Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Vision language models are unreliable at trivial spatial cognition

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Tested vision-language models answer logically equivalent left/right questions inconsistently, with accuracy ranging from 12% to 100% depending on wording, object order, and whether the relation holds.

desk verdict A useful benchmark with a solid core finding—VLMs are unstable across logically equivalent spatial prompts—but the 'unreliable' claim is stronger than the evidence, and the paper needs a human baseline and code/data release. read the letter →

arxiv 2504.16061 v1 pith:ZXK52PLZ submitted 2025-04-22 cs.CV cs.AI

Sangeet Khemlani , Tyler Tran , Nathaniel Gyory , Anthony M. Harrison , Wallace E. Lawson , Ravenna Thielstrom , Hunter Thompson , Taaren Singh
show 1 more author
J. Gregory Trafton
This is my paper · ORCID
classification cs.CVcs.AI
keywords vision-languagemodelsspatialcognitionleft/rightrelationspromptreliabilitysyntheticbenchmarkrelationalreasoningimage-textevaluationTableTest
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

The paper sets out to establish that state-of-the-art vision-language models are unreliable at trivial spatial cognition: recognizing which of two objects is left or right of the other in an uncluttered image. Using a new synthetic dataset called TableTest and eight families of simple prompts, the authors show that accuracy swings with variations that should be meaningless, including left-versus-right wording, the order in which objects are named, whether the asked relation is true or false, and the presence of an irrelevant relation. Each tested model falls below a 90% adequacy threshold on at least two prompt families, and no model is consistent across them. The point matters because a model that answers one phrasing correctly but fails an equivalent one cannot be trusted in applications that depend on stable spatial judgments.

What carries the argument

TableTest is the central instrument: a synthetic benchmark of 4,032 two-object images, plus a larger set of three-object scenes, built from 64 Objaverse objects rendered in Blender on a uniform table with object centers placed 200 pixels from the table's center. The paper pairs each image with eight prompt families whose variations, including left/right wording, object order, truth and falsity, and the insertion of an irrelevant relation, are meant to be logically equivalent. The benchmark's work is to hold the visual scene constant while varying only the surface form of the query, so any accuracy shift isolates a model's instability rather than a confound in the image.

What would settle it

Give neurotypical adults the exact TableTest images with all eight prompt families and their variations; if their accuracy moves with wording, object order, or truth value by more than a small margin, the assumed human ceiling of 1.0 fails and the models' variation cannot be called unreliable by the paper's own standard.

Watch

Extended reading notes

Core claim

Across three readily available vision-language models, the paper finds that accuracy on trivial left/right recognition depends on nearly every surface variable in the prompt. Overall accuracy spans 0.12 to 1.00 across eight prompt families; every model falls below 90% on at least two multimodal prompts, and factors the paper treats as logically irrelevant, namely whether the query says left or right, whether objects are named in image order, whether the queried relation is true or false, and whether an unrelated relation like "inside of" is added, reliably change accuracy. The models also err in different directions from one another, so the variation is not a single stable bias. The paper's conclusion is that these models do not construct coherent spatial representations even for relations a human resolves instantly.

Load-bearing premise

The central conclusion rests on the premise that logically equivalent phrasings of a left/right question should produce identical accuracy, together with the paper's assumption, which is not backed by human data, that neurotypical adults answer all these probes at ceiling.

Editorial extensions

If this is right

  • A high score on a single spatial prompt does not establish spatial competence: the same model can be near ceiling on one logically equivalent phrasing and near chance on another.
  • Evaluation protocols for VLMs should report reliability across equivalent prompt variations, not only accuracy on a fixed template, because surface format measurably changes results.
  • Fine-tuning on a single prompt is unlikely to generalize to other phrasings unless the model forms a coherent spatial representation; errors reappear as soon as the wording changes.
  • Other spatial relations, such as above, below, inside, and between, are likely to exhibit the same instability, since training captions rarely enumerate all relations in an image.

