Pith. sign in

REVIEW 1 major objections 2 minor 1 cited by

HyperVis: Continuous Latent Visual Relational Graphs on the Lorentz Hyperboloid for Compositional Reasoning

T0 review · 1 major / 2 minor · reviewed 2026-06-28 · grok-4.3

Pith's one-line read Projecting dense visual relation tensors onto a Lorentz hyperboloid with IoA-driven entailment cones improves compositional reasoning in vision-language models.

desk verdict HyperVis shows modest gains on GQA and SugarCrepe by embedding a dense visual relation tensor into Lorentz space with custom losses, but the claim that this tensor supplies usable hierarchy rests on an untested assumption. read the letter →

arxiv 2606.06100 v1 pith:5EL3HMM6 submitted 2026-06-04 cs.CV

classification cs.CV
keywords hyperbolicembeddingvisualrelationscompositionalreasoningvision-languagemodelsLorentzhyperboloidentailmentconesscenegraphsGQA
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 demonstrates that adding discrete scene graph triplets from an off-the-shelf generator reduces GQA accuracy from 60.38% to 58.86% because the text labels clash with continuous visual features. It replaces this with a dense O(N²) visual relation tensor computed directly from class-agnostic region proposals through spatially-biased cross-attention. The tensor is embedded in hyperbolic space where hierarchy is maintained by IoA-based entailment cones and exterior-angle repulsion. Used as a training regularizer the hyperbolic losses improve GQA to 61.03%; used as an inference encoder the resulting prefix tokens raise SugarCrepe to 79.94%. Curvature settles at κ=4.0 instead of collapsing, and a Euclidean ablation shows the compositionality lift is specific to the hyperbolic geometry.

What carries the argument

The Lorentz hyperboloid with IoA-driven entailment cones and exterior-angle repulsion that enforces hierarchy on the continuous visual relation tensor.

What would settle it

Training the identical relational pipeline in Euclidean space and finding no reduction in the SugarCrepe compositionality gain relative to the hyperbolic version would falsify the necessity of hyperbolic geometry.

Watch

Extended reading notes

Core claim

HyperVis computes a dense O(N²) visual relation tensor from class-agnostic region proposals via spatially-biased cross-attention, projects the tensor onto the Lorentz hyperboloid, and enforces hierarchy through IoA-driven entailment cones together with exterior-angle repulsion. This continuous relational structure serves as both a training-time regularizer that shapes LoRA representations for generative VQA and an inference-time encoder that supplies hyperbolic prefix tokens for discriminative scoring, with the learned curvature stabilizing at κ=4.0.

Load-bearing premise

The dense O(N²) visual relation tensor computed from spatially-biased cross-attention on class-agnostic region proposals faithfully represents inter-object relationships that benefit from hyperbolic hierarchy enforcement.

Editorial extensions

If this is right

  • The hyperbolic relational losses act as a training regularizer that raises GQA from 57.21% (LoRA alone) to 61.03%.
  • Hyperbolic prefix tokens at inference time raise SugarCrepe by 6.25 percentage points over baseline.
  • The relational pipeline alone regularizes comparably in Euclidean space (GQA 60.81%), but the compositionality improvement is specifically hyperbolic (+4.58pp on SugarCrepe).
  • Entailment loss remains approximately six times higher when the same pipeline is trained in Euclidean space.

Reading between the lines

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

  • The approach could be tested on other relational tasks such as visual entailment or referring expression comprehension where inter-object structure matters.
  • If the dense tensor proves reliable, future VLMs might drop external scene-graph generators entirely in favor of learned continuous relations.
  • The high stable curvature suggests that visual scene hierarchies occupy a different geometric regime than the near-Euclidean embeddings common in prior hyperbolic VLMs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 2 minor

Summary. The paper claims that discrete scene-graph triplets from off-the-shelf SGG degrade VLM compositional performance (GQA drops from 60.38% to 58.86%), and proposes HyperVis: a dense O(N²) visual-relation tensor computed via spatially-biased cross-attention on class-agnostic region proposals, projected onto the Lorentz hyperboloid, with hierarchy enforced by IoA-driven entailment cones and exterior-angle repulsion. HyperVis is used both as a training regularizer (improving GQA to 61.03%) and as an inference-time encoder (SugarCrepe 79.94%, +6.25 pp). The learned curvature stabilizes at κ=4.0; a Euclidean ablation shows comparable regularization but a 4.58 pp gap on SugarCrepe and ~6× higher entailment loss, supporting that the compositionality gain is specifically hyperbolic.

