Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

The paper claims that explicit geometry-supervised patch alignment plus action-then-answer reinforcement lifts a 3-billion-parameter multimodal model 14.2 points on average across three multi-image spatial reasoning benchmarks and matches m

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

2026-08-03 03:10 UTC pith:AXR2FR4U

load-bearing objection Potentially useful training recipe, but the headline SPAR-Bench-MV result is not yet cleanly established because training and evaluation data come from the same SPAR family, and the split is never disclosed. the 5 major comments →

arxiv 2602.08735 v3 pith:AXR2FR4U submitted 2026-02-09 cs.CV

From Correspondence to Actions: Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models

Masanari Oi , Koki Maeda , Ryuto Koike , Daisuke Oba , Nakamasa Inoue , Naoaki Okazaki This is my paper
classification cs.CV
keywords multi-image spatial reasoningcross-view correspondencepatch-level alignmentaction-then-answer reasoningGRPO reinforcement learningviewpoint transformationmultimodal large language modelsverifiable rewards
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 is trying to establish that the missing ingredient in multi-image spatial reasoning for multimodal large language models is not model size but explicit training signals for two human-like mechanisms: seeing which regions in different views are the same place, and composing a sequence of viewpoint changes. To make that case, it introduces HATCH, a training recipe with two stages. The first stage uses training-only camera poses and depth maps to build soft patch-to-patch correspondence targets and drives the image encoder to align features for corresponding regions. The second stage forces the model to write explicit viewpoint-transition actions (turns and moves) before the answer and reinforces both action and answer with verifiable rewards. On three multi-image benchmarks, a 3-billion-parameter model trained this way improves by 14.2 points on average and reaches parity with much larger systems on two benchmarks; a sympathetic reader would care because it suggests spatial reasoning can be bought with better supervision rather than more parameters.

Core claim

The central claim is that cross-view correspondence can be supervised directly at the patch level, and that explicit action generation then converts alignment into reliable answers. PaStA uses depth maps, camera intrinsics, and poses to build a symmetric patch-overlap matrix, turns it into a soft target, and trains the image encoder so patch-feature similarities reproduce it. ActoR makes the model emit JSON camera actions before the answer and optimizes those actions with group relative policy optimization under answer, action, and format rewards. Reported gains: 24.9→53.6, 37.8→50.2, and 25.6→27.0 on the three benchmarks; ablations show both stages matter, with PaStA more important for view

What carries the argument

The load-bearing object is the symmetric patch-overlap matrix S: entry S[i,j] measures how consistently patch i in one view projects into patch j in the other, using depth and camera poses. PaStA converts S into a soft target distribution and trains the vision encoder to match it from patch-feature cosine similarities. The second mechanism is the explicit action sequence: the model must output JSON camera operations for every image pair before the answer, and GRPO rewards action accuracy, answer accuracy, and format. Together they force correspondence and stepwise viewpoint composition into the training signal.

Load-bearing premise

The headline gains rest on the unstated assumption that the 10,000 multi-image training samples selected from the source dataset are disjoint from, or at least distributionally distinct from, the evaluation samples, and that the comparison models were trained on matched data; the paper does not report such an exclusion.

What would settle it

Re-run HATCH on the same 10,000-sample split, then evaluate on a held-out multi-view test set from which every training sample and near-duplicate has been removed, while training comparably sized baselines on the same split; if the reported multi-view gain (for example, the +17.8 point rise on the primary benchmark) collapses to noise, the central claim is falsified, and if it persists, the overlap concern is not the source.

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

If this is right

  • A 3-billion-parameter model can match much larger proprietary systems on multi-view spatial reasoning, so capacity is not the binding constraint.
  • Camera poses and depth are used only at training time; inference needs only images and text, so the recipe fits existing deployment pipelines.
  • Explicit JSON actions outperform free-form natural-language chain-of-thought as intermediate reasoning on the tested spatial questions, because actions are quantitative and verifiable.
  • PaStA and ActoR are complementary: removing PaStA mainly hurts viewpoint-change questions, while removing ActoR mainly hurts depth and distance questions.
  • The single-image benchmarks improve along with the multi-image ones, so the training does not trade away monocular spatial understanding.