Reading between the lines

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

  • We infer that reliability across equivalent prompt variants is a more informative metric than peak accuracy for choosing a model in practice: a model that is consistently adequate on every phrasing may be preferable to one that is very accurate on a single phrasing and near chance on others.
  • We infer that the pattern of left/right and object-order biases could be formalized as a response-bias model, letting engineers predict on which unseen prompt variations a given model will fail before running a full benchmark.
  • We infer that systems chaining multiple VLM calls, such as an image caption followed by a relational query, may compound or cancel these prompt-format errors, so reliability should also be measured at the pipeline level.
  • We infer that a human baseline on the same TableTest prompts is the natural way to set the "trivial" threshold; until such data exist, the paper's ceiling of 1.0 should be treated as a stipulated standard rather than a measured one.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces TableTest, a synthetic benchmark of 3D-rendered table scenes with known ground-truth left/right relations, and uses it to probe three VLMs (BLIP, Llama 3.2, Idefics2) across eight prompt formats, each with surface variations that the authors treat as logically equivalent. In multimodal evaluations, accuracies range from .12 to 1.00 across prompts; the authors report stable main effects of left vs. right wording, object order, true vs. false statements, and the presence of an irrelevant relation, and argue that because these variations should not affect a system with a coherent spatial representation, the models are unreliable at trivial spatial cognition. Text-only control conditions generally show higher accuracy, and open-form prompts (6-8) are scored with a permissive custom metric described in Appendix A.

Significance. If the empirical claims hold, the paper makes a useful contribution: a synthetic benchmark with exact ground truth, systematic and reproducible prompt variants, and nonparametric object-level analyses showing that VLM performance is sensitive to surface factors. The cross-model consistency of the main effects is a strength, as is the deliberate focus on trivial tasks with simple stimuli. However, the central interpretation depends on an unmeasured human ceiling and on a scoring procedure partially derived from the tested models' own outputs, so the significance is conditional on those issues being resolved. The work would be more valuable if the benchmark and scoring code were released and if human norming data were collected.

major comments (3)
  1. [§1.1, Figs 2-3, §3.3.9] The paper's central inference, from 'accuracy varies across prompt formulations' to 'models lack stable spatial representations,' depends on the assertion that neurotypical adults are at ceiling on all eight prompts and that every manipulated variable is logically irrelevant. These claims are asserted in Section 1.1 and in the figure captions ('Humanlike performance anticipated to be at ceiling'), but no human data on the exact stimuli or prompts are reported. If adults show even a small true-vs-false asymmetry or a left/right advantage on these materials, the observed VLM differences would not demonstrate the absence of stable spatial representations; they would show that the models track difficulty or response biases that humans also share. I recommend adding a human norming study on the same images and prompts, or a clearly applicable published baseline, and tempering the concluding claim accordingly.
  2. [§3.1] There is a factual slip in the Prompt 1 example. The text says that for Figure 1b (blender to the right of a cupcake) a VLM should respond false to the statements 'the cupcake is to the left of the blender' and 'the blender is to the right of the cupcake'; both statements are true for that image. Please correct the example and, more importantly, verify the ground-truth relation labels in the Prompt 1 evaluation code, because the reported true-vs-false accuracy asymmetry is central to the paper's conclusions.
  3. [Appendix A] The open-form accuracy metric is partly defined by the evaluated models' own responses. Acceptable object labels are entered into the lookup table whenever a VLM 'routinely generated' them five or more times, and the metric explicitly counts a response as correct unless it fails to recognize the relation. This creates a circularity that can inflate accuracy on Prompts 6-8 and makes those prompts not directly comparable with the closed-form prompts. The acceptable-alternative list should be constructed independently of the tested models (e.g., from human annotation, WordNet synonyms, or the Objaverse labels), and the full-sentence exception in Prompt 6 scoring should be reported as a separate analysis.
minor comments (4)
  1. [Table 2, Figs 2-3] The paper reports aggregate proportions without confidence intervals or effect sizes, and the unit of analysis for the Wilcoxon tests is implicit. Please report bootstrapped confidence intervals or variance across objects, and state explicitly whether the tests are run on per-object accuracy scores.
  2. [§3.3.6 and §1] In §3.3.6, 'see 2f' should be 'see Figure 2f'; in Section 1, 'no affect on performance' should be 'no effect on performance.'
  3. [References] The paper refers to Llama 3.2 but cites the Llama 2 paper [25]. Please cite the correct model release or technical report for Llama 3.2.
  4. [§3.2] The statement that 'adequate performance should be ≥ 90%' introduces an arbitrary threshold. Please justify it or explicitly label it as a descriptive criterion rather than a statistical standard.

