Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Can Multimodal Large Language Models Understand Spatial Relations?

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a new bounding-box-free, perspective-changing multiple-choice benchmark of 5,392 spatial questions scores today's best multimodal model at 48.14% accuracy, far below the 98.40% human level, with text-only performance…

desk verdict A useful new benchmark for spatial reasoning in MLLMs, but the in-image perspective labels need a sharper definition of observer orientation before the headline numbers are fully trustworthy. read the letter →

arxiv 2505.19015 v2 pith:D42YTSRA submitted 2025-05-25 cs.CV cs.MM

Jingping Liu , Ziyan Liu , Zhedong Cen , Yan Zhou , Yinan Zou , Weiyan Zhang , Haiyun Jiang , Tong Ruan This is my paper
classification cs.CVcs.MM
keywords spatialrelationreasoningmultimodallargelanguagemodelsbenchmarkmultiple-choicequestionansweringperspectivesubstitutionobjective-worldcoordinatesCOCO2017evaluationmethodology
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 tries to establish that today's multimodal large language models do not yet understand spatial relations the way humans do, and that previous benchmarks hid that gap. It constructs SpatialMQA, a human-annotated multiple-choice benchmark of 5,392 questions from COCO2017 images, built without bounding boxes and anchored to an explicit objective-world coordinate system. The benchmark deliberately mixes out-of-image, first-person, and third-person perspectives and filters out every question that could be answered from commonsense alone. On this test the strongest open model, instruction-tuned SpaceLLaVA, reaches 48.14% accuracy and GPT-4o reaches 40.20%, while humans reach 98.40% with images and only 24.40% with text alone. If the benchmark is sound, the wide gap is a concrete, reproducible diagnosis of where spatial reasoning fails.

What carries the argument

The load-bearing object is the spatial coordinate system defined in Table 2: the observer is the origin, gravity points down the negative Z-axis, the X-axis runs left-to-right, and the Y-axis runs back-to-front. Each of the six relations is an inequality on these axes, e.g., the subject is 'left of' the object when $x_s < x_o$ and 'on/above' when $z_s > z_o$. This converts perspective questions into one objective reference frame, lets annotators label first- and third-person views without changing the rule, and makes the benchmark's ground truth mechanical rather than impressionistic. Around it sits a three-round annotation procedure (annotate, check, review) and three question-template families that enforce the no-bbox, no-prior-knowledge, and perspective-substitution properties.

What would settle it

Re-annotate a random sample of, say, 200 SpatialMQA test items with a fresh team using the same Table 2 rules and compare labels. If pairwise agreement is far below the reported 98.40% human accuracy, or if flipping a question from first-person to third-person changes the SCS-derived label, the ground truth is not stable enough to support the claimed human-model gap.

Watch

Extended reading notes

Core claim

The central claim is that SpatialMQA is the first spatial-relation benchmark that combines four properties at once: no bounding boxes, objective-world labels, perspective substitution, and no prior-knowledge answers. Under that combination, the best model's accuracy is 48.14%, far below the 98.40% human score, and text-only human accuracy drops to 24.40%, essentially random. The authors attribute the failure to perspective substitution in particular, since errors of that type are the most frequent in their error analysis, followed by misrecognizing subjects and objects and lack of commonsense reasoning. The benchmark's value is diagnostic: each question is grounded in a six-relation coordinate system, so low scores can be traced to specific axes (X, Y, or Z) and specific question perspectives.

Load-bearing premise

The whole measurement rests on the assumption that the coordinate system in Table 2 tells annotators unambiguously which of the six relations holds for any image from any of the three perspectives; if 'front' or 'left' is genuinely underdetermined in even a modest fraction of images, both the 48.14% model score and the 98.40% human score become noisy, and the benchmark would reward annotation agreement more than spatial understanding.