Where Pith is reading between the lines

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

  • One consequence the authors leave implicit: the same geometry-derived soft target plus verifiable-action recipe may transfer to other perception tasks that decompose into correspondence plus composition, such as object tracking across viewpoints or embodied navigation.
  • Because the action vocabulary is limited to rotations and translations, a natural extension is object-centric actions (for example, 'move toward object X'), which might help the benchmark categories where all models sit near chance.
  • The grid-resolution result—accuracy peaks at a 4-by-4 patch grid and falls at finer resolutions—suggests adaptive or attention-based correspondence supervision could push the method further; the paper itself lists attention supervision as future work.
  • A natural stress test is to re-run with an explicit train/test split and matched baselines, since the training data is drawn from the same benchmark suite as the evaluation.

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

5 major / 5 minor

Summary. The paper proposes HATCH, a two-stage training framework for multi-image spatial reasoning in multimodal large language models. In the first stage, PaStA uses camera poses, intrinsics, and depth maps to build a patch-level spatial-overlap matrix and aligns patch features across views through a cross-entropy objective. In the second stage, ActoR trains the model to emit explicit JSON viewpoint-transition actions before the final answer, using cold-start SFT followed by GRPO with action, answer, and format rewards. Experiments on SPAR-Bench-MV, MindCube-Tiny, and MMSI-Bench report substantial improvements over Qwen2.5-VL-3B and competitive performance against much larger models, together with ablations and single-image benchmark results.

Significance. If the reported numbers are trustworthy, the contribution is practically valuable: explicit geometry-supervised correspondence is a clean and low-cost training signal, and action-then-answer GRPO is a novel, interpretable intermediate objective. The MindCube-Tiny gain (+14.9) is on a benchmark outside the SPAR training suite, and the component ablations are broadly consistent with the paper's mechanistic story. However, the headline SPAR-Bench-MV result is not yet established because the paper does not demonstrate train/eval disjointness, selects a hyperparameter on the evaluation benchmark, and mixes full-benchmark and tiny-subset scores in a headline comparison. These issues must be resolved before the central claim can be accepted.

major comments (5)
  1. [§4.1, footnote 1, Appendix E.1] The paper trains on 10,000 multi-image samples 'selected from SPAR-7M' and evaluates on SPAR-Bench-MV, defined as multi-image samples 'extracted from SPAR-Bench'. Since SPAR-7M and SPAR-Bench come from the same SPAR suite, the paper must state explicitly that the SPAR-Bench-MV evaluation samples were excluded from the 10,000-sample training pool. It also never states whether the same-backbone baselines (SpatialLadder-3B, Spatial-MLLM-4B) were trained on matched, overlapping, or absent SPAR-family data. Without this information, the +17.8 point gain on SPAR-Bench-MV and the claim of matching GPT-5.2 are confounded by possible data exposure. Please provide the exact disjointness procedure and, ideally, report results on a held-out split constructed before training.
  2. [Figure 4 / §4.3] The PaStA grid resolution n is selected by maximizing average accuracy on SPAR-Bench-MV, which is the same benchmark used for the main result and ablations. This is hyperparameter selection on the evaluation set, and the reported numbers with n=4 are therefore optimistically biased. Please select n on a validation split or fix it a priori; if that is infeasible, report the sensitivity across n for all benchmarks and qualify the headline results accordingly.
  3. [Table 1 / §4.2] GPT-5.2 and Gemini-3-Pro are evaluated on SPAR-Bench-Tiny-MV (gray italic entries), not on the full SPAR-Bench-MV, yet the text states that 'HATCH (53.6%) matches GPT-5.2 (52.6%) on SPAR-Bench-MV'. These scores are not directly comparable. Either evaluate HATCH on the same SPAR-Bench-Tiny-MV subset, or rephrase the proprietary-model comparison and clearly mark it as approximate.
  4. [§3.3, Eq. (2)] For image patches with no geometric overlap, S[i,:]=0, so p(j|i) becomes a uniform distribution over all patches in the other image. Unless such rows are masked or explicitly handled, the cross-entropy loss in Eqs. (4)-(5) trains the encoder to produce uniform similarity for non-corresponding patches, which is contrary to the intended alignment. Please specify how zero-overlap rows are treated and provide an ablation or analysis showing the effect of masking them.
  5. [Tables 1-4 / §4.1] No error bars, confidence intervals, or multiple-seed results are reported. On MMSI-Bench the overall gain is only +1.4 points (25.6 to 27.0), and several ablation differences are between 0 and 4 points; without variance estimates these differences cannot be interpreted. Please report at least three independent training runs or bootstrap confidence intervals for the key comparisons, especially the small MMSI-Bench gain and the ablation deltas.
minor comments (5)
  1. [Table 1 / Table 4] The baseline name 'LLaV A-OneVision-4B' contains an awkward spacing; use the standard name consistently.
  2. [§4.2] 'Qwen-2.5-VL-3B' appears in one place while the rest of the paper uses 'Qwen2.5-VL-3B'; standardize the notation.
  3. [Appendix D] The hyperparameter values for τ1, τ2, t, τt, and τr are not listed in the implementation-details appendix. Please provide these values for reproducibility.
  4. [Figure 5 caption] The caption says 'success and failure cases' but the surrounding text and figure appear to show two successful examples. Clarify whether a failure case is included and what it illustrates.
  5. [Appendix B.2] The example JSON action list ends with a trailing comma after the final 'turn_down_deg' object. Although likely a formatting artifact, it should be cleaned up to avoid ambiguity in the schema description.