Circularity Check

1 steps flagged · score 2.0 of 10

No substantive circularity: the main relational benchmark is externally defined, and the only self-referential step is the Appendix A scoring rubric that harvests acceptable object labels from the same VLMs under evaluation.

  1. self definitional [Appendix A, 'Assessing accuracy for Prompts 6-8']
    "If a VLM routinely generated a semantic alternative in place of an expected description 5+ times, that alternative was entered into a lookup table. If a response described an object or its semantic alternative, the object name was replaced by a placeholder, and that placeholder was used to assess accuracy."

    For the open-form prompts (6-8), the set of acceptable object labels is not fixed in advance by Objaverse or an independent human standard; it is induced from the outputs of the same VLMs being scored. Thus a model's own systematic mislabeling can be promoted to an 'acceptable variation' and then counted as correct, making the accuracy metric partly defined by the models it is used to measure. This does not affect the central closed-form left/right comparisons in Prompts 1-5, where the correct alternatives are fixed by the image geometry, but it is a genuine self-referential element in the open-ended scoring procedure.

full rationale

The paper's main claim is an empirical benchmark result, not a derived prediction: images and prompt variations are defined independently of model outputs, and accuracy is measured against the known left/right placement of objects in TABLE TEST. No fitted parameter is subsequently renamed as a prediction, and no uniqueness theorem or load-bearing result is imported from the authors' prior work. The normative assumption that logically equivalent prompts should yield identical accuracy and that human performance would be at ceiling is an unmeasured external standard, not a circular construction; it may be a correctness or validity risk, but it is not a reduction of the output to the input. The only detectable circular step is Appendix A, where acceptable object names for Prompts 6-8 are obtained from the same models' response frequencies. That lookup is lenient and model-dependent, but it applies only to open-ended object naming, not to the core left/right relational judgments, and it does not manufacture the observed unreliability. Accordingly, the central finding remains independently grounded despite a minor self-referential scoring element.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The benchmark introduces no fitted physical parameters or invented entities. The evaluation relies on two hand-set thresholds and several domain assumptions about human competence, logical equivalence, rendering unambiguity, and the validity of text-only controls.

free parameters (2)
  • acceptable performance threshold = 0.90
    The paper declares that adequate performance should be at least 90% accuracy, a hand-set benchmark for humanlike competence with no empirical human calibration (Section 3.2).
  • semantic alternative frequency threshold = 5
    Appendix A enters a model-generated alternative into the accuracy lookup table only if it occurs 5 or more times in that model's responses, an arbitrary cutoff that shapes reported accuracy for open-form prompts.
assumptions (4)
  • domain assumption Neurotypical adults perform at ceiling on the TableTest relation tasks
    The paper defines 'trivial' spatial cognition as behavior expected of most adults (Section 1.1) and assumes human accuracy is 1.0 without running human subjects.
  • domain assumption Logically equivalent prompt formulations should produce identical VLM accuracy
    This normative invariance principle is the basis for the reliability claim and appears throughout the General Discussion. It is an assumed standard of what reliability means, not a measured property.
  • domain assumption A 200px horizontal offset in rendered scenes guarantees unambiguous egocentric left/right relations
    Data construction places object centers 200px left or right of the table center (Section 2), and the paper assumes this yields unambiguous, uniformly scaled spatial relations for all 64 objects.
  • domain assumption Text-only vignettes that state spatial positions are a valid control condition
    Two-sentence vignettes are used as text-only analogues of images (Section 3.1); the authors note in Appendix B.7 that the text-only case can reduce to text-matching, which limits the control condition's validity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision language models are unreliable at trivial spatial cognition." pith.science (2026). https://pith.science/paper/ZXK52PLZ

@misc{pith2026250416061,
  author       = {Pith},
  title        = {Pith review of: Vision language models are unreliable at trivial spatial cognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXK52PLZ}},
  note         = {Machine review of arXiv:2504.16061}
}
read the original abstract

Vision language models (VLMs) are designed to extract relevant visuospatial information from images. Some research suggests that VLMs can exhibit humanlike scene understanding, while other investigations reveal difficulties in their ability to process relational information. To achieve widespread applicability, VLMs must perform reliably, yielding comparable competence across a wide variety of related tasks. We sought to test how reliable these architectures are at engaging in trivial spatial cognition, e.g., recognizing whether one object is left of another in an uncluttered scene. We developed a benchmark dataset -- TableTest -- whose images depict 3D scenes of objects arranged on a table, and used it to evaluate state-of-the-art VLMs. Results show that performance could be degraded by minor variations of prompts that use logically equivalent descriptions. These analyses suggest limitations in how VLMs may reason about spatial relations in real-world applications. They also reveal novel opportunities for bolstering image caption corpora for more efficient training and testing.

