Pith. sign in

REVIEW 4 major objections 5 minor 33 references

GeoAnchor claims that breaking 3D spatial reasoning into position, direction, and scene-geometry latents lets a 2-billion-parameter model beat much larger vision-language systems on 3D reasoning benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:07 UTC pith:GKQJ6OFX

load-bearing objection Real method, decent ablations, but the SOTA claim rests on in-domain benchmarks and an uncontaminated eval is needed. the 4 major comments →

arxiv 2607.13454 v2 pith:GKQJ6OFX submitted 2026-07-15 cs.CV cs.AI

GeoAnchor: Collaborative Reasoning via Latent Decomposition for 3D Spatial Understanding

Hao Li , Han Fang , Zixin Pan , Xin Wei , Hongbo Sun , Jinglin Xu , Zhiyu Lin , Ye Yuan
show 3 more authors
Zhongjiang He Yu Yu Hao Sun
This is my paper
classification cs.CV cs.AI
keywords 3D spatial reasoningmultimodal large language modelslatent reasoningposition tokensdirection tokensgeometry tokensinterpretable reasoningvision-language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that 3D spatial understanding from a single 2D image fails when models are forced to verbalize geometry into discrete text, and also fails when a single continuous latent must carry all spatial information. GeoAnchor splits spatial evidence into three recombinable continuous latents—position for object grounding, direction for relational orientation, and geometry for scene structure—interleaved with ordinary text so the model can reason in both modalities. Trained in four stages that move from local perception to global context and then to adaptive latent selection via reinforcement learning, the resulting 2-billion-parameter model reports state-of-the-art accuracy on two in-domain 3D reasoning benchmarks and a third out-of-domain benchmark, surpassing its base model by about 21 percentage points on average. If correct, this means compact open models can outperform far larger proprietary systems on geometry-sensitive tasks like robotics, driving, and AR, simply by giving the model the right structure for continuous spatial evidence.

Core claim

On its own terms, the paper's central claim is that 3D spatial reasoning should be conducted as an interleaved text-latent trajectory rather than as pure text or a single latent. GeoAnchor introduces three typed latent tokens: a position token that is linearly decoded to 3D coordinates and trained with Smooth L1 loss, a direction token decoded to a 3D unit vector and trained with cosine loss, and a geometry token aligned to coarse pooled features from a pretrained geometry model through a soft-coverage loss that encourages every scene feature to be represented by at least one token without forcing dense token-wise correspondence. The text tokens handle semantic planning; the latents carry co

What carries the argument

The load-bearing machinery is the decomposed latent sequence: an interleaved trajectory O = t1 ⊕ z1 ⊕ ... ⊕ zk−1 ⊕ tk in which each latent token is a fixed-length block of continuous hidden states projected back into the model's embedding space by a projector. Position and direction tokens are averaged over their hidden states and mapped by linear heads to 3D points and directions, giving interpretable local evidence; the geometry token is projected into several geometry tokens and aligned to a coarse multi-scale pooled feature map using a soft-coverage loss with a balance term that prevents collapse. This structure lets the model compose local and global evidence dynamically, and the final

Load-bearing premise

The load-bearing premise is that the pseudo-depth-derived 3D positions and directions used to supervise the local tokens are faithful enough to true geometry—if monocular depth is systematically biased in the scenes that matter, the latents learn the annotator's errors, not space.

What would settle it

Train the identical pipeline using ground-truth depth (e.g., LiDAR or synthetic renderers) for the local 3D supervision while keeping everything else fixed, and compare on the same benchmarks; if accuracy does not improve or changes little, the reported gains may be tied to the pseudo-depth statistics rather than to true 3D understanding. Separately, evaluate the trained model on a held-out benchmark constructed from non-ScanNet scene distributions to test for contamination.