Circularity Check

0 steps flagged

No circularity found: PaStA and ActoR are supervised by external geometry and benchmark labels; the SPAR train/eval overlap is a data-contamination concern, not a derivation-level circularity.

full rationale

The paper's derivation chain is not circular. PaStA's target correspondence distribution p(j|i) is constructed from camera intrinsics, camera poses, and depth maps via the overlap matrix S (Eqs. 1-2), while q(j|i) is the encoder's predicted similarity distribution (Eq. 3); the cross-entropy loss (Eqs. 4-5) therefore supervises the encoder against externally computed geometric ground truth, not against the quantity it is meant to predict. ActoR's actions are generated from relative camera poses in cold-start SFT, and its rewards (Eq. 8) compare predicted actions to gold pose-derived actions and answers to benchmark labels, again providing independent supervision. No equation in Section 3 sets a model output equal to its own training target, and no load-bearing argument is reduced to a self-citation; the cited benchmarks and prior methods are external works. The main concerns are evaluation-validity issues rather than circularity: HATCH trains on 10,000 samples 'selected from SPAR-7M' while SPAR-Bench-MV is 'extracted from SPAR-Bench', with no stated train/eval disjointness, and the PaStA grid size n is selected using SPAR-Bench-MV accuracy (Figure 4). These could undermine the empirical claim if evaluation items overlap training data, but they do not make the method's derivation equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on one tuned-on-test hyperparameter (grid n), several unreported temperatures/thresholds, and an unstated train/test disjointness assumption for the SPAR family. No new physical entities are introduced; the action JSON format is an output representation, not an entity.

free parameters (6)
  • PaStA grid size n = 4
    Tuned by maximizing average accuracy on SPAR-Bench-MV (Figure 4); selecting a hyperparameter on the evaluation benchmark inflates the reported result.
  • PaStA temperatures τ1, τ2 = not reported
    Softmax temperatures in Eqs. (2)-(3) control target sharpness and feature-similarity sharpness; no values given, so the alignment objective is not fully specified.
  • depth consistency threshold t = not reported
    Defines geometric consistency in the overlap matrix in §3.3; no value reported, though it controls which projections count as corresponding.
  • action reward temperatures τt, τr = not reported
    Scale the pose error into Ract-acc in Eq. (10); no values reported, and reward scale affects GRPO behavior.
  • reward weights λ1, λ2, λ3 = 1, 1, 1
    Set equal by hand (§D.1); chosen rather than derived.
  • learning rate / KL coefficient = 1e-5 / 0.01
    Fixed training hyperparameters reported in §D.1; standard but hand-chosen.
axioms (4)
  • domain assumption Human multi-image spatial reasoning is well modeled by cross-view correspondence plus stepwise viewpoint transformation
    Taken from cognitive refs in §1; motivates the two objectives but is not derived or tested.
  • domain assumption Camera poses and depth maps in the training data are accurate enough for patch projection and action supervision
    PaStA's overlap matrix and ActoR's teacher actions and action rewards all rest on this; no sensitivity analysis.
  • ad hoc to paper The fixed atomic action vocabulary and the three-stage execution order (rotate, translate, rotate) can express the needed viewpoint transitions
    Design choice in §3.4/Appendix B.3; there are infinitely many equivalent trajectories between two views, and no theorem guarantees this discretization covers them.
  • domain assumption The SPAR-Bench-MV evaluation samples are disjoint from the SPAR-7M training samples
    Unstated in §4.1; training data is selected from SPAR-7M and the test set is extracted from the same SPAR benchmark family, so disjointness is load-bearing for the headline gain.