Figures

Figures reproduced from arXiv: 2504.16061 by the authors.

Figure 1
Figure 1. Examples from the TA [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Proportion accuracy from multimodal evaluations of VLM performance on spatial recognition prompts (see [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Proportion accuracy from text-only evaluations of VLM performance on spatial recognition prompts (see [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Evaluation of Multilingual Ability to Use Spatial Deictic Expressions in Vision-Language Models

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Open vision-language models fail to select spatial demonstratives based on object distance in a human-like manner across four languages.

Reference graph

Works this paper leans on

29 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Amant, J Gregory Trafton, et al

    Leslie M Blaha, Mitchell Abrams, Sarah A Bibyk, Claire Bonial, Beth M Hartzler, Christopher D Hsu, Sangeet Khemlani, Jayde King, Robert St. Amant, J Gregory Trafton, et al. Understanding is a process. Frontiers in Systems Neuroscience, 16:800280, 2022. 8

    work page 2022

  2. [2]

    Spatial reasoning

    Ruth MJ Byrne and Philip N Johnson-Laird. Spatial reasoning. Journal of memory and language , 28(5): 564–575, 1989. 2

    work page 1989

  3. [3]

    Spatialbot: Precise spatial understanding with vision language models

    Wenxiao Cai, Iaroslav Ponomarenko, Jianhao Yuan, Xiaoqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642 ,

    arXiv

  4. [4]

    Spa- tialvlm: Endowing vision-language models with spa- tial reasoning capabilities

    Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spa- tialvlm: Endowing vision-language models with spa- tial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 14455–14465, 2024

    work page 2024

  5. [5]

    Large language models are visual reasoning coor- dinators

    Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coor- dinators. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  6. [6]

    Spatialrgpt: Grounded spatial rea- soning in vision language model

    An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial rea- soning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 1

    arXiv 2024

  7. [7]

    What makes mental modeling difficult? normative data for the mul- tidimensional relational reasoning task

    Robert A Cortes, Adam B Weinberger, Griffin A Co- laizzi, Grace F Porter, Emily L Dyke, Holly O Keaton, Dakota L Walker, and Adam E Green. What makes mental modeling difficult? normative data for the mul- tidimensional relational reasoning task. Frontiers in psychology, 12:668256, 2021. 2

    work page 2021

  8. [8]

    Spatial com- munication systems across languages reflect universal action constraints

    Kenny R Coventry, Harmen B Gudde, Holger Dies- sel, Jacqueline Collier, Pedro Guijarro-Fuentes, Mila Vulchanova, Valentin Vulchanov, Emanuela Todisco, Maria Reile, Merlijn Breunesse, et al. Spatial com- munication systems across languages reflect universal action constraints. Nature human behaviour , 7(12): 2099–2110, 2023. 2

    work page 2023

Show all 29 references
  1. [9]

    Objaverse: A universe of annotated 3d ob- jects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d ob- jects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  2. [10]

    Scene-llm: Extending language model for 3d visual understanding and reasoning

    Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Wen- han Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 1

    2024 arXiv

  3. [11]

    Exploring the fron- tier of vision-language models: A survey of current methodologies and future directions

    Akash Ghosh, Arkadeep Acharya, Sriparna Saha, Vinija Jain, and Aman Chadha. Exploring the fron- tier of vision-language models: A survey of current methodologies and future directions. arXiv preprint arXiv:2404.07214, 2024. 1

    2024

  4. [12]

    Spatial lan- guage and spatial representation.Cognition, 55(1):39– 84, 1995

    William G Hayward and Michael J Tarr. Spatial lan- guage and spatial representation.Cognition, 55(1):39– 84, 1995. 2

    1995

  5. [13]

    Correctness comparison of chatgpt-4, gemini, claude-3, and copilot for spatial tasks

