Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Disentangling 3D Modeling from Spatial Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read DiSR shows that strong spatial reasoning can be achieved by explicitly separating 3D perception from reasoning, using frozen experts to build per-object 3D evidence and a LoRA-tuned LLM to reason over that evidence alone.

desk verdict DiSR's disentanglement result is real and worth engaging, but the headline SOTA claim is confounded by model scale and the SPAR-Bench* number is partly fitted. read the letter →

arxiv 2608.05242 v2 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 spatialreasoning3DperceptionstructuredevidencedisentanglementLoRAVQAinterpretabilitydataefficiency
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 sets out to prove that spatial reasoning does not have to be learned jointly with 3D perception. It constructs explicit per-object 3D evidence—center, extent, size descriptors, and front/left axes in a shared gravity-aligned coordinate frame—using frozen expert perception models, serializes that evidence as text, and fine-tunes a large language model with LoRA to answer spatial questions from the text alone. On 3DSRBench and SPAR-Bench the authors report new state-of-the-art overall accuracy, beating the previous best methods by 3.77% and 1.70% with 0.33M training samples and 59 GPU hours, a fraction of the compute used by the strongest joint-training baseline. If true, this establishes that a modular perception/reasoning split is a data-efficient alternative to billion-scale 3D VQA training, and it makes spatial errors inspectable because evidence and reasoning are separable.

What carries the argument

The load-bearing object is the structured 3D evidence record $e_i = (c_i, d_i, \sigma_i, \rho_i, f_i, l_i)$: object center, 3D extent, two size descriptors, and front/left axes, all in one gravity-aligned coordinate system. This textual interface converts a spatial question into an algebraic comparison problem, so the reasoning model can be trained with standard autoregressive language modeling over evidence text. The LoRA-tuned Qwen3-VL-8B backbone, together with the question parser that produces persistent object slots, is the second mechanism that carries the argument.

What would settle it

Take two images of scenes with identical per-object centers, extents, size descriptors, and front/left axes but different containment, support, or occlusion relations, and ask DiSR a question whose correct answer differs; if it answers correctly, it is relying on information absent from its stated evidence, and if it fails, the evidence interface is provably incomplete for that class of spatial questions.

Watch

Extended reading notes

Core claim

The central claim is that explicitly disentangling 3D perception from spatial reasoning is an effective and data-efficient paradigm for spatial intelligence. DiSR reconstructs the physical world into structured 3D evidence $E(I)$, with one tuple per object composed of center, 3D extent, size descriptors, and front/left axes in a shared gravity-aligned frame; the frozen experts are SAM for masks, Metric3D for metric depth, WildCamera and PerspectiveFields for camera geometry, and OrientAnything for object orientation. The LLM never sees pixels: it reasons over a serialized form of $E(I)$ after LoRA fine-tuning on 0.33M samples drawn from Open Images, with parsing and reasoning trained as two supervised tasks. This configuration reaches 67.62% overall on 3DSRBench and 46.33% on SPAR-Bench, and controlled diagnostics show that when the required 3D evidence is replaced by ground truth, CV-Bench-3D Depth and Distance accuracy reaches 100%, which the authors read as evidence that the remaining errors are perception-side rather than reasoning-side.

Load-bearing premise

The argument holds only if every spatial relation the benchmarks ask about can be recovered from per-object centers, extents, size descriptors, and front/left axes in a shared gravity-aligned frame, and if the frozen depth, camera, and orientation experts are accurate enough to supply those values—a premise the paper's own appendix shows is violated for small objects.

Editorial extensions

If this is right

  • Perception upgrades transfer without retraining: replacing a backbone LLM or the grounding module improves overall spatial accuracy while the pipeline stays the same.
  • Evidence-level diagnostics isolate failures: with ground-truth grounding, CV-Bench-3D accuracy rises by 1.34% and 1.66% on Depth and Distance tasks, while ground-truth evidence reaches 100%, pinning residual error on perception rather than reasoning.
  • The method preserves general visual reasoning: DiSR matches its base model across MMBench, GQA, POPE, SEED, and RealWorldQA, unlike spatial-specialist baselines that trade general ability for spatial ability.
  • Spatial skills no longer require billion-scale 3D VQA data: 0.33M LoRA samples on one RTX 4080 Super reproduce and exceed results obtained with 2B spatial QA pairs and 1,536 H100 GPU hours.