Editorial extensions

If this is right

  • Any model that scores near 48% or below on SpatialMQA cannot be credited with human-like spatial grounding; the benchmark gives a single number that separates image-grounded reasoning from text-only guessing.
  • Instruction tuning on objective-world spatial data is a promising lever, since LoRA tuning lifted SpaceLLaVA from 31.32% to 48.14% accuracy.
  • Evaluation reports should separate X-, Y-, and Z-axis relations and the three question perspectives, because models are uneven across them (SpaceLLaVA ranges from 31.41% on Z-axis items to 56.00% on X-axis items).
  • Few-shot comparisons need to control whether in-context examples match the question type; GPT-4o's accuracy drops when examples are misaligned, while Gemini's rises with more examples.
  • The near-random text-only scores (24.40% for humans, 17.20-26.40% for models) confirm that SpatialMQA measures image understanding, not world knowledge.

Reading between the lines

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

  • Beyond the paper: because the spatial coordinate system is a formal rule, synthetic image generators or renderers could produce unlimited training samples with exact labels, testing whether data scale closes the human-model gap.
  • Beyond the paper: the coordinate rules imply a consistency check — asking the same question from first-person and third-person perspectives should produce complementary labels; a model that fails that swap reveals axis-based guessing rather than spatial grounding.
  • Beyond the paper: the coordinate system could be extended from six discrete relations to continuous 'how far' judgments, since it already encodes direction via coordinate differences but discards distance.
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

3 major / 4 minor

Summary. This paper presents SpatialMQA, a multiple-choice question-answering benchmark for spatial relation reasoning built on COCO2017. The benchmark contains 5,392 human-annotated samples covering six relations (left of, right of, in front of, behind, on/above, below) across three perspective types: out-of-image, first-person in-image, and third-person in-image. The authors claim that SpatialMQA avoids reliance on bounding boxes, filters out questions answerable from prior knowledge without an image, and tests perspective substitution. They evaluate several open- and closed-source MLLMs and report that the best model (SpaceLLaVA with LoRA) achieves 48.14% accuracy, far below human performance at 98.40%. The paper also includes analyses of question types, answer types, ICL alignment, image dependence, and error categories.

Significance. If the ground-truth labels are reliable, SpatialMQA addresses a real gap in spatial reasoning benchmarks by combining objective-world reference, perspective substitution, and a multiple-choice format that prevents bbox-based shortcuts. The released code, human evaluation, text-only control (24.40%), and the large human-model gap are strengths that support the benchmark's usefulness for future research. However, the benchmark's validity rests on the consistency of annotation conventions for in-image perspectives, which are currently under-specified, and on the comparability of evaluation protocols across model families. These issues must be resolved before the reported accuracy gap can be attributed to spatial reasoning ability.

