Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Disentangling 3D Modeling from Spatial Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Explicitly separating 3D perception from reasoning — frozen experts build structured evidence, a LoRA-tuned LLM reasons over it — is a scalable, effective alternative to end-to-end spatial modeling.

desk verdict Frozen perception + LoRA LLM over serialized 3D evidence is a real alternative to end-to-end spatial training; the missing evidence-accuracy analysis on the SOTA benchmarks is the main reviewable gap. read the letter →

arxiv 2608.05242 v1 pith:O3RDJMH6 submitted 2026-08-05 cs.LG cs.CV

Haoze Sun , Jiequan Cui , Qingshan Xu , Richang Hong This is my paper · ORCID
classification cs.LGcs.CV
keywords spatialreasoning3DperceptionstructuredevidenceLoRAfine-tuningvision-languagemodelsVQAbenchmarksmetricdepthestimationmodularity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DiSR is a framework that separates the problem of spatial reasoning into two stages: first, frozen off-the-shelf perception models reconstruct an image into structured 3D evidence (object centers, sizes, orientations in a shared gravity-aligned frame); second, a LoRA-tuned LLM answers questions by reading only a serialized text version of that evidence, with no image input. The paper's central claim is that this explicit separation is a scalable, effective alternative to end-to-end models that jointly learn perception and reasoning through large-scale 3D VQA training. On 3DSRBench and SPAR-Bench, DiSR sets new state-of-the-art results (67.62% and 46.33%, outperforming the previous best by 3.77% and 1.70%), while training on 0.33M samples for 59 GPU-hours on a single RTX 4080 SUPER. The paper argues this shows strong spatial reasoning does not require jointly learning 3D geometry, and that the explicit intermediate representation makes errors diagnosable: with ground-truth 3D evidence, the reasoning module reaches near-perfect accuracy.

What carries the argument

The load-bearing object is the structured 3D evidence tuple $e_i = (c_i, d_i, \sigma_i, \rho_i, f_i, l_i)$ in a shared gravity-aligned coordinate system, serialized into text. It functions as a fixed, interpretable interface between the frozen perception stack and the LoRA-tuned LLM: the LLM only needs to perform compositional reasoning over explicit numbers and axes, while perception errors remain visible and attributable. This interface is what makes the disentanglement claim testable — the paper can swap ground-truth evidence in and show reasoning is at ceiling.

What would settle it

Give DiSR the same frozen perception stack but replace its small-object evidence with higher-quality depth or masks (or use ground-truth 3D evidence as the paper already does) on CV-Bench-3D Depth; if accuracy on the smallest-object quintile does not rise from 84.17% toward the 89.17% of Qwen3-VL-8B-Instruct, the claim that perception quality is the bottleneck would be wrong. Conversely, if an end-to-end MLLM trained on large-scale 3D VQA still beats DiSR on a benchmark designed to have no 2D shortcuts even when DiSR receives ground-truth evidence, the separation paradigm would lose its central advantage.

Watch

Extended reading notes

Core claim

The central discovery is that spatial reasoning can be bought by composition rather than learned end-to-end. DiSR's perception stack — SAM for masks, Metric3D for metric depth, WildCamera and Perspective Fields for a shared gravity-aligned camera frame, and OrientAnything for object orientation — converts each queried object into a tuple $e_i = (c_i, d_i, \sigma_i, \rho_i, f_i, l_i)$ of 3D center, extent, size descriptors, and front/left axes. This evidence is serialized as text and fed, together with the question, to an LLM that has been lightly adapted with LoRA; the LLM never sees the image. Over 0.33M training samples and 59 GPU-hours, this pipeline reaches 67.62% on 3DSRBench and 46.33% on SPAR-Bench, surpassing large-scale spatial models such as HiSpatial and SpatialReasoner, and it remains competitive on CV-Bench-3D while preserving the base LLM's general reasoning scores. A diagnosability study shows that the reasoning stage is essentially solved once accurate 3D evidence is supplied: replacing predicted evidence with ground truth yields about 100% accuracy on CV-Bench-3D Depth and Distance, locating the remaining bottleneck in perception, especially for small objects.

Load-bearing premise

The load-bearing premise is that off-the-shelf perception models recover accurate metric depth, camera pose, masks, and orientations in a shared gravity-aligned frame, and the paper's own size-stratified results show this premise breaks down for small objects in low-quality images, with a 15.00% evidence-construction error rate in the smallest-object group versus 1.67% in the largest.

