao, Xingyu Li, Xue Liu, Ian Reid, and Xiaodan Liang. Spatialdreamer: Incentivizing spatial reasoning via active mental imagery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7176–7187, 2026. 1, 2 on Pith
- [11] 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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14455–144... on Pith
- [12] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models?Advances in Neural Information Processing Systems, 37:27056–27087, 2024. 4, C.2 on Pith
- [13] Xinyuan Chen, Haoyu Guo, Shi Guo, Bingqi Jiang, Chunhua Shen, Xing Shen, Tianfan Xue, Yufei Xue, Mulin Yu, Weinan Zhang, et al. A definition and roadmap for world models.arXiv preprint arXiv:2607.06401, 2026. 2 on PithPith's reviewarXiv
- [14] Zhixun Chen, Ming Li, Yuxuan Huang, Yali Du, Meng Fang, and Tianyi Zhou. Atlas: Agent tuning via learning critical steps. InFindings of the Association for Computational Linguistics: ACL 2025, pages 25334–25349, 2025. A.3 on Pith
- [15] 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.Advances in Neural Information Processing Systems, 37:135062–135093, 2024. 1, 2, A.2 12 SpatialASM: Interna... on Pith
- [16] Seokju Cho, Ryo Hachiuma, Abhishek Badki, Hang Su, Byung-Kwan Lee, Chan Hee Song, Sifei Liu, Subhashree Radhakrishnan, Seungryong Kim, Yu-Chiang Frank Wang, et al. Spatialclaw: Rethinking action interface for agentic spatial reasoning.arXiv preprint arXiv:2606.13673, 2026. A.3 on PithPith's reviewarXiv
- [17] Amirhossein Dabiriaghdam, Shayan Vassef, Mohammadreza Bakhtiari, Yasamin Medghalchi, Ilker Haci- haliloglu, Mesrob Ohannessian, Lele Wang, and Giuseppe Carenini. Vamps: Visual-assisted mathematical problem solving benchmark.arXiv preprint arXiv:2606.04244, 2026. A.1 on PithPith's reviewarXiv
- [18] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In2017 IEEE conference on computer vision and pattern recognition (CVPR), pages 2432–2443. IEEE, 2017. 2 on Pith
- [19] Yalun Dai, Hao Li, Shulin Tian, Runmao Yao, Yuhao Dong, Fangzhou Hong, Zhaoxi Chen, Fangfu Liu, Baoliang Tian, Dingwen Zhang, et al. S-agent: Spatial tool-use elicits reasoning for spatial intelligence. arXiv preprint arXiv:2606.20515, 2026. A.3 on PithPith's reviewarXiv
- [20] NianchenDeng, LixinGu, ShenglongYe, YinanHe, ZheChen, SongzeLi, HaominWang, JinhuiYin, QiWei, Tianshuo Yang, et al. Internspatial: A comprehensive dataset for spatial reasoning in vision-language models. InInternational Conference on Learning Representations, volume 2026, page... on Pith
show 65 more
- [21] Shuangrui Ding, Xuanlang Dai, Long Xing, Shengyuan Ding, Ziyu Liu, Yang JingYi, Penghui Yang, Zhixiong Zhang, Xilin Wei, Xinyu Fang, et al. Wildclawbench: A benchmark for real-world, long-horizon agent evaluation.arXiv preprint arXiv:2605.10912, 2026. A.1 on PithPith's reviewarXiv
- [22] Mengfei Du, Binhao Wu, Zejun Li, Xuanjing Huang, and Zhongyu Wei. EmbSpatial-Bench: Benchmarking spatial understanding for embodied tasks with large vision-language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Sho... on Pith
- [23] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. VLMEvalKit: An open-source toolkit for evaluating large multi-modality models. InProceedings of the 32nd ACM international conference on multimedia,... on Pith