pith-pipeline@v1.3.0-alltime-deepseek · 17520 in / 15093 out tokens · 164076 ms · 2026-08-03T03:10:03.575589+00:00 · methodology

0 comments
read the original abstract

While multimodal large language models (MLLMs) have made substantial progress in single-image spatial reasoning, multi-image spatial reasoning, which requires integration of information from multiple viewpoints, remains challenging. Cognitive studies suggest that humans address such tasks through two mechanisms: cross-view correspondence, which identifies regions across different views that correspond to the same physical locations, and stepwise viewpoint transformation, which composes relative viewpoint changes sequentially. However, existing studies incorporate these mechanisms only partially and often implicitly, without explicit supervision for both. We propose Human-Aware Training for Cross-view correspondence and viewpoint cHange (HATCH), a training framework with two complementary objectives: (1) Patch-Level Spatial Alignment, which encourages patch representations to align across views for spatially corresponding regions, and (2) Action-then-Answer Reasoning, which requires the model to generate explicit viewpoint transition actions before predicting the final answer. Experiments on three benchmarks demonstrate that HATCH consistently outperforms baselines of comparable size by a clear margin and achieves competitive results against much larger models, while preserving single-image reasoning capabilities.

Figures

Figures reproduced from arXiv: 2602.08735 by Daisuke Oba, Koki Maeda, Masanari Oi, Nakamasa Inoue, Naoaki Okazaki, Ryuto Koike.

Figure 1
Figure 1. Figure 1: Two cognitive mechanisms underlying multi-image spa￾tial reasoning: (a) cross-view correspondence, identifying regions across views that correspond to the same physical locations; (b) stepwise viewpoint transformation, composing relative viewpoint changes (e.g., rotations) in a sequential manner. answer by integrating information across the views. Unlike single-image settings, this task demands not only in… view at source ↗
Figure 2
Figure 2. Figure 2: HATCH pipeline overview. HATCH consists of two components: (i) Patch-Level Spatial Alignment (PaStA) to learn cross-view correspondence, (ii) Action-then-Answer Reasoning (ActoR) to perform stepwise viewpoint transformation via explicit actions. Feature-Level Alignment. Feature-level alignment be￾tween vision and language is widely used to learn trans￾ferable visual representations and enable zero-shot gen… view at source ↗
Figure 4
Figure 4. Figure 4: Grid resolution analysis for PaStA. SPAR-Bench-MV average accuracy improves up to n = 4 and drops for n ≥ 5, indicating that overly fine grids hurt correspondence learning [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics during GRPO training of ActoR. Action and QA accuracy rewards are shown for HATCH (yellow) and an ablated variant without PaStA (blue) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative success and failure cases for different reasoning modalities. Compared with natural language reasoning, action-based reasoning (HATCH) yields explicit, quantitative camera operations that more directly support correct multi-image spatial inference [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Vesta: A Generalist Embodied Reasoning Model

    cs.RO 2026-06 unverdicted novelty 6.0

    Vesta is a unified embodied generalist model that outperforms specialist baselines by over 20% on average and improves real-world robotic task success by over 35%.