Reading between the lines

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

  • Beyond the paper: because the evidence tuple is the complete input to the reasoner, the same architecture can answer spatial queries that are not in the current benchmarks—proximity thresholds, path relations, or volumetric comparisons—by adding fields to $e_i$ and retraining only the small LoRA adapter.
  • Beyond the paper: the appendix's 2D-shortcut analysis suggests CV-Bench-3D's high joint-model scores partly reflect image-plane heuristics; a benchmark rebuilt around 2D-conflict cases would likely widen DiSR's measured advantage, though this is our reading, not the paper's conclusion.
  • Beyond the paper: the scene-as-text interface points toward a practical route for embodied agents, where a robot could query a frozen perception stack for object-centric evidence and reason over it in natural language, potentially lowering data requirements for manipulation and navigation policies.
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 / 6 minor

Summary. The paper proposes DiSR, a two-stage spatial reasoning framework that explicitly separates 3D perception from reasoning. A frozen stack of expert perception models (SAM, Metric3D, WildCamera, PerspectiveFields, OrientAnything) reconstructs each queried object into structured 3D evidence: object center, 3D extent, size descriptors, and front/left axes in a shared gravity-aligned coordinate frame (Eq. 5). A Qwen3-VL-8B-Instruct backbone is used for question parsing and object grounding, and an LLM is fine-tuned with LoRA on 0.33M generated samples to answer questions from the serialized evidence alone, with no image input at reasoning time. The paper reports 67.62% on 3DSRBench (vs. HiSpatial-3B at 63.85%), 46.33% on SPAR-Bench (vs. Qwen3-VL-8B at 44.63%), and a further SPAR-Bench variant DiSR-8B-LoRA* at 57.56% after a 2,816-example adaptation set. On CV-Bench-3D, DiSR reaches 92.25%, comparable to its base model but below HiSpatial-3B. Controlled diagnostics show that replacing predicted object grounding with ground truth gives modest gains, while replacing the 3D evidence with ground truth yields near-100% accuracy on CV-Bench-3D Depth and Distance. The paper also reports preserved general reasoning performance and a large reduction in training cost relative to HiSpatial (59 RTX 4080 SUPER GPU-hours vs. 1,536 H100 GPU-hours).

Significance. If the main results hold, the paper is a valuable existence proof that a disentangled perception-plus-reasoning pipeline can match or exceed end-to-end spatial VQA training at far lower cost. The internal diagnostic with ground-truth evidence reaching near-100% accuracy is a strong, causal attribution of errors to perception rather than reasoning, and the Appendix B analysis of 2D shortcuts in CV-Bench-3D is a useful methodological contribution. The interpretability and modularity claims are also credible and are supported by the cross-backbone experiment in Table 7 and the grounding analysis in Table 6. However, the significance of the headline SOTA claims is currently weakened by two issues: the main 3DSRBench comparison is between an 8B-base system and a 3B-base specialist, and the SPAR-Bench* result is produced by a benchmark-targeted adaptation set rather than a held-out evaluation. These issues are fixable with additional experiments or with qualified claims, but they must be addressed before the paper's central contribution can be assessed at face value.