- [24] Xiang Feng, Jiawei Zhou, Zhangfeng Huang, Kewei Wang, Shanshan Ye, Jinxin Hu, Zulong Chen, Yong Luo, and Jing Zhang. Docscope: Benchmarking verifiable reasoning for trustworthy long-document understanding.arXiv preprint arXiv:2605.08888, 2026. A.1 on PithPith's reviewarXiv
- [25] 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, pages 148–166. Springer, 2024. 1 on Pith
- [26] Yuanyuan Gao, Hao Li, Yifei Liu, Xinhao Ji, Yuning Gong, Yuanjun Liao, Fangfu Liu, Manyuan Zhang, Yuchen Yang, Dan Xu, et al. Holi-spatial: Evolving video streams into holistic 3d spatial intelligence. arXiv preprint arXiv:2603.07660, 2026. 1, 2, 4, A.2, B.2.1, B.3, C.3 on PitharXiv
- [27] Renjie Gu, Kaichen Zhou, Yan Luo, and Mengyu Wang. GeoWorld-VLM: Geometry from world models for vision-language models.arXiv preprint arXiv:2605.16713, 2026. 2 on PithPith's reviewarXiv
- [28] Zefeng He, Xiaoye Qu, Yafu Li, Tong Zhu, Siyuan Huang, and Yu Cheng. Diffthinker: Towards generative multimodal reasoning with diffusion models.arXiv preprint arXiv:2512.24165, 2025. C.1 on PitharXiv
- [29] Yining Hong, Jiageng Liu, Han Yin, Manling Li, Leonidas Guibas, Li Fei-Fei, Jiajun Wu, and Yejin Choi. Esi-bench: Towards embodied spatial intelligence that closes the perception-action loop.arXiv preprint arXiv:2605.18746, 2026. 2, 4, A.1, C.2 on PithPith's reviewarXiv
- [30] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metricdepthandsurfacenormalestimation.IEEETransactionsonPatternAnalysisandMachine... on Pith
- [31] Amita Kamath, Jack Hessel, and Kai-Wei Chang. What’s “up” with vision-language models? investigating their struggle with spatial reasoning. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9161–9175, 2023. 2 on Pith
- [32] Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, et al. Mapanything: Universal feed-forward metric 3d reconstruction; map-anything. github. io. In2026 International Confe... on Pith
- [33] Juil Koo, Daehyeon Choi, Sangwoo Youn, Phillip Y Lee, and Minhyuk Sung. Toward ambulatory vision: Learning visually-grounded active view selection.arXiv preprint arXiv:2512.13250, 2025. 2 on PitharXiv
- [34] Phillip Y Lee, Jihyeon Je, Chanho Park, Mikaela Angelina Uy, Leonidas Guibas, and Minhyuk Sung. Perspective-aware reasoning in vision-language models via mental imagery simulation. InProceedings of the IEEE/CVF international conference on computer vision, pages 9241–9251, 2025. 2 on Pith
- [35] Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, et al. ViewSpatial-Bench: Evaluating multi-perspective spatial localization in vision-language models. InEuropean Conference on Computer Vision, ... on Pith
- [36] Haoyuan Li, Qihang Cao, Tao Tang, Kun Xiang, Zihan Guo, Jianhua Han, JiaWang Bian, Hang Xu, and Xiaodan Liang. Thinking with geometry: Active geometry integration for spatial reasoning.arXiv preprint arXiv:2602.06037, 2026. 2, A.1 on PithPith's reviewarXiv
- [37] Hongxing Li, Dingming Li, Zixuan Wang, Yuchen Yan, Hang Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu, Jun Xiao, and Yueting Zhuang. Spatialladder: Progressive training for spatial reasoning in vision- language models. InInternational Conference on Learning Representations, volu... on Pith