Editorial extensions

If this is right

  • A spatial reasoning system can reach state-of-the-art benchmark scores without large-scale 3D VQA training; 0.33M samples and 59 GPU-hours suffice for the reasoning component.
  • Improving perception (depth, camera pose, masks, orientation) directly improves spatial reasoning, because the reasoning module is already near ceiling; the CV-Bench-3D gain from 92.25% to 93.75% with ground-truth grounding illustrates this.
  • The framework is modular: swapping the base LLM (Qwen2.5-VL-7B to Qwen3-VL-8B) improves 3DSRBench from 65.52% to 67.62% without redesign.
  • General visual reasoning is preserved after spatial fine-tuning: DiSR matches its base model on MMBench, GQA, POPE, SEED, and RealWorldQA, unlike some prior spatial specialists that degrade or lose grounding.
  • On benchmarks where 2D image-plane shortcuts are uninformative (3DSRBench multi-object relations), explicit 3D evidence gives the largest gains, e.g., about 12% over HiSpatial on multi-object accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If the small-object perception bottleneck is removed (e.g., with higher-resolution depth or multi-frame geometry), a DiSR-style pipeline should surpass end-to-end MLLMs on CV-Bench-3D as well, since the reasoning stage is already at 100% with perfect evidence; this can be tested by swapping in stronger perception modules.
  • Editorial inference: The paper's 2D-shortcut analysis implies that some end-to-end MLLM scores on CV-Bench-3D partly reflect image-plane heuristics; a cleaner evaluation of spatial ability would use DiSR's conflict subsets, where the shortcut misleads, as a diagnostic.
  • Editorial inference: The same 'serialize geometry as text' recipe may extend beyond Q&A to robotics or AR, where a planner could consume object centers and axes as text and reason about reachability or occlusion without learning a 3D encoder.
  • Editorial inference: The question parser and grounding module are still an MLLM; replacing them with a lightweight detector/parser could make the whole pipeline fully modular and further cut cost, while keeping the evidence interface unchanged.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DiSR, a two-stage framework for spatial reasoning that explicitly separates 3D perception from reasoning. In the first stage, off-the-shelf perception models (SAM, Metric3D, WildCamera, PerspectiveFields, OrientAnything) construct a structured 3D evidence representation E(I) containing object centers, extents, sizes, and front/left orientation axes in a shared gravity-aligned coordinate frame. In the second stage, a frozen LLM equipped with LoRA is fine-tuned to answer questions from a serialized textual version of this evidence, without seeing the image. Training uses 0.33M generated question-answer samples from Open Images, requiring about 59 GPU-hours on a single RTX 4080 SUPER. The paper reports state-of-the-art results on 3DSRBench (67.62%) and SPAR-Bench (46.33%), competitive results on CV-Bench-3D, preserved general reasoning performance on five benchmarks, and a diagnostic analysis showing that errors can be attributed to question parsing, object grounding, or 3D evidence construction.

Significance. If the empirical claims hold, the paper demonstrates a genuinely different paradigm from end-to-end spatial VQA training: a frozen perception stack plus a small LoRA-tuned LLM can outperform models trained on billions of spatial QA pairs. The strengths of the paper are real: the controlled diagnostic in Table 5 separates parsing, grounding, and evidence errors; the 2D-shortcut analysis in Appendix B is careful and informative; the training cost is remarkably low; and the cross-backbone result in Table 7 supports modularity. The main risk is that the central claim depends entirely on the accuracy of the structured 3D evidence, and the paper measures that accuracy only on CV-Bench-3D, where it is shown to fail for small objects. Without a benchmark-specific evidence-quality analysis for 3DSRBench and SPAR-Bench, the SOTA claims are not yet fully supported. The comparison to a 3B baseline and the absence of error bars further weaken the quantitative conclusions.