major comments (3)
  1. [Section 3.2 and Table 2] The spatial coordinate system is defined with 'the observer as the origin' and the X/Y axes tied to the observer's left/right/front/back, but no rule is given for establishing the observer's orientation when the observer is inside the image. Since Table 3 reports that 60% of samples are in-image (Q2 first-person and Q3 third-person), the labels 'left of', 'right of', 'in front of', and 'behind' are not uniquely determined for objects without an intrinsic facing direction (e.g., a keyboard or bottle in Q2) or for living beings whose facing direction is not specified (Q3). The paper does not report inter-annotator agreement, so there is no evidence that the three annotators, two checkers, and reviewer all applied the same orientation convention. This under-specification is load-bearing because the human 98.40% accuracy and the model accuracies are measured against labels whose correctness depends on this unspecified convention.
  2. [Section 4 and Section 5.1] The train/dev/test split is stated only as a 7:1:2 ratio, without specifying whether the split is by image or by sample. If multiple samples from a COCO2017 image appear in both the training and test sets, instruction-tuned open-source models (e.g., SpaceLLaVA LoRA, Table 4) could benefit from image-level leakage, which would inflate the reported 48.14% accuracy. The authors should clarify the split granularity and, if the split is sample-level, re-split at the image level and re-report the affected results.
  3. [Section 5.1 and Table 4] The closed-source models (GPT-4o, Gemini-1.5-flash) and the human evaluation are run on 'randomly select 500 samples', while the open-source models are evaluated on the full test set of 1,076 samples. Table 4 presents all models in one table without indicating that the closed-source and human numbers are based on a subset. Unless the same stratified 500-sample subset is used for all these evaluations and is representative of the full test set, the comparison between GPT-4o (40.20%), Gemini (35.40%–38.00%), and SpaceLLaVA LoRA (48.14%) is not valid. This should be clarified, and the closed-source results should either be reported on the full test set or clearly separated.
minor comments (4)
  1. [Section 1, paragraph 2] 'To fully the scene of four students' is missing a verb; it should be 'To fully understand the scene of four students'.
  2. [Throughout] The model names 'LLaV A' and 'SpaceLLaV A' contain an unintended space (likely a LaTeX rendering issue); please fix throughout.
  3. [Section 5.3 and Table 5] The human evaluation is based on 500 samples, but no confidence intervals or per-question-type sample sizes are reported in Table 5, making it difficult to assess the stability of the 98.40% estimate.
  4. [Table 2] The formal conditions for 'in front of' and 'behind' use ys and yo without explicitly stating that these are signed distances along the observer's Y-axis; a one-sentence clarification would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SpatialMQA is an externally evaluated benchmark with human-labeled ground truth, so model and human accuracies are not forced by construction.

full rationale

This paper constructs a new benchmark and evaluates MLLMs against it; there is no formal derivation whose conclusion is equivalent to its inputs. The central quantities (human accuracy 98.40%, best MLLM 48.14%) are measurements on held-out test samples against independently human-annotated ground truth. The annotation guidelines define the spatial coordinate system; labels are produced by a three-round human process and checked by separate annotators, not fitted to the models being evaluated. The claim that questions cannot be answered from prior knowledge is empirically tested via a text-only condition (24.40%, near random), not assumed into the benchmark construction in a way that would force the accuracy results. No parameter in the paper is fitted to the reported accuracies, no uniqueness theorem is imported from the authors' own prior work, and no known result is renamed. Concern about possible ambiguity in the in-image coordinate system (how an inanimate first-person observer is oriented) is a validity or annotation-reliability limitation, not circularity, because it does not make the reported numbers equal to the benchmark's definition by construction. Therefore the paper is not circular.

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

The paper introduces no new theoretical entities or fitted parameters. The central claim rests on domain assumptions about the correctness of human annotations and the representativeness of the sample.

assumptions (3)
  • domain assumption The spatial coordinate system in Table 2 defines the correct answer for every question.
    Section 2 and Table 2 introduce a coordinate system with gravity pointing down and six spatial relations, but the orientation of the observer for out-of-image and third-person perspectives is not fully specified, making it an assumption that annotators applied it consistently.
  • domain assumption Human annotation is the gold standard for spatial relations.
    The benchmark uses human annotations as ground truth without independent verification against a formal definition or external benchmark, assuming human judgments are correct.
  • domain assumption The 5,392 samples are a representative and unbiased sample of spatial relations in COCO images.
    The annotation process selected 30,000 images and filtered based on annotator decisions, so sample selection is assumed to capture a fair distribution of relations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Multimodal Large Language Models Understand Spatial Relations?." pith.science (2026). https://pith.science/paper/D42YTSRA

@misc{pith2026250519015,
  author       = {Pith},
  title        = {Pith review of: Can Multimodal Large Language Models Understand Spatial Relations?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D42YTSRA}},
  note         = {Machine review of arXiv:2505.19015}
}
read the original abstract