- [38] Xudong Li, Mengdan Zhang, Peixian Chen, Jiaxi Tan, Zihao Huang, Jingyuan Zheng, Yan Zhang, Xiawu Zheng, Xing Sun, and Rongrong Ji. OmniView-Space: Reinforcing spatial reasoning via multi-perspective spatial mapping.arXiv preprint arXiv:2607.00881, 2026. 1 on PithPith's reviewarXiv
- [39] Haotong Lin, Sili Chen, Junhao Liew, Donny Y Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025. A.3 on PithPith's reviewarXiv
- [40] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2216... on Pith
- [41] Jianhui Liu, Haoze Sun, Wenbo Li, Yanbing Zhang, Rui Yang, Zhiliang Zhu, Yijun Yang, Shenghe Zheng, Nan Jiang, Jiaxiu Jiang, et al. Openspatial: A principled data engine for empowering spatial intelligence. arXiv preprint arXiv:2604.07296, 2026. 1, A.2 on PithPith's reviewarXiv
- [42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean conference on computer vision, pages 38–55. Springer,... on Pith
- [43] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? InEuropean conference on computer vision, pages 216–233. Springer, 2024. 1 on Pith
- [44] Yuhong Liu, Beichen Zhang, Yuhang Zang, Yuhang Cao, Long Xing, Xiaoyi Dong, Haodong Duan, Dahua Lin, and Jiaqi Wang. Spatial-ssrl: Enhancing spatial understanding via self-supervised reinforcement learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Patter... on Pith
- [45] Wufei Ma, Haoyu Chen, Guofeng Zhang, Yu-Cheng Chou, Jieneng Chen, Celso De Melo, and Alan Yuille. 3DSRBench: A comprehensive 3d spatial reasoning benchmark. In2025 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6924–6934. IEEE, 2025. 4, C.2 on Pith
- [46] OpenAI. GPT Image 2 model, 2026. C.5 on Pith
- [47] Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, and Xu Sun. Spacer: Reinforcing mllms in video spatial reasoning.arXiv preprint arXiv:2504.01805, 2025. 2, A.2, C.3 on PithPith's reviewarXiv
- [48] Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. 1, 4 on Pith
- [49] Qwen Team. Qwen3.6-27B: Flagship-level coding in a 27b dense model, April 2026. B.2.2 on Pith
- [50] Qwen Team. Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026. B.2.2, C.5 on Pith
- [51] Arijit Ray, Jiafei Duan, Ellis Brown, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, et al. Sat: Dynamic spatial aptitude training for multimodal language models.arXiv preprint arXiv:2412.07755, 2024. 2 on PitharXiv
- [52] Juneyoung Ro, Namwoo Kim, and Yoonjin Yoon. How well do vision-language models understand cities? a comparative study on spatial reasoning from street-view images. In2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 6535–6544. IEEE, 2025. A.2 on Pith
- [53] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 1, 3.3 on PithPith's reviewarXiv
- [54] Roger N Shepard and Jacqueline Metzler. Mental rotation of three-dimensional objects.Science, 1971. 1 on Pith
- [55] HunyuanWorld Team, Zhenwei Wang, Yuhao Liu, Junta Wu, Zixiao Gu, Haoyuan Wang, Xuhui Zuo, Tianyu Huang, Wenhuan Li, Sheng Zhang, et al. Hunyuanworld 1.0: Generating immersive, explorable, and interactive 3d worlds from words or pixels.arXiv preprint arXiv:2507.21809, 2025. 2 on PithPith's reviewarXiv
- [56] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, Jianfeng Cai, Xinyuan Cai, Peizhou Cao, Yuxuan Cao, Ziwei Chai, Y Charles, et al. Kimi k3: Open frontier intelligence.arXiv preprint arXiv:2607.24653, 2026. A.1 on PitharXiv