  2. Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models

    cs.CV 2026-06 unverdicted novelty 6.0

    An agentic VLM approach with dynamic cognitive maps and Spatial Assertion Codes reaches 80.5% accuracy on MindCube, gaining 29.5 points on rotation tasks via dense-reward RL.

Reference graph

Works this paper leans on

21 extracted references · 14 linked inside Pith · cited by 2 Pith papers

  1. [1]

    LLaV A- OneVision-1.5: Fully Open Framework for Democratized Multimodal Training.arXiv preprint 2509.23661,

    An, X., Xie, Y ., Yang, K., Zhang, W., Zhao, X., Cheng, Z., Wang, Y ., Xu, S., Chen, C., Zhu, D., Wu, C., Tan, H., Li, C., Yang, J., Yu, J., Wang, X., Qin, B., Wang, Y ., Yan, Z., Feng, Z., Liu, Z., Li, B., and Deng, J. LLaV A- OneVision-1.5: Fully Open Framework for Democratized Multimodal Training.arXiv preprint 2509.23661,

    Pith/arXiv arXiv

  2. [4]

    DeepSeek-R1: Incentivizing Reasoning Ca- pability in LLMs via Reinforcement Learning.arXiv preprint 2501.12948,

    DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Ca- pability in LLMs via Reinforcement Learning.arXiv preprint 2501.12948,

    Pith/arXiv arXiv

  3. [6]

    Spatial Reasoning with Vision- Language Models in Ego-Centric Multi-View Scenes

    Gholami, M., Rezaei, A., Weimin, Z., Mao, S., Zhou, S., Zhang, Y ., and Akbari, M. Spatial Reasoning with Vision- Language Models in Ego-Centric Multi-View Scenes. arXiv preprint 2509.06266,

    arXiv

  4. [7]

    9 Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models Gu, X., Lin, T.-Y ., Kuo, W., and Cui, Y

    Accessed: 2026-01-27. 9 Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models Gu, X., Lin, T.-Y ., Kuo, W., and Cui, Y . Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. InProc. International Conference on Learn- ing Representations (ICLR),

    2026

  5. [9]

    3DRS: MLLMs Need 3D-Aware Representation Supervision for Scene Understanding.arXiv preprint 2506.01946,

    Huang, X., Wu, J., Xie, Q., and Han, K. 3DRS: MLLMs Need 3D-Aware Representation Supervision for Scene Understanding.arXiv preprint 2506.01946,

    arXiv

  6. [10]

    SpatialLadder: Progressive Training for Spatial Reasoning in Vision- Language Models.arXiv preprint 2510.08531,

    Li, H., Li, D., Wang, Z., Yan, Y ., Wu, H., Zhang, W., Shen, Y ., Lu, W., Xiao, J., and Zhuang, Y . SpatialLadder: Progressive Training for Spatial Reasoning in Vision- Language Models.arXiv preprint 2510.08531,

    arXiv

  7. [12]

    Coarse Correspondences Elicit 3D Spacetime Understanding in Multimodal Language Model.arXiv preprint 2408.00754,

    Liu, B., Dong, Y ., Wang, Y ., Rao, Y ., Tang, Y ., Ma, W.- C., and Krishna, R. Coarse Correspondences Elicit 3D Spacetime Understanding in Multimodal Language Model.arXiv preprint 2408.00754,

    Pith/arXiv arXiv

  8. [13]

    GPT-4o System Card.arXiv preprint 2410.21276,

    OpenAI. GPT-4o System Card.arXiv preprint 2410.21276,

    Pith/arXiv arXiv

  9. [14]

    OpenAI GPT-5 System Card.arXiv preprint 2601.03267,

    OpenAI. OpenAI GPT-5 System Card.arXiv preprint 2601.03267,

    Pith/arXiv arXiv

  10. [15]

    SpaceR: Reinforcing MLLMs in Video Spatial Reasoning.arXiv preprint 2504.01805,

    Ouyang, K., Liu, Y ., Wu, H., Liu, Y ., Zhou, H., Zhou, J., Meng, F., and Sun, X. SpaceR: Reinforcing MLLMs in Video Spatial Reasoning.arXiv preprint 2504.01805,

    Pith/arXiv arXiv

  11. [17]

    Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence

    Wu, D., Liu, F., Hung, Y .-H., and Duan, Y . Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence. InProc. Annual Conference on Neural In- formation Processing Systems (NeurIPS), 2025a. Wu, H., Huang, X., Chen, Y ., Zhang, Y ., Wang, Y ., and Xie, W. SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence.arXiv pre...