Spatial relation reasoning is a crucial task for multimodal large language models (MLLMs) to understand the objective world. However, current benchmarks have issues like relying on bounding boxes, ignoring perspective substitutions, or allowing questions to be answered using only the model's prior knowledge without image understanding. To address these issues, we introduce SpatialMQA, a human-annotated spatial relation reasoning benchmark based on COCO2017, which enables MLLMs to focus more on understanding images in the objective world. To ensure data quality, we design a well-tailored annotation procedure, resulting in SpatialMQA consisting of 5,392 samples. Based on this benchmark, a series of closed- and open-source MLLMs are implemented and the results indicate that the current state-of-the-art MLLM achieves only 48.14% accuracy, far below the human-level accuracy of 98.40%. Extensive experimental analyses are also conducted, suggesting the future research directions. The benchmark and codes are available at https://github.com/ziyan-xiaoyu/SpatialMQA.git.

Figures

Figures reproduced from arXiv: 2505.19015 by the authors.

Figure 1
Figure 1. Samples from spatial relation reasoning bench [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distributions of subject and object types. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Distribution of error types. form similarly to random selection and significantly underperform MLLMs with I+Q+O. This indicates that our benchmark heavily relies on image inputs and cannot depend solely on the model’s prior knowledge. 2) MLLMs with Q+O still perform significantly better than random selection (17.20%) when given a fixed set of six options. This is be￾cause some of the options in this set contradict c… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Error examples. (a), (b), (c), and (d) describe examples of the IRSO, FRS, LCR, and IILN types, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: First-round annotation page in our tool. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 7
Figure 7. Figure 7: Human evaluation page in our tool [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. AnatomiX, an Anatomy-Aware Grounded Multimodal Large Language Model for Chest X-Ray Interpretation

    cs.CV 2026-01 conditional novelty 7.0 of 10

    AnatomiX, a two-stage anatomy-first multimodal LLM for chest X-ray interpretation, reports >25% relative gains on anatomy grounding and grounded captioning, but some aggregate benchmark numbers are internally inconsis...

  2. Spatial-aware Vision Language Model for Autonomous Driving

    cs.CV 2025-12 conditional novelty 6.0 of 10

    LVLDrive fuses LiDAR point clouds into a pre-trained vision-language driving model through a zero-initialized gated Q-Former and a spatial-aware QA dataset, improving open-loop planning and 3D grounding over image-onl...

Reference graph

Works this paper leans on

37 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Anja Belz, Adrian Muscat, Pierre Anguill, Mouhamadou Sow, Ga \'e tan Vincent, and Yassine Zinessabah. 2018. Spatialvoc2k: A multilingual dataset of images with annotations and features for spatial relations between objects. In Proceedings of the 11th International Conference on Natural Language Generation, pages 140--145

    work page 2018

  5. [5]

    Wenxiao Cai, Yaroslav Ponomarenko, Jianhao Yuan, Xiaoqi Li, Wankou Yang, Hao Dong, and Bo Zhao. 2024. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642

    arXiv 2024

  6. [6]

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

    2024

  7. [7]

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

    arXiv 2024

  8. [8]

    Sijie Cheng, Zhicheng Guo, Jingwen Wu, Kechen Fang, Peng Li, Huaping Liu, and Yang Liu. 2023. Can vision-language models think from a first-person perspective? arXiv preprint arXiv:2311.15596

    arXiv 2023

Show all 37 references
  1. [9]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. 2024. Instructblip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36

    2024

  2. [10]

    Mengfei Du, Binhao Wu, Zejun Li, Xuanjing Huang, and Zhongyu Wei. 2024. Embspatial-bench: Benchmarking spatial understanding for embodied tasks with large vision-language models. arXiv preprint arXiv:2406.05756

    2024 arXiv

  3. [11]

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. 2023. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394

    2023 arXiv

  4. [12]

    Haoxiang Gao, Yaqian Li, Kaiwen Long, Ming Yang, and Yiqing Shen. 2024. A survey for foundation models in autonomous driving. arXiv preprint arXiv:2402.01105