major comments (4)
  1. [Method, Eq. (5) and Appendix Table A4] The entire performance ceiling of DiSR is set by the accuracy of the structured evidence E(I) in Eq. (5). Table 5 shows that when ground-truth 3D evidence is supplied, the reasoner reaches 100% on CV-Bench-3D, and Table A4 shows that the automatic evidence-construction error rate is 15.00% for the smallest-object group versus 1.67% for the largest. However, the only direct measurement of E(I) quality is on CV-Bench-3D Depth. For 3DSRBench and SPAR-Bench, where the paper claims new state-of-the-art results, no analogous evidence-accuracy analysis is reported. Since object sizes, scene types, and relation categories differ across these benchmarks, the headline gains could in principle arise from reasoning over inaccurate evidence plus text-level regularities rather than from accurate geometric evidence. I request a benchmark-specific evidence-quality evaluation (e.g., size-stratified construction error rates or systematic failure-case inspection) for 3DSRBench and SPAR-Bench, or an explicit transfer argument from the CV-Bench-3D measurement.
  2. [Experiments, Tables 1 and 2] The comparisons supporting the SOTA claims are partially confounded by base-model size. Table 1 compares DiSR-8B-LoRA, built on Qwen3-VL-8B-Instruct, to HiSpatial-3B, a 3B model, while Table 2 compares it to Qwen3-VL-8B-Instruct on SPAR-Bench. The reported margins (3.77% on 3DSRBench, 1.70% on SPAR-Bench) come from a single run with no error bars, repeated seeds, or significance tests. The cross-backbone result in Table 7 helps for 3DSRBench because DiSR-Qwen2.5-7B reaches 65.52%, still above HiSpatial-3B, but no equivalent result is given for SPAR-Bench. To make the SOTA claim convincing, please add multi-seed statistics and, where possible, a matched-size spatial-specialist baseline.
  3. [Experiments, Table 2] On SPAR-Bench, the overall average of 46.33% is driven by metric estimation (49.37% MRA versus Qwen3-VL-8B-Instruct's 33.55%), while on relational selection DiSR scores only a 40.25% average versus 66.81% for Qwen3-VL-8B-Instruct, including 26.37% on ObjRel-OO. The paper acknowledges the training-data coverage gap in the text, but the abstract and introduction still present this as a state-of-the-art spatial reasoning result without this important qualification. The claim should be narrowed, and the relational-selection deficiency should be analyzed beyond the single-sentence explanation, for example by reporting the distribution of training questions over the SPAR-Bench task types.
  4. [Method and Implementation Details] The role of Qwen3-VL-8B-Instruct as question parser and object grounder is underspecified. Equations (2) and (4) describe a planner and a grounding module, and the Implementation Details say 'for object grounding in deriving 3D evidence, we use both the vision encoder and LLM backbone, while only the LLM is adopted for reasoning and question parsing.' It is not stated whether the grounding and parsing at inference use the frozen base model without LoRA or the LoRA-adapted model, nor whether the reported 59 GPU-hours include any adaptation of the grounding path. This matters for the disentanglement claim because part of the 'perception' work is performed by a large vision-language model rather than by the small specialized expert models. Please clarify the exact parameter states and inference-time components.
minor comments (5)
  1. [Experiments, Table 3] The HiSpatial-3B (RGB) row reports only the average (95.58) without Depth and Distance scores; please provide the full breakdown for comparability.
  2. [Experiments, Table 6] The entry 'HiSpatial-3B (RGB-XYZ) fails' is informal; it should be replaced with a precise statement such as 'does not produce valid bounding boxes under the evaluation protocol' or with the measured failure rate.
  3. [Method, Eq. (5)] The notation e_i = (c_i, d_i, sigma_i, rho_i, f_i, l_i) introduces sigma_i and rho_i as 'object-level size descriptors' without defining them; please state explicitly whether they are volume, surface area, axis-aligned bounding-box dimensions, or something else.
  4. [Data Collection and Training] The sentence 'Since images are not required during training, we also construct synthetic data that approximates the distribution of the collected dataset' is vague; please specify how the synthetic data are generated, how many samples are synthetic, and how they are combined with the Open Images-derived data.
  5. [Appendix C, Table A5] The two-stage training procedure (Stage 1 with 112,157 instances and Stage 2 with 216,000 instances) is not explained in the main text; please describe what each stage learns and why the decomposition is necessary.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims are external benchmark accuracies, and its diagnostic ablations are presented as error attribution rather than as predictions derived from fitted inputs.