Significance. If the central results hold, the work supplies concrete evidence that strongly curved hyperbolic geometry can regularize continuous visual features for relational compositionality without discrete semantic bottlenecks, together with an explicit Euclidean control and a reproducible curvature-stabilization observation. The availability of code and the two-mode (regularizer vs. encoder) evaluation strengthen the contribution.

major comments (1)
  1. [Method (tensor construction and entailment-cone section)] The load-bearing assumption that the dense O(N²) visual-relation tensor computed from class-agnostic proposals encodes hierarchical semantic relations amenable to IoA-driven entailment cones is not directly tested. The Euclidean ablation shows only a modest GQA difference (60.81% vs 61.03%) while the SugarCrepe gap is larger; however, no diagnostic (e.g., relation-label correlation or hierarchy-recovery metric on the tensor itself) is reported to confirm that the tensor contains the hierarchical structure the cones are intended to exploit rather than merely acting as a differentiable spatial regularizer.
minor comments (2)
  1. [Abstract / §4] The abstract states that SGG “backfires” with a single pair of numbers; a short table or figure showing the degradation across multiple seeds or datasets would strengthen the claim.
  2. [Method] Notation for the spatially-biased cross-attention and the precise definition of IoA-driven entailment cones should be expanded with a short equation block for reproducibility.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the significance of the results. We address the single major comment below.

read point-by-point responses
  1. Referee: [Method (tensor construction and entailment-cone section)] The load-bearing assumption that the dense O(N²) visual-relation tensor computed from class-agnostic proposals encodes hierarchical semantic relations amenable to IoA-driven entailment cones is not directly tested. The Euclidean ablation shows only a modest GQA difference (60.81% vs 61.03%) while the SugarCrepe gap is larger; however, no diagnostic (e.g., relation-label correlation or hierarchy-recovery metric on the tensor itself) is reported to confirm that the tensor contains the hierarchical structure the cones are intended to exploit rather than merely acting as a differentiable spatial regularizer.

    Authors: We agree that a direct diagnostic would strengthen the claim. The Euclidean ablation and the ~6× higher entailment loss provide indirect evidence that the cones exploit hierarchy rather than acting only as a spatial regularizer, but these are performance-level indicators. In the revision we will add an explicit diagnostic on the tensor: Pearson correlation of tensor entries against ground-truth hierarchical relations extracted from GQA annotations, plus a hierarchy-recovery accuracy on synthetic entailment structures. This will be reported in the method and experiments sections. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical results and ablations are independent of inputs

full rationale

The paper introduces HyperVis as a proposed architecture that computes a visual relation tensor from class-agnostic proposals, projects to the Lorentz model, and applies IoA-driven cones plus repulsion losses. Reported gains on GQA (as regularizer) and SugarCrepe (as encoder) are measured outcomes of training runs, not quantities derived by algebraic reduction from the same inputs. The curvature value κ=4.0 is obtained by optimization and reported as an observed stabilization; the Euclidean ablation supplies a separate control that isolates the hyperbolic contribution on SugarCrepe without any fitted parameter being relabeled as a prediction. No self-citations, uniqueness theorems, or ansatzes are invoked to close the argument, and no equation equates a claimed result to its own construction. The derivation chain therefore remains self-contained against external benchmarks.

Assumptions & free parameters 1 free parameters · 1 assumptions · 0 invented entities

The paper relies on standard assumptions of hyperbolic geometry and the validity of the cross-attention for relations, with the main addition being the specific losses and the projection.

free parameters (1)
  • curvature κ = 4.0
    Stabilizes at this value during training, indicating it is fitted to the visual data.
assumptions (1)
  • domain assumption Hyperbolic space with high curvature is necessary for representing hierarchical visual relations from continuous features
    Used to justify why κ=4.0 is meaningful and why Euclidean underperforms on compositionality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperVis: Continuous Latent Visual Relational Graphs on the Lorentz Hyperboloid for Compositional Reasoning." pith.science (2026). https://pith.science/paper/5EL3HMM6

@misc{pith2026260606100,
  author       = {Pith},
  title        = {Pith review of: HyperVis: Continuous Latent Visual Relational Graphs on the Lorentz Hyperboloid for Compositional Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5EL3HMM6}},
  note         = {Machine review of arXiv:2606.06100}
}
abstract