    2024 arXiv

  5. [13]

    Ankit Goyal, Kaiyu Yang, Dawei Yang, and Jia Deng. 2020. Rel3d: A minimally contrastive benchmark for grounding spatial relations in 3d. Advances in Neural Information Processing Systems, 33:10514--10525

    2020

  6. [14]

    Zixian Guo, Bowen Dong, Zhilong Ji, Jinfeng Bai, Yiwen Guo, and Wangmeng Zuo. 2023. Texts as images in prompt tuning for multi-label image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2808--2817

    2023

  7. [15]

    Zachary Hawes and Daniel Ansari. 2020. What explains the relationship between spatial and mathematical skills? a review of evidence from brain and behavior. Psychonomic bulletin & review, 27:465--482

    2020

  8. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  9. [17]

    Drew A Hudson and Christopher D Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700--6709

    2019

  10. [18]

    Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE CVPR, pages 2901--2910

    2017

  11. [19]

    Hugo Lauren c on, Daniel van Strien, Stas Bekman, Leo Tronchon, Lucile Saulnier, Thomas Wang, Siddharth Karamcheti, Amanpreet Singh, Giada Pistilli, Yacine Jernite, et al. 2023. Introducing idefics: An open reproduction of state-of-the-art visual language model, 2023. URL http...

    2023

  12. [20]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730--19742. PMLR

    2023

  13. [21]

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

    2022

  14. [22]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, ...

    2014

  15. [23]

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

    2023

  16. [24]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024. Visual instruction tuning. Advances in neural information processing systems, 36

    2024

  17. [25]

    Weihao Liu, Fangyu Lei, Tongxu Luo, Jiahe Lei, Shizhu He, Jun Zhao, and Kang Liu. 2023 b . Mmhqa-icl: Multimodal in-context learning for hybrid question answering over text, tables and images. arXiv preprint arXiv:2309.04790

    2023 arXiv

  18. [26]

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2021. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602

    2021 arXiv

  19. [27]

    Michael J Proulx, Orlin S Todorov, Amanda Taylor Aiken, and Alexandra A de Sousa. 2016. Where am i? who am i? the relation between spatial cognition, social cognition and individual differences in the built environment. Frontiers in psychology, 7:158846

    2016

  20. [28]

    Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. 2023. Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning. In The Twelfth International Conference on Learning Representations

    2023

  21. [29]

    Alane Suhr, Mike Lewis, James Yeh, and Yoav Artzi. 2017. A corpus of natural language for visual reasoning. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 217--223

    2017

  22. [30]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2023 arXiv

  23. [31]

    Junjie Wang, Wei Li, Yinjian Wang, Ran Tao, and Qian Du. 2023. Representation-enhanced status replay network for multisource remote-sensing image classification. IEEE Transactions on Neural Networks and Learning Systems

    2023

  24. [32]

    Lei Wang, Yi Hu, Jiabang He, Xing Xu, Ning Liu, Hui Liu, and Heng Tao Shen. 2024. T-sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38...

    2024

  25. [33]

    Chuan Wen, Dinesh Jayaraman, and Yang Gao. 2024. Can transformers capture spatial relations between objects? arXiv preprint arXiv:2403.00729

    2024 arXiv

  26. [34]

    Kaiyu Yang, Olga Russakovsky, and Jia Deng. 2019. Spatialsense: An adversarially crowdsourced benchmark for spatial relation recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2051--2060

    2019

  27. [35]

    Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. 2023. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178

    2023 arXiv

  28. [36]

    Daoan Zhang, Junming Yang, Hanjia Lyu, Zijian Jin, Yuan Yao, Mingkai Chen, and Jiebo Luo. 2024. Cocot: Contrastive chain-of-thought prompting for large multimodal models with multiple image inputs. arXiv preprint arXiv:2401.02582

    2024 arXiv

  29. [37]

    Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. 2023. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168--5191

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.