full rationale

DiSR's derivation chain is not circular. The structured 3D evidence E(I) in Eq. (5) is constructed by frozen off-the-shelf perception models, and the LLM is fine-tuned to reason over that evidence. The headline claims—state-of-the-art results on 3DSRBench and SPAR-Bench—are external benchmark numbers, not quantities that equal their inputs by construction. No parameter is fitted to a benchmark subset and then reported as a prediction of that same subset. The ground-truth ablations in Table 5 and Table A4 (e.g., 'DiSR-8B-LoRA w/ gt of required 3D evidence 100.0 100.0') are explicitly diagnostic: they attribute errors to perception rather than claiming the 100% result as the system's actual performance. The paper's own size-stratified analysis in Table A4 acknowledges that small-object 3D evidence construction fails (15.00% error rate), which is an honest limitation rather than a circular justification. No load-bearing argument reduces to a self-citation: the cited prior works are baselines and building blocks, not author-supplied uniqueness theorems or ansatz justifications. The use of Qwen3-VL-8B-Instruct as both a component and a baseline is a design choice, not a circular step, because the reported gains are measured against that same model on external benchmarks. The weakest assumption—that frozen perception models produce accurate metric evidence on the SOTA benchmarks—is an unverified empirical premise, not a circularity. Missing measurements on 3DSRBench/SPAR-Bench perception quality are a correctness-risk concern, not evidence that the derivation reduces to its own inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reliability of pretrained perception models and on benchmark validity, not on derived physical constants or a mathematical derivation. No new physical entity is postulated. The structured textual evidence is a serialization format, not a new entity. Training hyperparameters are ordinary design choices. The main uncharged dependency is perception accuracy for small objects.

free parameters (4)
  • LoRA rank = 128
    Chosen PEFT hyperparameter listed in Appendix C; affects adapter capacity but is not fitted to a scientific target.
  • LoRA scaling factor = 256
    Chosen PEFT hyperparameter; standard implementation choice, not derived from data.
  • Learning rate = 2e-5
    Chosen optimization hyperparameter used in both training stages.
  • Stage-2 training instance count = 216,000
    Training data volume is a design choice; the authors state it does not cover the SPAR-Bench relational selection distribution.
assumptions (4)
  • domain assumption Off-the-shelf perception models (SAM, Metric3D, WildCamera, Perspective Fields, OrientAnything) produce sufficiently accurate metric depth, camera pose, masks, and orientations in a shared gravity-aligned coordinate frame.
    Invoked in Method, Structured 3D Evidence Construction, Eq. (5). Appendix Table A4 shows this fails for small objects, with 15.00% evidence construction error versus 1.67% for large objects.
  • domain assumption The serialized textual evidence E_T preserves all geometric information needed to answer the spatial questions.
    The reasoning LLM sees only E_T through Eqs. (6) and (7); if the text drops or distorts information, no LLM reasoning ability can recover it.
  • domain assumption The evaluation benchmarks measure spatial reasoning rather than solvable 2D image-plane shortcuts.
    The authors themselves show in Appendix B that CV-Bench-3D Depth and Distance contain strong 2D heuristics, so benchmark validity is not uniform across the three datasets.
  • ad hoc to paper The 0.33M training samples cover the distribution of test question types.
    The authors state that SPAR-Bench relational selection underperforms because the generated training data cannot cover that distribution, making this an explicit and acknowledged assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangling 3D Modeling from Spatial Reasoning." pith.science (2026). https://pith.science/paper/O3RDJMH6

@misc{pith2026260805242,
  author       = {Pith},
  title        = {Pith review of: Disentangling 3D Modeling from Spatial Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O3RDJMH6}},
  note         = {Machine review of arXiv:2608.05242}
}
read the original abstract

In this work, we explore an alternative paradigm for spatial reasoning by explicitly disentangling 3D perception from reasoning, rather than jointly acquiring implicit 3D perception and reasoning through large-scale training. Our key observation is that modern perception models excel at estimating continuous 3D geometry, whereas large language models (LLMs) are particularly effective at compositional and symbolic reasoning. Motivated by these complementary strengths, we propose the Disentangled Spatial Reasoner (DiSR), a simple yet effective framework that reconstructs the physical world into structured 3D evidence using off-the-shelf expert perception models and fine-tunes an LLM with LoRA to perform reasoning solely over this explicit geometric evidence. Without large-scale 3D VQA training or complex tool-use policies, DiSR achieves competitive performance on popular spatial reasoning benchmarks. Beyond its strong performance, DiSR offers improved interpretability, modularity, and computational efficiency, demonstrating that explicit separation of perception and reasoning is a scalable and effective alternative paradigm to end-to-end modeling for spatial intelligence.

Figures

Figures reproduced from arXiv: 2608.05242 by the authors.

Figure 1
Figure 1. Comparison of spatial reasoning paradigms. (a) Methods jointly acquire implicit geometric perception and spatial [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of DiSR. Object Grounding. Given the object plan P, we take Qwen3-VL-8B-Instruct as the grounding module to local￾ize each queried object reference mi in the image and iden￾tify its corresponding region ri . The grounded regions are represented as: R = {(qi : ri)} N i=1, (4) where the slot identifier qi preserves the correspondence be￾tween the language query and the visual region. It bridges linguistic ref… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 9 canonical work pages

  1. [1]

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

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

  2. [2]

    Advances in Neural Information Processing Systems , volume=

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

  3. [3]

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

    Spatialllm: A compound 3d-informed design towards spatially-intelligent large multimodal models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Spatialreasoner: Towards explicit and generalizable 3d spatial reasoning , author=. Advances in Neural Information Processing Systems , volume=

    work page

  5. [5]

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

    HiSpatial: Taming Hierarchical 3D Spatial Understanding in Vision-Language Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    3d-llm: Injecting the 3d world into large language models , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

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

    3d-vista: Pre-trained transformer for 3d vision and text alignment , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  8. [8]

    European Conference on Computer Vision , pages=

    Pointllm: Empowering large language models to understand point clouds , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

Show all 37 references
  1. [9]

    arXiv preprint arXiv:2311.12871 , year=

    An embodied generalist agent in 3d world , author=. arXiv preprint arXiv:2311.12871 , year=

    arXiv

  2. [10]

    Advances in neural information processing systems , volume=

    Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors , author=. Advances in neural information processing systems , volume=

  3. [11]

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

    Spacetools: Tool-augmented spatial reasoning via double interactive rl , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  4. [12]

    arXiv preprint arXiv:2510.07181 , year=

    TIGeR: Tool-Integrated Geometric Reasoning in Vision-Language Models for Robotics , author=. arXiv preprint arXiv:2510.07181 , year=

  5. [13]

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

    Geometrically-constrained agent for spatial reasoning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  6. [14]

    arXiv preprint arXiv:2508.02324 , year=

    Qwen-image technical report , author=. arXiv preprint arXiv:2508.02324 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    arXiv

  8. [16]

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

    Segment anything , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  9. [17]

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

    Metric3d: Towards zero-shot metric 3d prediction from a single image , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  10. [18]

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

    Perspective fields for single image camera calibration , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [19]

    arXiv preprint arXiv:2412.18605 , year=

    Orient anything: Learning robust object orientation estimation from rendering 3d models , author=. arXiv preprint arXiv:2412.18605 , year=

    arXiv

  12. [20]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  13. [21]

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

    3dsrbench: A comprehensive 3d spatial reasoning benchmark , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  14. [22]

    Advances in Neural Information Processing Systems , volume=

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms , author=. Advances in Neural Information Processing Systems , volume=

  15. [23]

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

    Gqa: A new dataset for real-world visual reasoning and compositional question answering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [24]

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

    Improved baselines with visual instruction tuning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [25]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Spatialbot: Precise spatial understanding with vision language models , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  18. [26]

    International Conference on Learning Representations , volume=

    Grounding multimodal large language models to the world , author=. International Conference on Learning Representations , volume=

  19. [27]

    Lin, Ji and Yin, Hongxu and Ping, Wei and Molchanov, Pavlo and Shoeybi, Mohammad and Han, Song , booktitle=

  20. [28]

    Advances in Neural Information Processing Systems , volume=

    Visual Instruction Tuning , author=. Advances in Neural Information Processing Systems , volume=

  21. [29]

    Advances in Neural Information Processing Systems , volume=

    From flatland to space: Teaching vision-language models to perceive and reason in 3d , author=. Advances in Neural Information Processing Systems , volume=

  22. [30]

    Advances in Neural Information Processing Systems , volume=

    Tame a wild camera: In-the-wild monocular camera calibration , author=. Advances in Neural Information Processing Systems , volume=

  23. [31]

    European conference on computer vision , pages=

    Mmbench: Is your multi-modal model an all-around player? , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  24. [32]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Evaluating object hallucination in large vision-language models , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  25. [33]

    arXiv preprint arXiv:2307.16125 , year=

    Seed-bench: Benchmarking multimodal llms with generative comprehension , author=. arXiv preprint arXiv:2307.16125 , year=

    arXiv

  26. [34]

    2024 , howpublished =

    2024

  27. [35]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  28. [36]

    arXiv preprint arXiv:1811.00982 , year=

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale , author=. arXiv preprint arXiv:1811.00982 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.