Vision-Language Models (VLMs) struggle with compositional reasoning that requires understanding inter-object relationships. A natural remedy is to inject explicit scene graph triplets $\langle s, p, o \rangle$ from an off-the-shelf scene graph generator (SGG), but we show this backfires: discrete text labels collide with the continuous visual modality, degrading GQA accuracy from 60.38\% to 58.86\%. We propose \textbf{HyperVis}, which bypasses the SGG semantic bottleneck entirely. From $N$ class-agnostic region proposals, we compute a dense $O(N^2)$ visual relation tensor via spatially-biased cross-attention, project it onto a Lorentz hyperboloid, and enforce hierarchy through spatial physics, namely IoA-driven entailment cones and exterior-angle repulsion. We discover that HyperVis contributes in two complementary ways: (1) as a \emph{training-time regularizer}, the hyperbolic relational losses shape LoRA representations that improve generative VQA (GQA 61.03\% vs.\ 57.21\% for LoRA fine-tuning without relational losses, recovering and surpassing the baseline); and (2) as an \emph{inference-time relational encoder}, hyperbolic prefix tokens boost discriminative compositional scoring (SugarCrepe 79.94\%, $+$6.25pp over baseline). The learned curvature stabilises at $\kappa{=}4.0$, an order of magnitude above prior hyperbolic VLMs where $\kappa$ typically collapses toward zero, indicating that continuous visual features genuinely require the exponential volume of strongly curved space. A controlled Euclidean ablation confirms this decomposition: the relational pipeline regularises LoRA comparably in flat space (GQA 60.81\%), but the compositionality gain is specifically hyperbolic (SugarCrepe $+$4.58pp over Euclidean), with entailment loss ${\sim}6{\times}$ higher in Euclidean training. Codes are available at TBA.

Figures

Figures reproduced from arXiv: 2606.06100 by the authors.

Figure 1
Figure 1. Overview of HyperVis. No SGG, no predicate vocabulary, the object relationships are continuous mono-modal [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. IoA-driven hierarchical geometry in hyperbolic space. Spatial containment in image space (left) maps to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Training dynamics. (a) Curvature κ with vs. without angle loss. (b) Entailment loss: Lorentz vs. Euclidean. (c) GQA accuracy curves. yellow school bus Image & Question Retrieved (Top-K = 3) Q. What is sitting on the table in front of the microwave? bowl of fruit on table 0.86 GT: fruit fruit on plate on table 0.71 table in kitchen 0.39 Q. What color is the bus driving down the wet street? 0.89 GT: yellow wet street … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative behaviour of the hyperbolic Top- [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: (Left) Fixed versus learned curvature analysis, where GQA accuracy [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Why prefix tokens hurt generation but help scoring. The L2-norm mismatch distorts autoregressive attention [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Learnable gates in VLM auxiliary paths close themselves; non-optional, norm-matched injection plus hyperbolic geometric losses preserves relational accuracy where soft prompts trade it away.

Reference graph

Works this paper leans on

46 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1]

    Making the V in VQA matter: Elevating the role of image understanding in visual question answering

    Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. InCVPR, 2017

  2. [2]

    GQA: A new dataset for real-world visual reasoning and composi- tional question answering

    Drew A Hudson and Christopher D Manning. GQA: A new dataset for real-world visual reasoning and composi- tional question answering. InCVPR, 2019

  3. [3]

    Winoground: Probing vision and language models for visio- linguistic compositionality

    Tristan Thrush, Ryan Jiang, Max Bartolo, et al. Winoground: Probing vision and language models for visio- linguistic compositionality. InCVPR, 2022

  4. [4]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, et al. Learning transferable visual models from natural language supervision. In ICML, 2021

  5. [5]

    BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InICML, 2023

  6. [6]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InNeurIPS, 2024

  7. [7]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuqian Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024

  8. [8]

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

Show all 46 references
  1. [9]

    Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality

    Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. InNeurIPS, volume 36, pages 31096–31116, 2023

  2. [10]

    SA-VQA: Structured alignment of visual and semantic representations for visual question answering.arXiv:2201.10654, 2022

    Peixi Xiong et al. SA-VQA: Structured alignment of visual and semantic representations for visual question answering.arXiv:2201.10654, 2022

  3. [11]

    LLaV A-SG: Leveraging scene graphs as visual semantic expression in vision-language models.ICASSP, 2025

    Jingyi Wang, Jianzhong Ju, Jian Luan, and Zhidong Deng. LLaV A-SG: Leveraging scene graphs as visual semantic expression in vision-language models.ICASSP, 2025

  4. [12]

    Compositional chain-of-thought prompting for large multimodal models

    Chancharik Mitra, Brendan Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. InCVPR, 2024

  5. [13]

    Hyperbolic image-text representations

    Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic image-text representations. InICML, 2023

  6. [14]

    Accept the modality gap: An exploration in the hyperbolic space

    Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. InCVPR, 2024

  7. [15]

    Show and tell: A neural image caption generator

    Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. InCVPR, 2015

  8. [16]

    Multimodal compact bilinear pooling for visual question answering and visual grounding

    Akira Fukui, Dong Huk Park, Daylen Yang, et al. Multimodal compact bilinear pooling for visual question answering and visual grounding. InEMNLP, 2016

  9. [17]

    VL-BERT: Pre-training of generic visual-linguistic representations

    Weijie Su, Xizhou Zhu, Yue Cao, et al. VL-BERT: Pre-training of generic visual-linguistic representations. In ICLR, 2020

  10. [18]

    ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. InNeurIPS, 2019

  11. [19]

    UNITER: Universal image-text representation learning

    Yen-Chun Chen, Linjie Li, Licheng Yu, et al. UNITER: Universal image-text representation learning. InECCV, 2020

  12. [20]

    Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12192, 2024

    Peng Wang, Shuai Bai, Hao Sinian, et al. Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12192, 2024

  13. [21]

    InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InCVPR, 2024

  14. [22]

    Scene graph generation by iterative message passing

    Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. InCVPR, 2017

  15. [23]

    Graphical contrastive losses for scene graph generation

    Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph generation. InCVPR, 2019

  16. [24]

    Panoptic scene graph generation

    Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. InECCV, 2022

  17. [25]

    From pixels to graphs: Open-vocabulary scene graph generation with vision-language models

    Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. InCVPR, 2024. 13 HyperVis: Continuous Latent Visual Relational Graphs

  18. [26]

    Incorporating structured representations into pretrained vision and language models using scene graphs

    Roei Herzig, Amir Mendelson, Leonid Karlinsky, et al. Incorporating structured representations into pretrained vision and language models using scene graphs. InEMNLP, 2023

  19. [27]

    EGTR: Extracting graph from transformer for scene graph generation

    Jinbae Im, JeongYeon Nam, Nokyung Park, Hyungmin Lee, and Seunghyun Park. EGTR: Extracting graph from transformer for scene graph generation. InCVPR, 2024

  20. [28]

    Leveraging predicate and triplet learning for scene graph generation

    Jiankai Li et al. Leveraging predicate and triplet learning for scene graph generation. InCVPR, 2024

  21. [29]

    Visual genome: Connecting language and vision using crowdsourced dense image annotations.IJCV, 123:32–73, 2017

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations.IJCV, 123:32–73, 2017

  22. [30]

    Neural motifs: Scene graph parsing with global context

    Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. InCVPR, 2018

  23. [31]

    Poincaré embeddings for learning hierarchical representations

    Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. InNeurIPS, 2017

  24. [32]

    Hyperbolic image embeddings

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. InCVPR, 2020

  25. [33]

    Hyperbolic neural networks

    Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. InNeurIPS, 2018

  26. [34]

    Inferring concept hierarchies from text corpora via hyperbolic embeddings

    Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximillian Nickel. Inferring concept hierarchies from text corpora via hyperbolic embeddings. InACL, 2019

  27. [35]

    Order-embeddings of images and language

    Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. Order-embeddings of images and language. In ICLR, 2016

  28. [36]

    Compositional entailment learning for hyperbolic vision-language models

    Avik Pal, Max van Spengler, Guido Maria D’Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. InICLR, 2025

  29. [37]

    Hyperbolic safety-aware vision-language models

    Tobia Poppi, Tejaswi Kasarla, Pascal Mettes, Lorenzo Baraldi, and Rita Cucchiara. Hyperbolic safety-aware vision-language models. InCVPR, 2025

  30. [38]

    HyperET: Efficient training in hyperbolic space for multi-modal large language models

    Zelin Peng, Zhengqin Xu, Qingyang Liu, Xiaokang Yang, and Wei Shen. HyperET: Efficient training in hyperbolic space for multi-modal large language models. InNeurIPS, 2025

  31. [39]

    Multi-relational Poincaré graph embeddings

    Ivana Balazevic, Carl Allen, and Timothy Hospedales. Multi-relational Poincaré graph embeddings. InNeurIPS, 2019

  32. [40]

    Low-dimensional hyperbolic knowledge graph embeddings

    Ines Chami, Adva Wolf, Da-Cheng Crouse, et al. Low-dimensional hyperbolic knowledge graph embeddings. In ACL, 2020

  33. [41]

    RotatE: Knowledge graph embedding by relational rotation in complex space

    Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. RotatE: Knowledge graph embedding by relational rotation in complex space. InICLR, 2019

  34. [42]

    Mind the gap: Under- standing the modality gap in multi-modal contrastive representation learning

    Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Under- standing the modality gap in multi-modal contrastive representation learning. InNeurIPS, 2022

  35. [43]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. InECCV, 2022

  36. [44]

    LLaMA-Adapter V2: Parameter-efficient visual instruction model

    Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xiangyu Yue, Hongsheng Li, and Yu Qiao. LLaMA-Adapter V2: Parameter-efficient visual instruction model. In arXiv preprint arXiv:2304.15010, 2023

  37. [45]

    World Scientific, 2005

    Abraham Albert Ungar.Analytic hyperbolic geometry: Mathematical foundations and applications. World Scientific, 2005

  38. [46]

    Why is winoground hard? investigating failures in visuolinguistic compositionality

    Anuj Diwan, Layne Berry, Eunsol Choi, David Harwath, and Kyle Mahowald. Why is winoground hard? investigating failures in visuolinguistic compositionality. InEMNLP, 2022. 14

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.