    Hartwig H Hochmair, Levente Juh ´asz, and Takoda Kemp. Correctness comparison of chatgpt-4, gemini, claude-3, and copilot for spatial tasks. Transactions in GIS, 2024. 1

    2024

  6. [14]

    What’s” up” with vision-language models? investi- gating their struggle with spatial reasoning

    Amita Kamath, Jack Hessel, and Kai-Wei Chang. What’s” up” with vision-language models? investi- gating their struggle with spatial reasoning. arXiv preprint arXiv:2310.19785, 2023. 1, 2, 8

    2023 arXiv

  7. [15]

    Space to reason: A spatial theory of human thought

    Markus Knauff. Space to reason: A spatial theory of human thought. Mit Press, 2013. 2

    2013

  8. [16]

    Whence and whither in spatial language and spatial cognition? Be- havioral and brain sciences, 16(2):255–265, 1993

    Barbara Landau and Ray Jackendoff. Whence and whither in spatial language and spatial cognition? Be- havioral and brain sciences, 16(2):255–265, 1993. 2

    1993

  9. [17]

    What matters when building vision- language models? arXiv preprint arXiv:2405.02246,

    Hugo Laurenc ¸on, L´eo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision- language models? arXiv preprint arXiv:2405.02246,

    arXiv

  10. [18]

    BLIP: Bootstrapping language-image pre- training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre- training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888–12900. PMLR, 2022. 3

    2022

  11. [19]

    A survey of attacks on large vision-language models: Resources, advances, and fu- ture trends

    Daizong Liu, Mingyu Yang, Xiaoye Qu, Pan Zhou, Yu Cheng, and Wei Hu. A survey of attacks on large vision-language models: Resources, advances, and fu- ture trends. arXiv preprint arXiv:2407.07403, 2024. 1

    2024 arXiv

  12. [20]

    Visual spatial reasoning

    Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635–651, 2023. 1 9

    2023

  13. [21]

    A survey on hallucination in large vision-language models

    Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 1

    2024 arXiv

  14. [22]

    Zero-shot visual reasoning by vision-language mod- els: Benchmarking and analysis

    Aishik Nagar, Shantanu Jaiswal, and Cheston Tan. Zero-shot visual reasoning by vision-language mod- els: Benchmarking and analysis. In 2024 Interna- tional Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2024. 2

    2024

  15. [23]

    A theory and a computational model of spatial reasoning with pre- ferred mental models

    Marco Ragni and Markus Knauff. A theory and a computational model of spatial reasoning with pre- ferred mental models. Psychological review, 120(3): 561, 2013. 2

    2013

  16. [24]

    Sparkle: Master- ing basic spatial capabilities in vision language mod- els elicits generalization to composite spatial reason- ing

    Yihong Tang, Ao Qu, Zhaokai Wang, Dingyi Zhuang, Zhaofeng Wu, Wei Ma, Shenhao Wang, Yunhan Zheng, Zhan Zhao, and Jinhua Zhao. Sparkle: Master- ing basic spatial capabilities in vision language mod- els elicits generalization to composite spatial reason- ing. arXiv preprint arX...

  17. [25]

    Llama 2: Open foundation and fine- tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine- tuned chat models. arXiv preprint arXiv:2307.09288,

    arXiv

  18. [26]

    Learning phys- ical parameters from dynamic scenes

    Tomer D Ullman, Andreas Stuhlm ¨uller, Noah D Goodman, and Joshua B Tenenbaum. Learning phys- ical parameters from dynamic scenes. Cognitive psy- chology, 104:57–82, 2018. 2

    2018

  19. [27]

    When and why vision-language models behave like bags-of-words, and what to do about it?, 2023

    Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it?, 2023. 2

    2023

  20. [28]

    Vision-language models for vision tasks: A sur- vey

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A sur- vey. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2024. 1

    2024

  21. [29]

    the [blank] is to the left of the [blank]

    Zheyuan Zhang, Fengyuan Hu, Jayjun Lee, Freda Shi, Parisa Kordjamshidi, Joyce Chai, and Ziqiao Ma. Do vision-language models represent space and how? evaluating spatial frame of reference under ambigui- ties. arXiv preprint arXiv:2410.17385, 2024. 2 10 A. Assessing accuracy fo...

    arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.