- [57] Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Ziwei Chai, Y Charles, HS Che, Cheng Chen, et al. Kimi k2. 5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276, 2026. C.5 on PithPith's reviewarXiv
- [58] Qwen Team. Qwen2.5-vl, January 2025. C.3 on Pith
- [59] Kangrui Wang, Linjie Li, Zhengyuan Yang, Shiqi Chen, Zihan Wang, Li Fei-Fei, Jiajun Wu, Leonidas Guibas, Lijuan Wang, and Manling Li. Planning with the views.arXiv preprint arXiv:2605.29563, 2026. 2 on PithPith's reviewarXiv
- [60] Qineng Wang, Baiqiao Yin, Pingyue Zhang, Jianshu Zhang, Kangrui Wang, Zihan Wang, Jieyu Zhang, Keshigeyan Chandrasegaran, Han Liu, Ranjay Krishna, et al. Spatial mental modeling from limited views. InThe Fourteenth International Conference on Learning Representations, 2025. 4,... on Pith
- [61] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20697–20709. IEEE, 2024. A.2 on Pith
- [62] Wenqi Wang, Reuben Tan, Pengyue Zhu, Jianwei Yang, Zhengyuan Yang, Lijuan Wang, Andrey Kolobov, Jianfeng Gao, and Boqing Gong. Site: towards spatial intelligence thorough evaluation. In2025 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9058–9069. IEEE, 2025. 2 on Pith
- [63] Xingrui Wang, Wufei Ma, Tiezheng Zhang, Celso M de Melo, Jieneng Chen, and Alan Yuille. Spatial457: A diagnostic benchmark for 6d spatial reasoning of large multimodal models. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24669–24679. IEEE... on Pith
- [64] Shenxi Wu, Yuhong Liu, Haosong Zhang, Tongjin Zou, Yanxun Zhang, Gaochang Chen, Dun Liang, Jiaqi Wang, Zhecan James Wang, Yuhang Zang, et al. Scidocbench: A workflow-centered benchmark and data pipeline for scientific document understanding.arXiv preprint arXiv:2609.05141, 2026. A.1 on PithPith's reviewarXiv
- [65] xAI. Grok-1.5 vision preview.https://x.ai/news/grok-1.5v, 2024. 4, C.2 on Pith
- [66] Yicheng Xiao, Wenhu Zhang, Lin Song, Yukang Chen, Wenbo Li, Nan Jiang, Tianhe Ren, Haokun Lin, Wei Huang, Haoyang Huang, et al. Spatialedit: Benchmarking fine-grained image spatial editing.arXiv preprint arXiv:2604.04911, 2026. C.5 on PithPith's reviewarXiv
- [67] Long Xing, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Jianze Liang, Qidong Huang, Jiaqi Wang, Feng Wu, and Dahua Lin. Caprl: Stimulating dense image caption capabilities via reinforcement learning. In International Conference on Learning Representations, volume 2026, pages 13066–13... on Pith
- [68] Haotian Xu, Yue Hu, Zhengqiu Zhu, Chen Gao, Ziyou Wang, Junreng Rao, Wenhao Lu, Weishi Li, Quanjun Yin, and Yong Li. Citycube: Benchmarking cross-view spatial reasoning on vision-language models in urban environments. InProceedings of the 64th Annual Meeting of the Association... on Pith
- [69] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10632–10643. IEEE, 2025. 1 on Pith
- [70] Penghui Yang, Long Xing, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Yibin Wang, Yujie Zhou, Jiazi Bu, Jianze Liang, Qidong Huang, et al. Caprl++: Unified reinforcement learning with verifiable rewards for dense image and video captioning.arXiv preprint arXiv:2606.09393, 2026. 1, A.1 on PithPith's reviewarXiv
- [71] Rui Yang, Ziyu Zhu, Yanwei Li, Jingjia Huang, Shen Yan, Siyuan Zhou, Zhe Liu, Xiangtai Li, Shuangye Li, Wenqian Wang, et al. Visual spatial tuning. InEuropean Conference on Computer Vision, pages 192–211. Springer, 2026. 1, 2, 4, A.2, B.2.1, C.3 on Pith