major comments (4)
  1. [Table 1 and Table 7] The headline 3DSRBench comparison is not scale-matched: DiSR-8B-LoRA uses an 8B-parameter base model (Qwen3-VL-8B-Instruct) while the previous best specialist, HiSpatial-3B, uses a 3B base. The within-backbone gain from 53.70% (Qwen3-VL-8B-Instruct) to 67.62% (DiSR-8B-LoRA) is a valid and useful control, but it does not establish a new state of the art against a scale-matched specialist. Table 7 shows DiSR-Qwen2.5 with a 7B base at 65.52%, which still uses a larger backbone than HiSpatial-3B. The paper should either add a HiSpatial or equivalent specialist at 7B-8B scale, or explicitly restrict the SOTA claim to comparisons with the reported baselines and discuss the scale confound.
  2. [Table 2 and Appendix D] The starred SPAR-Bench result, DiSR-8B-LoRA* at 57.56%, is presented in the body text as surpassing the previous best model by 12.93%, but Appendix D states that this model is produced by a post-freeze 2,816-example adaptation set in which half of the examples cover the two previously unseen SPAR-Bench task types. This is a benchmark-targeted fit, not a held-out evaluation, and it should not be used as evidence for a new state of the art. The unstarred 46.33% result is a legitimate held-out number, but it is only 1.70 points above the previous best and comes from a single seed. Please separate these two claims clearly and, if the starred result is kept, label it as an adaptation diagnostic rather than a SOTA result.
  3. [Method, Eq. (5) and Appendix B, Table A4] The structured evidence representation contains only object centers, 3D extents, size descriptors, and front/left axes. Spatial relations involving shape, support, containment, or occlusion cannot be expressed in this representation, and the paper's own size-stratified analysis on CV-Bench-3D Depth reports a 15.00% evidence-construction error rate on the smallest-object quintile, with low-quality images identified as a cause. This does not invalidate the benchmark results, but it substantially narrows the breadth of the paradigm claim stated in the Abstract and Introduction, which speaks of reconstructing the physical world and achieving spatial intelligence. The scope of the claims should be narrowed, or additional relation types and evidence modalities should be evaluated.
  4. [Tables 1 and 2, Appendix C] All reported accuracies come from a single random seed (seed 42, Appendix C) with no confidence intervals or repeated runs. The SPAR-Bench margin over the previous best in Table 2 is only 1.70 points, which is within the range of typical run-to-run variation for LoRA fine-tuning on a single GPU. Please provide multiple seeds or bootstrap confidence intervals for at least the headline 3DSRBench and SPAR-Bench numbers, so the reader can assess whether the differences are statistically meaningful.
minor comments (6)
  1. [Experimental Results, SPAR-Bench paragraph] The text contains a typo: 'Relatinal Selection' should be 'Relational Selection'.
  2. [Table 2] The table combines MRA (mean relative accuracy) and accuracy in one 'Avg.' column, but the weighting of the two metric types is not defined. Please clarify how the overall average is computed.
  3. [Abstract vs. Introduction] The Abstract says DiSR 'achieves competitive performance' while the Introduction states it 'achieves new state-of-the-art performance on 3DSRBench and SPAR-Bench.' These formulations are inconsistent and should be harmonized, especially given the scale and adaptation issues noted above.
  4. [Method, Eq. (6)] The evidence serialization function phi is never concretely specified. A worked example of the exact textual format passed to the LLM would substantially improve reproducibility.
  5. [Appendix D vs. main text] The main text says 'additional 2000 data' for relational selection, while Appendix D describes a 2,816-example adaptation set with half new task types and half replay data. These numbers and the composition of the adaptation set should be reconciled.
  6. [Reproducibility details] No code, data, or trained checkpoints are released, despite the central claims being purely empirical. If release is not possible, the paper should at least provide the full training-data construction script or a more detailed template of the serialized evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core benchmark results are held-out evaluations and the only task-adapted result (DiSR-8B-LoRA*) is explicitly disclosed and non-central.