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

If this is right

  • Text-only chain-of-thought does not consistently improve spatial reasoning; continuous latents are the effective carrier of geometric evidence.
  • Decomposed tokens outperform a single latent, so the gains come from separating object grounding, orientation, and scene structure rather than from latent reasoning per se.
  • The model generalizes to an out-of-domain viewpoint benchmark (10.7-point gain over the base model), suggesting the mechanism transfers beyond the training scene distribution.
  • The approach transfers to a different base model, indicating the latent-decomposition recipe is not tied to one backbone.

Where Pith is reading between the lines

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

  • If the local 3D labels inherit systematic bias from monocular depth estimation, the reported gains could partly reflect learning the pseudo-label pipeline; a cleaner test would retrain with ground-truth depth or LiDAR on a subset and compare.
  • The pattern-reward mechanism suggests an efficiency principle: models should invoke global scene context only when local evidence is insufficient—this could generalize to other reasoning domains beyond 3D.
  • The interpretability of position/direction latents (each attends to its own object) could be repurposed as a debugging tool for spatial failures in embodied agents.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes GeoAnchor, a 2B multimodal LLM for 3D spatial reasoning from a single image. It introduces an interleaved text-latent reasoning framework with three decomposed latent tokens — position, direction, and geometry — and a four-stage collaborative training strategy: local perception warm-up, spatial latent reasoning with explicit local/global losses, latent relaxation via text-only supervision, and GRPO with pattern-specific rewards. Evaluated on SPAR-Bench, SPBench, and ViewSpatial, the paper reports 68.4%, 69.7%, and 47.0% mean accuracy respectively, and claims state-of-the-art performance and robust generalization, including a 21.2% average improvement over the base Qwen3-VL-2B model.

Significance. If the empirical claims are supported, GeoAnchor contributes a plausible and well-motivated design: decomposing latent reasoning into interpretable local and global factors for 3D spatial understanding, with a carefully staged training curriculum. The paper provides extensive ablations (training stages, token lengths, pooling resolutions, different base models) and an explicit validation of pseudo-depth supervision in Appendix B.2.2. These are strengths. However, the headline claims of state-of-the-art and robust generalization rest on benchmarks that are not independent of the training data, and the central quantitative claims lack error bars. The technical idea is worth publishing if the evaluation is re-framed or supplemented with scene-disjoint and independent tests.

major comments (4)
  1. [§4.1 and App. A.1–A.3] The headline claims of 'state-of-the-art' and 'robust generalization' are not supported because the evaluation benchmarks are drawn from the same data sources as the training data. The model trains on 100k SPAR questions and 5k SpatialLadder-26K samples (§3.4, App. B.3), while SPAR-Bench is explicitly 'constructed from the SPAR-7M split' (App. A.1) and SPBench-SI is built with the same pipeline as SpatialLadder on the ScanNet validation set (App. A.2). ViewSpatial images use ScanNet validation and MS-COCO (App. A.3), and Stage 1 uses 10k ScanNet scenes (App. B.2.1) with no stated scene-disjointness check. The paper itself labels SPAR-Bench and SPBench as 'in-domain' (§4.1). Please report results on scene-disjoint splits (e.g., train on SPAR scenes, test on held-out scenes) and on an independent out-of-domain benchmark not derived from ScanNet/COCO. Without this, the SOTA claim is not est
  2. [§1, Table 2] The '21.2% margin' over the base model conflates generic SFT gains with the latent-decomposition contribution. Table 2 shows vanilla SFT already improves the base Qwen3-VL-2B by 24.0/5.2/4.6 points on SPAR/SPBench/ViewSpatial, and text CoT SFT by 27.7/0.0/5.9. The marginal gain of the full GeoAnchor over text CoT is 7.4/15.9/4.1 points, not 21.2%. The 21.2% figure is accurate only as a base-model improvement and is misleading as evidence for the method. Please report and emphasize the comparison against equal-compute SFT and CoT baselines, and temper the 'state-of-the-art' claim accordingly.
  3. [§4.2, Tables 1–5] All results are reported from a single run with no error bars, seeds, or significance tests. Given that several headline differences are small (e.g., ViewSpatial 47.0 vs. Gemini 45.5 in Table 1; SPBench Rel. 86.7 vs. SpatialLadder 81.6) and that the method involves stochastic RL (Stage 4), the empirical claims need at least 3 seeds with standard deviations and a paired significance test (e.g., bootstrap). Without this, it is impossible to assess whether the improvements over the strongest baselines are reliable.
  4. [App. B.2.2, Table 6] The pseudo-depth validation is reported only as aggregate mean/median error (0.09 m / 0.04 m) without specifying the validation subset, its size, or whether those scenes overlap with the evaluation benchmarks. Equation (1) averages depth over the entire bounding box, which can be biased for large or partially occluded objects. Since the position and direction latents are trained against these pseudo-depth-derived coordinates, this is a load-bearing component. Please describe the validation split, stratify errors by object size and depth range, and show that the downstream reasoning gains are not an artifact of the pseudo-depth pipeline (e.g., by training a small subset with ground-truth depth and comparing).
minor comments (5)
  1. [§3.4 vs App. B.3, Table 9] The text says 5k samples from SpatialLadder-26K are used, but Table 9 sums to 5,928 samples. Please reconcile the numbers.
  2. [Figure 2] The caption says 'Each global token aligns only with its most similar feature,' but Eq. (7)–(8) use a soft-coverage average assignment, not hard max similarity. Please align the caption with the actual objective.
  3. [Table 1] SpatialLadder's SpBench Abs. value is bolded but SpatialLadder is not the overall best; the bold/underline criteria should be clarified (e.g., bold for best overall per column, underline for second best overall).
  4. [§4.1] The protocol for numerical questions ('mean accuracy computed as the average accuracy across confidence thresholds from 0.5 to 0.9 with a step size of 0.05') is under-specified. Please clarify whether each generated answer is scored at each threshold or only the best, and how multiple valid outputs are handled.
  5. [App. D.1] Typo: 'illustrateed' should be 'illustrated.'

Circularity Check

2 steps flagged

The SOTA numbers on SPAR-Bench and SPBench are partly forced because those benchmarks come from the same data/pipeline as the training set; the decomposition method itself retains independent content, so the circularity is partial.

specific steps
  1. fitted input called prediction [Section 3.4 (Dataset Construction) and Appendix A.1]
    "For the spatial reasoning dataset, we use SPAR ... We sample 100k questions from SPAR ... In total, the resulting spatial reasoning dataset contains 105k samples. ... SPAR-Bench is constructed from the SPAR-7M split by selecting representative spatial tasks and manually verifying the resulting samples for quality control."

    The headline SPAR-Bench number (68.4%) is obtained after training on 100k SPAR questions sampled from SPAR-7M, while the benchmark is 'constructed from the SPAR-7M split' and uses the same task types (depth, distance, proximity, relation, view change; cf. App. A.1 Table 1 and App. B.3 Table 8). The paper never states that SPAR-Bench samples or scenes were excluded from the 100k training sample. Thus the benchmark score is not an independent test of 3D reasoning; it is an in-distribution fit produced by the same data pipeline and partly forced by the training input — the 'fitted input called prediction' pattern.

  2. fitted input called prediction [Section 3.4 (Dataset Construction) and Appendix A.2]
    "The paper further introduces two dedicated benchmarks built using the same pipeline on the ScanNet validation set, namely SPBench-SI and SPBench-MV. ... we further incorporate 5k samples from SpatialLadder-26k."

    SPBench-SI is created by the same pipeline that generated SpatialLadder-26K, and the training set includes 5k SpatialLadder-26K samples. The SPBench task types (relative direction/distance, absolute distance, object size) are exactly the SpatialLadder task categories. No disjointness between the 5k training samples and the SPBench pipeline is reported. The 69.7% SPBench result therefore measures learning of the training task/pipeline distribution rather than an external prediction — a second instance of fitted-input-called-prediction.

full rationale

The paper's own appendices establish that the two headline in-domain benchmarks are generated from the same data/pipeline as the training set: SPAR-Bench is 'constructed from the SPAR-7M split' after the model trains on 100k SPAR questions, and SPBench-SI is built from the SpatialLadder pipeline on ScanNet validation after training on 5k SpatialLadder-26K samples. No explicit exclusion of benchmark samples/scenes from training is reported, so the 68.4/69.7 SOTA numbers are partly measures of in-distribution fit rather than independent predictions. The ViewSpatial result is the main claimed out-of-domain support, but it also draws ScanNet validation images while training is ScanNet-heavy, and no scene-disjointness check is reported; thus the robust-generalization claim is weakly grounded. The core methodological derivation — decomposed position/direction/geometry latents with multi-stage training — is not definitionally tied to the benchmark numbers; ablations, the comparison to text CoT, and the Qwen2.5-VL instantiation supply independent content. This is partial circularity from benchmark/training overlap, not a definitional identity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

The central empirical claim rests on pseudo-label pipelines, VGGT-based supervision, and benchmark alignment. These are domain assumptions rather than formal axioms, and the method depends on several hand-tuned hyperparameters. The invented entities are not physical but are new learned representation components; they do not have falsifiable handles outside the paper's own benchmarks.

free parameters (5)
  • latent token lengths (lpos, ldir, lgeo) = 2, 2, 8
    Tuned by ablation in Table 13; no theoretical reason is given for why these lengths are optimal.
  • VGGT alignment temperature τ = not reported
    Appears in Eq. (7) but no value is given in Table 10 or elsewhere, so replication requires guessing.
  • multi-scale pooling resolutions = {1, 2, 4}
    Tuned by ablation in Table 14; {1,2,4,8} performs worse, so the choice is empirical.
  • loss weights (λt, λl, λg, λbal) = 1, 1, 0.1, 0.05
    Set by hand; no sensitivity analysis is reported for λt, λl, or λbal.
  • RL pattern-reward hyperparameters (rpattern, β, κ, μ) = 0.5, 0.01, 8, 0.2
    Chosen without a sweep; only the presence versus absence of the pattern reward is ablated.
axioms (4)
  • domain assumption Depth Anything v3 pseudo-depth, after back-projection, is accurate enough to serve as ground truth for local 3D supervision.
    Appendix B.2.2 and Table 6 justify this with mean error 0.09 m on a subset, but if depth estimates are biased in other scenes, the position/direction losses teach wrong geometry.
  • domain assumption VGGT final-layer features are a valid proxy for global 3D scene structure.
    Section 3.2 supervises the geometry token with VGGT features; no independent evidence is given that these features capture geometry beyond VGGT's own training.
  • domain assumption Qwen3-VL-32B-generated object boxes and referring expressions, after IoU≥0.5 self-check, are correct.
    Appendix B.2.1 uses a pseudo-label pipeline with 94.41% retention but no human validation beyond self-consistency.
  • domain assumption Evaluation benchmarks are not contaminated by the training distribution.
    Training uses SPAR and SpatialLadder-26K while SPAR-Bench and SPBench are built from the same sources/pipelines; overlap is never quantified.
invented entities (3)
  • position latent token zpos no independent evidence
    purpose: Encodes the estimated 3D position of a target object as a sequence of continuous hidden states, decoded by Wpos to (x, y, z).
    This is a learned network construct, not an independently observable quantity; evidence is internal ablations and attention maps, with no external falsifiable handle.
  • direction latent token zdir no independent evidence
    purpose: Encodes relative orientation between objects, decoded by Wdir to a 3D direction vector.
    Same as position token: an internal latent representation whose meaning is inferred from supervised losses and ablations, not from an external measurement.
  • geometry latent token zgeo no independent evidence
    purpose: Encodes global scene structure by soft-aligning to VGGT feature maps.
    Although it is supervised by VGGT features, its status as a geometry representation is only shown through t-SNE proximity and benchmark gains, both internal to the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 21747 in / 11565 out tokens · 109398 ms · 2026-08-02T05:07:21.027453+00:00 · methodology

0 comments
read the original abstract

Although multimodal large language models (MLLMs) have achieved remarkable progress, understanding 3D spatial relationships from 2D images remains a critical challenge. Existing methods primarily rely on symbolic text tokens, which inherently lack the fidelity to represent continuous geometric information. While recent methods use latent representations to enhance reasoning, relying on a single latent type cannot adapt to the diversity of spatial tasks, leading to misalignment in complex geometric scenarios. To address these limitations, we propose GeoAnchor, an interleaved text-latent reasoning framework. GeoAnchor decomposes 3D spatial information into three complementary components: position latents for object grounding, direction latents for relational orientation, and geometry latents for scene structure. These components are recombined in a structured space to construct local evidence while capturing global context, enabling dynamic and interpretable reasoning. Furthermore, we introduce a collaborative training strategy that guides the model from local spatial perception to comprehensive 3D understanding. Extensive experiments on diverse and complex 3D reasoning tasks demonstrate that GeoAnchor outperforms the state of the art, validating its effectiveness and generalization capabilities.

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

33 extracted references · 19 linked inside Pith

  1. [2]

    Spatialthinker: Reinforcing 3d reasoning in multimodal llms via spatial rewards.arXiv preprint arXiv:2511.07403,

    Hunar Batra, Haoqin Tu, Hardy Chen, Yuanze Lin, Cihang Xie, and Ronald Clark. Spatialthinker: Reinforcing 3d reasoning in multimodal llms via spatial rewards.arXiv preprint arXiv:2511.07403,

    Pith/arXiv arXiv

  2. [3]

    Soft tokens, hard truths.arXiv preprint arXiv:2509.19170,

    Natasha Butt, Ariel Kwiatkowski, Ismail Labiad, Julia Kempe, and Yann Ollivier. Soft tokens, hard truths.arXiv preprint arXiv:2509.19170,

    arXiv

  3. [4]

    Scaling spatial intelligence with multimodal foundation models.arXiv preprint arXiv:2511.13719,

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

    arXiv

  4. [6]

    Molmo2: Open weights and data for vision-language models with video understanding and grounding.arXiv preprint arXiv:2601.10611,

    Christopher Clark, Jieyu Zhang, Zixian Ma, Jae Sung Park, Mohammadreza Salehi, Rohun Tripathi, Sangho Lee, Zhongzheng Ren, Chris Dongjoo Kim, Yinuo Yang, et al. Molmo2: Open weights and data for vision-language models with video understanding and grounding.arXiv preprint arXiv:2601.10611,

    Pith/arXiv arXiv

  5. [9]

    Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction.arXiv preprint arXiv:2505.20279,

    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.arXiv preprint arXiv:2505.20279,

    Pith/arXiv arXiv

  6. [10]

    Spatial reasoning with vision-language models in ego-centric multi-view scenes.arXiv preprint arXiv:2509.06266,

    Mohsen Gholami, Ahmad Rezaei, Zhou Weimin, Sitong Mao, Shunbo Zhou, Yong Zhang, and Mohammad Akbari. Spatial reasoning with vision-language models in ego-centric multi-view scenes.arXiv preprint arXiv:2509.06266,

    arXiv

  7. [11]

    Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al

    URL https://storage.googleapis.com/deepmind-media/ Model-Cards/Gemini-3-Pro-Model-Card.pdf. Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, et al. Glm-4.5 v and glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning.arXiv preprint arXiv:2507.01006,

    Pith/arXiv arXiv

  8. [12]

    G2vlm: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning.arXiv preprint arXiv:2511.21688,

    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.arXiv preprint arXiv:2511.21688,

    arXiv

  9. [13]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276,

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

    Pith/arXiv arXiv

  10. [14]

    Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262,

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262,

    Pith/arXiv arXiv

  11. [15]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

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

    Pith/arXiv arXiv

  12. [16]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Pith/arXiv arXiv

  13. [18]

    Ssr: Enhancing depth perception in vision-language models via rationale-guided spatial reasoning.arXiv preprint arXiv:2505.12448, 2025a

    Yang Liu, Ming Ma, Xiaomin Yu, Pengxiang Ding, Han Zhao, Mingyang Sun, Siteng Huang, and Donglin Wang. Ssr: Enhancing depth perception in vision-language models via rationale-guided spatial reasoning.arXiv preprint arXiv:2505.12448, 2025a. Yuhong Liu, Beichen Zhang, Yuhang Zang, Yuhang Cao, Long Xing, Xiaoyi Dong, Haodong Duan, Dahua Lin, and Jiaqi Wang. ...

    arXiv

  14. [19]

    Spacer: Reinforcing mllms in video spatial reasoning.arXiv preprint arXiv:2504.01805,

    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,

    Pith/arXiv arXiv

  15. [20]

    Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747,

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747,

    Pith/arXiv arXiv

  16. [21]

    Chain-of-visual-thought: Teaching vlms to see and think better with continuous visual tokens.arXiv preprint arXiv:2511.19418,

    Yiming Qin, Bomin Wei, Jiaxin Ge, Konstantinos Kallidromitis, Stephanie Fu, Trevor Darrell, and XuDong Wang. Chain-of-visual-thought: Teaching vlms to see and think better with continuous visual tokens.arXiv preprint arXiv:2511.19418,

    arXiv

  17. [22]

    URLhttps://qwen.ai/blog?id=qwen3.5. 13 GeoAnchor: Collaborative Reasoning via Latent Decomposition for 3D Spatial Understanding Arijit Ray, Ahmed Abdelkader, Chengzhi Mao, Bryan A Plummer, Kate Saenko, Ranjay Krishna, Leonidas Guibas, and Wen-Sheng Chu. Mull-tokens: Modality-agnostic latent thinking.arXiv preprint arXiv:2512.10941,

    Pith/arXiv arXiv

  18. [23]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

    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,

    Pith/arXiv arXiv

  19. [24]

    Codi: Compressing chain-of-thought into continuous space via self-distillation

    Zhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu, Yali Du, and Yulan He. Codi: Compressing chain-of-thought into continuous space via self-distillation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 677–693,

    2025

  20. [25]

    Kimi-vl technical report.arXiv preprint arXiv:2504.07491,

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

    Pith/arXiv arXiv

  21. [27]

    Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence.arXiv preprint arXiv:2505.23747, 2025a

    Diankun Wu, Fangfu Liu, Yi-Hsin Hung, and Yueqi Duan. Spatial-mllm: Boosting mllm capabilities in visual-based spatial intelligence.arXiv preprint arXiv:2505.23747, 2025a. Junfei Wu, Jian Guan, Kaituo Feng, Qiang Liu, Shu Wu, Liang Wang, Wei Wu, and Tieniu Tan. Reinforcing spatial reasoning in vision-language models with interwoven thinking and visual dra...

    Pith/arXiv arXiv

  22. [28]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a

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

    Pith/arXiv arXiv

  23. [30]

    Hybrid latent reasoning via reinforcement learning.arXiv preprint arXiv:2505.18454,

    14 GeoAnchor: Collaborative Reasoning via Latent Decomposition for 3D Spatial Understanding Zhenrui Yue, Bowen Jin, Huimin Zeng, Honglei Zhuang, Zhen Qin, Jinsung Yoon, Lanyu Shang, Jiawei Han, and Dong Wang. Hybrid latent reasoning via reinforcement learning.arXiv preprint arXiv:2505.18454,

    arXiv

  24. [31]

    Actial: Activate spatial reasoning ability of multimodal large language models.arXiv preprint arXiv:2511.01618,

    Xiaoyu Zhan, Wenxuan Huang, Hao Sun, Xinyu Fu, Changfeng Ma, Shaosheng Cao, Bohan Jia, Shaohui Lin, Zhenfei Yin, Lei Bai, et al. Actial: Activate spatial reasoning ability of multimodal large language models.arXiv preprint arXiv:2511.01618,

    arXiv

  25. [32]

    Spatial understanding from videos: Structured prompts meet simulation data.arXiv preprint arXiv:2506.03642, 2025a

    Haoyu Zhang, Meng Liu, Zaijing Li, Haokun Wen, Weili Guan, Yaowei Wang, and Liqiang Nie. Spatial understanding from videos: Structured prompts meet simulation data.arXiv preprint arXiv:2506.03642, 2025a. Jiahui Zhang, Yurui Chen, Yanpeng Zhou, Yueming Xu, Ze Huang, Jilin Mei, Junhui Chen, Yu-Jie Yuan, Xinyue Cai, Guowei Huang, et al. From flatland to spac...

    arXiv

  26. [2008]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025a. Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, and Yisen Wang. Monet: Reason...

    arXiv

  27. [2014]

    Latent visual reasoning.arXiv preprint arXiv:2509.24251, 2025a

    Bangzheng Li, Ximeng Sun, Jiang Liu, Ze Wang, Jialian Wu, Xiaodong Yu, Hao Chen, Emad Barsoum, Muhao Chen, and Zicheng Liu. Latent visual reasoning.arXiv preprint arXiv:2509.24251, 2025a. 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 ...

    Pith/arXiv arXiv

  28. [2017]

    Graid: Enhancing spatial reasoning of vlms through high-fidelity data generation.arXiv preprint arXiv:2510.22118,

    Karim Elmaaroufi, Liheng Lai, Justin Svegliato, Yutong Bai, Sanjit A Seshia, and Matei Zaharia. Graid: Enhancing spatial reasoning of vlms through high-fidelity data generation.arXiv preprint arXiv:2510.22118,

    arXiv

  29. [2020]

    Opendrivevla: Towards end-to-end autonomous driving with large vision language action model.arXiv preprint arXiv:2503.23463,

    Xingcheng Zhou, Xuyuan Han, Feng Yang, Yunpu Ma, Volker Tresp, and Alois Knoll. Opendrivevla: Towards end-to-end autonomous driving with large vision language action model.arXiv preprint arXiv:2503.23463,

    arXiv

  30. [2023]

    Minicpm-v 4.5: Cooking efficient mllms via architecture, data, and training recipe

    Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, Wenshuo Ma, Zhihui He, Tianchi Cai, Weize Chen, Yuxiang Huang, Yuanqian Zhao, et al. Minicpm-v 4.5: Cooking efficient mllms via architecture, data, and training recipe. arXiv preprint arXiv:2509.18154,

    Pith/arXiv arXiv

  31. [2024]

    Think with 3d: Geometric imagination grounded spatial reasoning from limited views.arXiv preprint arXiv:2510.18632,

    Zhangquan Chen, Manyuan Zhang, Xinlei Yu, Xufang Luo, Mingze Sun, Zihao Pan, Yan Feng, Peng Pei, Xunliang Cai, and Ruqi Huang. Think with 3d: Geometric imagination grounded spatial reasoning from limited views.arXiv preprint arXiv:2510.18632,

    arXiv

  32. [2025]

    Llava-onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    Xiang An, Yin Xie, Kaicheng Yang, Wenkang Zhang, Xiuwei Zhao, Zheng Cheng, Yirui Wang, Songcen Xu, Changrui Chen, Didi Zhu, et al. Llava-onevision-1.5: Fully open framework for democratized multimodal training.arXiv preprint arXiv:2509.23661,

    Pith/arXiv arXiv

  33. [2026]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261,

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261,

    Pith/arXiv arXiv