    Pith/arXiv arXiv

  12. [19]

    MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence.arXiv preprint 2505.23764,

    Yang, S., Xu, R., Xie, Y ., Yang, S., Li, M., Lin, J., Zhu, C., Chen, X., Duan, H., Yue, X., Lin, D., Wang, T., and Pang, J. MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence.arXiv preprint 2505.23764,

    Pith/arXiv arXiv

  13. [20]

    L., Cao, Y ., and Narasimhan, K

    Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y ., and Narasimhan, K. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. InProc. Annual Conference on Neural Information Processing Systems (NeurIPS), 2023a. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y . ReAct: Synergizing Reasoning and Ac...

    arXiv

  14. [21]

    14 Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models D

    Rformat encourages the model to produce outputs that are directly consumable by downstream parsers and reward functions. 14 Human-Like Multi-Image Spatial Reasoning in Multi-modal Large Language Models D. Implementation Details D.1. Hyperparameter Settings This subsection summarizes the hyperparameters used across different training stages. Unless otherwi...

    2025

  15. [1983]

    K., Wu, Y ., and Guo, D

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y . K., Wu, Y ., and Guo, D. DeepSeek- Math: Pushing the Limits of Mathematical Reasoning in Open Language Models.arXiv preprint 2402.03300,

    Pith/arXiv arXiv

  16. [1997]

    VLM-3R: Vision-Language Models Augmented with Instruction- Aligned 3D Reconstruction.arXiv preprint 2505.20279,

    Fan, Z., Zhang, J., Li, R., Zhang, J., Chen, R., Hu, H., Wang, K., Qu, H., Wang, D., Yan, Z., Xu, H., Theiss, J., Chen, T., Li, J., Tu, Z., Wang, Z., and Ranjan, R. VLM-3R: Vision-Language Models Augmented with Instruction- Aligned 3D Reconstruction.arXiv preprint 2505.20279,

    Pith/arXiv arXiv

  17. [2004]

    G 2VLM: Geome- try Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning.arXiv preprint 2511.21688,

    Hu, W., Lin, J., Long, Y ., Ran, Y ., Jiang, L., Wang, Y ., Zhu, C., Xu, R., Wang, T., and Pang, J. G 2VLM: Geome- try Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning.arXiv preprint 2511.21688,

    arXiv

  18. [2021]

    Xu, R., Wang, W., Tang, H., Chen, X., Wang, X., Chu, F.-J., Lin, D., Feiszli, M., and Liang, K. J. Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models.arXiv preprint 2505.17015,

    Pith/arXiv arXiv

  19. [2022]

    Improved Visual-Spatial Reasoning via R1- Zero-Like Training.arXiv preprint 2504.00883,

    Liao, Z., Xie, Q., Zhang, Y ., Kong, Z., Lu, H., Yang, Z., and Deng, Z. Improved Visual-Spatial Reasoning via R1- Zero-Like Training.arXiv preprint 2504.00883,

    Pith/arXiv arXiv

  20. [2024]

    SD-VLM: Spatial Measuring and Un- derstanding with Depth-Encoded Vision-Language Mod- els

    Chen, P., Lou, Y ., Cao, S., Guo, J., Fan, L., Wu, Y ., Yang, L., Ma, L., and Ye, J. SD-VLM: Spatial Measuring and Un- derstanding with Depth-Encoded Vision-Language Mod- els. InProc. Annual Conference on Neural Information Processing Systems (NeurIPS), 2025a. Chen, Z., Wang, W., Cao, Y ., Liu, Y ., Gao, Z., Cui, E., Zhu, J., Ye, S., Tian, H., Liu, Z., Gu...

    Pith/arXiv arXiv

  21. [2025]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y ., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y ., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., and Lin, J. Qwen2.5-VL Technical Report. arXiv preprint 2502.13923,

    Pith/arXiv arXiv