full rationale

The paper's derivation chain is self-contained: a frozen perception stack constructs structured 3D evidence E(I) via Eq. 5, an LLM with LoRA reasons over the serialized evidence E_T via Eq. 7, and the LLM is trained on QA pairs generated from Open Images. The reported 3DSRBench and SPAR-Bench scores for DiSR-8B-LoRA are held-out evaluations rather than quantities fitted from the benchmarks themselves. The SPAR-Bench* result (Table 2, Appendix D) is trained on additional data covering the evaluation task types, but it is clearly marked with an asterisk, described as a post-freeze diagnostic, and is not the basis of the abstract's 1.70% margin; it is a disclosed benchmark adaptation, not a disguised prediction. The oracle experiment in Table 5 (100% with ground-truth 3D evidence) is an intentionally constructed diagnostic upper bound used for error attribution, not a claimed predictive result, so it does not make the central claim circular. There are no load-bearing self-citations: the citations to SpatialReasoner (Ma et al. 2026) concern data-construction protocol and are by different authors. The paper's own appendix acknowledges perception limitations (e.g., 15.00% evidence-construction error on the smallest-object quintile of CV-Bench-3D Depth), which is a correctness/robustness concern, not evidence of circularity. The scale mismatch between DiSR-8B and HiSpatial-3B is an interpretation confound, not a circular-derivation issue. No equation or parameter in the paper reduces by construction to its own input, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities or fitted physical constants, so the ledger records experimental design choices and domain assumptions. The most consequential choices are the evidence schema (Eq. 5), the frozen expert stack, the un-released 0.33M training data, and the benchmark-targeted 2,816-example adapter behind the SPAR-Bench* number.

free parameters (3)
  • LoRA hyperparameters (rank 128, scaling 256, dropout 0.05, LR 2e-5) = rank=128, alpha=256, dropout=0.05
    Chosen by hand without sensitivity analysis; they influence all reported results but are not fitted to benchmark answers. Listed as free parameters because the paper provides no ablation showing robustness.
  • Training dataset size and split (112,157 parsing + 216,000 reasoning samples) = 0.33M total
    The efficiency claim rests on this quantity; the exact construction follows Ma et al. 2026 and is not fully specified, so the effective data distribution is an uncontrolled choice.
  • SPAR-Bench adaptation set (2,816 examples) = 2,816 (text also says 'additional 2000 data')
    The DiSR-8B-LoRA* result (57.56) is obtained after training on examples covering two SPAR-Bench task types. This is a hand-picked, benchmark-targeted dataset; its composition and overlap with evaluation are not described. The paper gives inconsistent counts (2000 vs 2816).
assumptions (6)
  • domain assumption Specialized perception models provide accurate continuous 3D geometry (depth, pose, orientation) for images in the evaluation benchmarks.
    Invoked in Method/Structured 3D Evidence Construction; Appendix B shows this fails for small/low-resolution objects, so the assumption is load-bearing and partially contradicted.
  • domain assumption All spatial relations queried by the benchmarks can be derived from per-object center, extent, size, front axis and left axis in a shared gravity-aligned coordinate frame (Eq. 5).
    The serialized evidence E(I) contains only these attributes; relations involving support, containment, occlusion, or shape-dependent geometry are not represented. This is the interface that makes the LLM reasoning possible.
  • domain assumption The question parser and object grounding module (Qwen3-VL-8B-Instruct) reliably identifies every object referenced by the question.
    Used in Eq. 2-4 and grounding pipeline; evaluated only on a COCO unique-instance subset (Table 6), not on the three spatial benchmarks.
  • standard math Depth back-projection and rigid alignment (solving for metric 3D centers and orientations) are standard linear algebra and are assumed correct given accurate depth, intrinsics, and gravity orientation.
    Invoked in 'Reconstructing the Physical Scene to Derive Evidence'; the math itself is routine, but its output validity depends on the perception inputs.
  • domain assumption Camera intrinsics and gravity-aligned orientation from WildCamera/PerspectiveFields are accurate enough for back-projection into a shared coordinate frame.
    Back-projection and coordinate transformation in Method; no error analysis of camera calibration on benchmark images.
  • ad hoc to paper Training data generated from Open Images following Ma et al. 2026, plus synthetic data, approximates the evaluation distribution well enough for LoRA fine-tuning.
    The paper states 'Following previous work (Ma et al. 2026)...' and 'we also construct synthetic data that approximates the distribution of the collected dataset' without releasing prompts or generation code; this data distribution is central to the SPAR-Bench relational-selection gap.

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 ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 7 linked inside Pith

  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=

  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=

  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=

  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=

  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 10, 2026 · model on record in the stance chip above.