- [72] Sihan Yang, Runsen Xu, Yiman Xie, Sizhe Yang, Mo Li, Jingli Lin, Chenming Zhu, Xiaochen Chen, Haodong Duan, Xiangyu Yue, et al. Mmsi-bench: A benchmark for multi-image spatial intelligence. In International Conference on Learning Representations, volume 2026, pages 157051–1570... on Pith
- [73] Yuncong Yang, Jiageng Liu, Zheyuan Zhang, Siyuan Zhou, Reuben Tan, Jianwei Yang, Yilun Du, and Chuang Gan. Mindjourney: Test-time scaling with world models for spatial reasoning.Advances in Neural Information Processing Systems, 38:109855–109885, 2026. 1, 2, A.3 on Pith
- [74] Shoubin Yu, Yue Zhang, Zun Wang, Jaehong Yoon, Huaxiu Yao, Mingyu Ding, and Mohit Bansal. When and how much to imagine: Adaptive test-time scaling with world models for visual spatial reasoning. arXiv preprint arXiv:2602.08236, 2026. 1, 2, A.3 on PithPith's reviewarXiv
- [75] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF conference on co... on Pith
- [76] Beichen Zhang, Yuhong Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Haodong Duan, Yuhang Cao, Dahua Lin, and Jiaqi Wang. Booststep: Boosting mathematical capability of large language models via improved single-step reasoning.arXiv preprint arXiv:2501.03226, 2025. A.1 on PithPith's reviewarXiv
- [77] Jiahui Zhang, Yurui Chen, Yueming Xu, Ze Huang, Jilin Mei, Chunhui Chen, Yanpeng Zhou, Yu-Jie Yuan, Xinyue Cai, Guowei Huang, et al. From flatland to space: Teaching vision-language models to perceive and reason in 3d.Advances in Neural Information Processing Systems, 38, 2026. 2 on Pith
- [78] Wanyue Zhang, Wenxiang Wu, Wang Xu, Jiaxin Luo, Helu Zhi, Yibin Huang, Shuo Ren, Zitao Liu, and Jiajun Zhang. World2vlm: Distilling world model imagination into vlms for dynamic spatial reasoning. arXiv preprint arXiv:2604.26934, 2026. 1, 2, A.2 16 SpatialASM: Internalizing Ac... on PithPith's reviewarXiv
- [79] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1... on Pith
- [80] 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 arXiv:2601.13029, 2026. 1, 2, A.3 on PithPith's reviewarXiv
- [81] Duo Zheng, Yanyang Li, Liwei Wang, et al. Learning from videos for 3d world: Enhancing mllms with 3d vision geometry priors.Advances in neural information processing systems, 38:20560–20586, 2026. 2, A.1 on Pith
- [82] Jensen Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. In2025 IEEE/CVF International Conference on Computer Vision (ICCV), pa... on Pith
- [83] Yang Zhou, Zixuan Huang, Sunzhu Li, Zhuo Yang, Chen Zhang, Shunian Chen, Caijun Yan, Jianyao Xu, Shunyu Liu, Weijie Fu, et al. Spatialcli: Learning to reason with spatial tools, then without them.arXiv preprint arXiv:2607.27703, 2026. 1, A.3 on PithPith's reviewarXiv
- [84] Chenming Zhu, Jingli Lin, Yilin Long, Peizhou Cao, Tai Wang, Jiangmiao Pang, and Xihui Liu. Thinking with imagination: Agentic visual spatial reasoning with world simulators.arXiv preprint arXiv:2606.06476, on PithPith's reviewarXiv
- [2026] 1, 2, A.3 17 SpatialASM: Internalizing Active Mentalization for Spatial Reasoning A. More Related Work A.1. Proactive Spatial Intelligence LVLMs have exhibited outstanding performance on various tasks, such as image and video captioning [67, 70], math reasoning [17, 76], agent... on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment