pith. sign in

arxiv: 2606.06714 · v2 · pith:5FGSFHM5new · submitted 2026-06-04 · 💻 cs.CV

Anchored, Not Graded: Vision-Language Models Fail at Slant-from-Texture Perception

Qian Zhang , Michal Golovanevsky , Fulvio Domini , James Tompkin This is my paper

Pith reviewed 2026-07-02 22:54 UTC · model grok-4.3

classification 💻 cs.CV
keywords vision-language modelsslant perceptiontexture cuesanchoringgeometric perceptionmodel evaluationfine-tuningpsychophysics
0
0 comments X

The pith

Vision-language models anchor slant-from-texture predictions to a few discrete angles instead of producing graded outputs.

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

The paper examines whether VLMs can match human competence at perceiving surface slant from texture, which produces reliable graded biases in psychophysical tests. It reports that multiple VLM families and scales, under both zero-shot and in-context prompting, restrict predictions to a small set of anchor angles such as 0°, ±25°, and ±45° with almost no dependence on field of view, optical slant, or surface curvature. Supervised fine-tuning reduces the severity of anchoring but leaves it intact. The authors locate the failure at the step that maps internal representations to language output rather than an absence of geometric encoding altogether.

Core claim

Across VLM families and scales, zero-shot and in-context prompting both produce slant predictions restricted to a small set of anchors (0°, ±25°, ±45°) that show little dependence on stimulus field of view, optical slant, or surface curvature; supervised fine-tuning partially remediates the failure but residual anchoring persists.

What carries the argument

Anchoring effect, in which slant outputs collapse to discrete fixed values independent of continuous changes in the texture stimulus.

If this is right

  • High-level vision-language benchmarks can be solved without sensitivity to low-level geometric cues such as slant.
  • The observed failure occurs at the representation-to-output language interface rather than from missing geometric encoding.
  • Supervised fine-tuning on slant tasks can reduce but does not remove the anchoring pattern.
  • This pattern differs from unsupervised CNNs, which reproduce several human-like graded biases on the same stimuli.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Training regimes that explicitly penalize discrete output distributions might be needed to elicit graded geometric judgments from VLMs.
  • Applications requiring precise surface orientation estimates, such as robotic grasping or augmented reality overlays, may encounter systematic errors traceable to this anchoring.
  • Similar anchoring could appear in other continuous perceptual dimensions (e.g., depth or curvature) when VLMs are asked to verbalize them.

Load-bearing premise

The chosen zero-shot and in-context prompting methods are sufficient to reveal the models' underlying perceptual representations rather than merely reflecting output formatting or language interface constraints.

What would settle it

A test in which the same VLM is shown identical slant-from-texture stimuli but asked to output a continuous numeric angle or to adjust a graded visual response, checking whether the output still collapses to the same small set of anchors.

Figures

Figures reproduced from arXiv: 2606.06714 by Fulvio Domini, James Tompkin, Michal Golovanevsky, Qian Zhang.

Figure 1
Figure 1. Figure 1: Slant-from-texture as a controlled perceptual task for VLMs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: VLMs anchor significantly on slant angle and sign prediction. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effects of prompt and task framing as boxplots of 95% Confidence Intervals [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fine tuning only partially helps. Qwen2.5-VL before and after supervised fine￾tuning: slant error decreases and anchoring weakens after SFT, but many outliers persist with high error. In-context learning does not help. In-context prompts do not significantly dif￾fer from other prompt families (Fig. 3b): Median slant errors remain above 40°, curvature-sign accuracy stays near chance, and mode percentages re… view at source ↗
Figure 5
Figure 5. Figure 5: SFT improves slant prediction but discrepancies remain. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise linear probe performance across the 32-layer Qwen2.5-VL-3B vision encoder. Optical slant is near ceiling from the first transformer layer (actual R 2 = 0.992–0.998; appears flat at this scale). FOV, physical slant, and curvature require progressively deeper processing, peaking around layers 17–19 before declining in the final layers. Embed = Conv3D patch projection output (before any transformer… view at source ↗
Figure 7
Figure 7. Figure 7: Layerwise probing across four VLMs, extending the single-model analysis in Qwen2.5-VL-3B to LLaVA-1.5-7B (CLIP-ViT-L/14), PaliGemma-3B (SigLIP-So400M), and Chameleon-7B (VQ-VAE). The result is consistent with the encoding being architecture-general, with the readout bottleneck arising at the language interface rather than in the vision tower. 3.3 Probing the Vision Module Pre-trained VLMs are trained on mi… view at source ↗
Figure 8
Figure 8. Figure 8: Layer-wise linear probe performance across the 32-layer Qwen2.5-VL-3B vision encoder. Optical slant is near ceiling from the first transformer layer (actual R 2 = 0.992–0.998; appears flat at this scale). FOV, physical slant, and curvature require progressively deeper processing, peaking around layers 17–19 before declining in the final layers. Embed = Conv3D patch projection output (before any transformer… view at source ↗
Figure 8
Figure 8. Figure 8: Readout bottleneck - VLM output anchors (top); LM-input vision tokens encode continuous slant (bottom). We extract the mean-pooled post-projector vision tokens (in LM embedding space) from Qwen, LLaVa, and PaliGemma and train a linear regressor on train image tokens to predict physical slant. The figure shows results of the held-out test images. Readout test. To test whether this is a true readout problem,… view at source ↗
Figure 9
Figure 9. Figure 9: Layerwise probing across four VLMs vision modules, extending the single-model analysis in Qwen2.5-VL-3B to LLaVA-1.5-7B (CLIP-ViT-L/14), PaliGemma-3B (SigLIP-So400M), and Chameleon-7B (VQ-VAE). The result is consistent with the encoding being architecture-general, with the readout bottleneck arising at the language interface rather than in the vision tower. Layer-wise probing. To trace how geometric inform… view at source ↗
Figure 9
Figure 9. Figure 9: Single attention head mean ablation on Qwen2.5-VL-3B (top) and Qwen2-VL-7B (bottom) slant prediction. (a) Baseline: predictions anchor at −45◦ . (b) Ablating L0 H9: predictions barely change. (c) Ablating L0 H23: anchoring spreads across more discrete values. (d) Ablating L5 H19: a new anchor at −23◦ appears that matches the prompt exam￾ple (prompt copying). No ablation produces continuous stimulus-depende… view at source ↗
Figure 10
Figure 10. Figure 10: VLM, CNN, and human curvature sign judgement accuracy versus optical slant, shown in matching axes for direct comparison. Left: Performance of VLM Qwen2.5-VL (3B) before supervised fine-tuning (SFT). Center left: After SFT. Qwen2.5-VL shows improved accuracy with increasing field of view, but still maintains high variance and is overall a poor proxy for human ability. Center right: Unsupervised CNN judgem… view at source ↗
Figure 11
Figure 11. Figure 11: No single attention head ablation produces continuous stimulus-dependent predic￾tions. We apply mean ablation on Qwen2.5-VL-3B (top) and Qwen2-VL-7B (bottom) slant prediction. (a) Baseline predictions; (b, c, d) Typical results of ablating attention heads at different layers. The prediction barely changes; some even induced prompt copying, e.g. ablating L5 H19, a new anchor at “−23◦ ” appears that matches… view at source ↗
Figure 11
Figure 11. Figure 11: SFT error analysis grid by slant angle and FOV, showing frequency of sign flips in predictions. Larger FOVs, larger optical slant images induced fewer flips; no curvature sign errors for concave FOV> 45°; convex-to-concave misjudge is more frequent than vice versa [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Scatter plot of model predictions across 12 recent VLMs. Each panel shows predictions from a different model, with mean average error (MAE) for physical slant, accuracy for curvature sign discrimination, and the number of unique predicted values (out of 400 stimuli) in the title [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 12
Figure 12. Figure 12: Scatter plot of model predictions across 12 recent VLMs. Each panel shows predictions from a different model, with mean average error (MAE) for physical slant, accuracy for curvature sign discrimination, and the number of unique predicted values (out of 400 stimuli) in the title. in visual math problems? In: European Conference on Computer Vision. pp. 169–186. Springer (2024) 16 49. Zhang, Y., Pan, J., Zh… view at source ↗
Figure 13
Figure 13. Figure 13: Under technical prompt settings, mode value and frequency of results across models in heatmap. 4 Are failures due to out-of-distribution stimuli? Our stimuli are synthetic polka-dot textures rendered under controlled conditions to isolate texture-gradient cues. A natural concern is that these images are out of distri￾bution (OOD) relative to VLM training data, which consists predominantly of nat￾ural phot… view at source ↗
Figure 13
Figure 13. Figure 13: Under technical prompt settings, mode value and frequency of results across models in heatmap. 4 Are failures due to out-of-distribution stimuli? Our stimuli are synthetic polka-dot textures rendered under controlled conditions to isolate texture-gradient cues. A natural concern is that these images are out of distribution (OOD) relative to VLM training data, which consists predominantly of natural photog… view at source ↗
Figure 14
Figure 14. Figure 14: With the in-context modifiers and natural language prompt settings, mode value and frequency of results across models in heatmap. Two observations argue against a pure OOD account. First, geometric probing confirms that the vision encoder extracts slant-relevant information from these stimuli at high fidelity (Sec. 3.3), ruling out a failure of low-level visual processing. Polka-dot and regular blob textu… view at source ↗
Figure 14
Figure 14. Figure 14: With the in-context modifiers and natural language prompt settings, mode value and frequency of results across models in heatmap. Second, the failure pattern persists after supervised fine-tuning (SFT) on the task stimuli: even after training on thousands of polka-dot images, anchoring is not elim￾inated and sign-of-curvature accuracy remains below ceiling. An OOD explanation would predict that SFT, by fa… view at source ↗
read the original abstract

Human perception of surface slant from texture exhibits systematic, graded biases that emerge reliably in psychophysical experiments. Prior work showed that unsupervised CNNs reproduce several human-like biases, while supervised CNNs do not. Do Vision-Language Models (VLMs) exhibit similar competences? Across multiple VLM families and model scales, zero-shot and in-context prompting both produce distinctive failures: slant is predicted at only a small set of anchors (e.g., 0\degree, $\pm$25\degree, $\pm$45\degree) with little dependence on stimulus field of view, optical slant, or surface curvature. Supervised fine-tuning partially remediates the failure, but residual anchoring persists. While success in high-level vision-language benchmarks might not require sensitivity to low-level geometric cues, we interpret anchoring as a failure at the representation-to-output language interface: not necessarily an absence of geometric encoding, but a failure to express it in a graded form.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that vision-language models (VLMs) across families and scales fail to produce graded slant estimates from texture, instead anchoring predictions to discrete values (0°, ±25°, ±45°) with little sensitivity to field of view, optical slant, or curvature under zero-shot and in-context prompting. Supervised fine-tuning partially reduces but does not eliminate anchoring. The authors interpret this as a representation-to-output interface failure rather than absent geometric encoding.

Significance. If the empirical pattern holds, the work usefully extends prior CNN findings on slant-from-texture to VLMs and underscores that high-level VLM benchmarks may not require low-level geometric competence. Strengths include the multi-model, multi-scale design and explicit discussion of the language-interface alternative. The result could motivate targeted improvements in how VLMs express continuous visual quantities.

major comments (2)
  1. [Abstract] Abstract and Discussion: the central claim of a 'failure at slant-from-texture perception' rests on the tested prompting regimes being sufficient to surface graded representations. Without experiments using alternative elicitation (direct numerical regression, chain-of-thought, or internal readout), the observed anchors remain compatible with intact internal encoding that standard language output cannot express continuously.
  2. [Results] Results section: the repeated claim of 'little dependence' on FOV, optical slant, and curvature requires quantitative support (e.g., correlation values or statistical tests) rather than qualitative description; without these, the distinctiveness of the anchoring pattern relative to stimulus variation is difficult to evaluate.
minor comments (2)
  1. Figure captions should explicitly label the stimulus parameters varied in each panel to allow readers to verify the independence claims.
  2. The abstract's LaTeX notation (\degree) should be rendered consistently as ° in the published version.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will revise the manuscript to incorporate quantitative analyses and additional experiments where feasible.

read point-by-point responses

  1. Referee: [Abstract] Abstract and Discussion: the central claim of a 'failure at slant-from-texture perception' rests on the tested prompting regimes being sufficient to surface graded representations. Without experiments using alternative elicitation (direct numerical regression, chain-of-thought, or internal readout), the observed anchors remain compatible with intact internal encoding that standard language output cannot express continuously.

    Authors: Our manuscript already frames the anchoring as potentially reflecting a representation-to-output interface failure rather than absent geometric encoding. The persistence of discrete anchors across both zero-shot and in-context prompting regimes is consistent with this interpretation. We agree that further elicitation methods would strengthen the evidence and will add chain-of-thought prompting experiments in the revision. Direct numerical regression can also be tested; internal readout is not straightforward for the black-box VLMs studied but will be noted as a limitation. revision: yes

  2. Referee: [Results] Results section: the repeated claim of 'little dependence' on FOV, optical slant, and curvature requires quantitative support (e.g., correlation values or statistical tests) rather than qualitative description; without these, the distinctiveness of the anchoring pattern relative to stimulus variation is difficult to evaluate.

    Authors: We agree that the claims require quantitative backing. The revised manuscript will include correlation coefficients, regression models, and statistical tests (such as ANOVA) to measure dependence on FOV, optical slant, and curvature, allowing a clearer assessment of how the anchoring pattern stands out from stimulus variation. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation of model outputs

full rationale

The paper reports direct empirical observations of VLM outputs under zero-shot and in-context prompting on slant-from-texture stimuli. No equations, fitted parameters, or derivations are present that could reduce any claim to its inputs by construction. The central claim rests on measured anchoring patterns across model families, with explicit acknowledgment of alternative interpretations (representation-to-output interface). No self-citation chains or ansatzes are invoked as load-bearing support. This is a standard non-circular empirical study.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions about model evaluation and psychophysical stimulus validity with no free parameters or invented entities introduced.

axioms (1)
  • domain assumption Prompting methods can elicit perceptual judgments from VLMs in a manner comparable to human psychophysics
    Invoked when interpreting zero-shot and in-context results as evidence of perceptual failure

pith-pipeline@v0.9.1-grok · 5705 in / 1133 out tokens · 16153 ms · 2026-07-02T22:54:33.115393+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 16 canonical work pages · 13 internal anchors

  1. [1]

    Understanding intermediate layers using linear classifier probes

    Alain, G., Bengio, Y.: Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644 (2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  2. [2]

    Large language model (2026),https://claude.ai/, accessed: 2026-05-11

    Anthropic: Claude model (may 2026 version) [large language model]. Large language model (2026),https://claude.ai/, accessed: 2026-05-11

    2026

  3. [3]

    Qwen Technical Report

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al.: Qwen technical report. arXiv preprint arXiv:2309.16609 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    PaliGemma: A versatile 3B VLM for transfer

    Beyer*,L.,Steiner*,A.,Pinto*,A.S.,Kolesnikov*,A.,Wang*,X.,Salz,D.,Neumann, M., Alabdulmohsin, I., Tschannen, M., Bugliarello, E., Unterthiner, T., Keysers, D., Koppula, S., Liu, F., Grycner, A., Gritsenko, A., Houlsby, N., Kumar, M., Rong, K., Eisenschlos, J., Kabra, R., Bauer, M., Bošnjak, M., Chen, X., Minderer, M., Voigtlaender, P., Bica, I., Balazevic...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    In: 2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA)

    Cai, W., Ponomarenko, I., Yuan, J., Li, X., Yang, W., Dong, H., Zhao, B.: Spatialbot: Precise spatial understanding with vision language models. In: 2025 IEEE Interna- tional Conference on Robotics and Automation (ICRA). pp. 9490–9498. IEEE (2025)

    2025

  6. [6]

    Journal of Vision20(7), 14–14 (2020)

    Chen, Z., Saunders, J.A.: Multiple texture cues are integrated for perception of 3d slant from texture. Journal of Vision20(7), 14–14 (2020)

    2020

  7. [7]

    Advances in Neural Information Processing Systems37, 135062–135093 (2024)

    Cheng, A.C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., Liu, S.: Spatialrgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems37, 135062–135093 (2024)

    2024

  8. [8]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Danier, D., Aygun, M., Li, C., Bilen, H., Mac Aodha, O.: Depthcues: Evaluating monocular depth perception in large vision models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20049–20059 (2025)

    2025

  9. [9]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy,A.,Beyer,L.,Kolesnikov,A.,Weissenborn,D.,Zhai,X.,Unterthiner,T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  10. [10]

    Advances in neural information processing systems27 (2014)

    Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems27 (2014)

    2014

  11. [11]

    arXiv preprint arXiv:2312.11370 (2023)

    Gao, J., Pi, R., Zhang, J., Ye, J., Zhong, W., Wang, Y., Hong, L., Han, J., Xu, H., Li, Z., et al.: G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370 (2023)

    work page arXiv 2023

  12. [12]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3828–3838 (2019)

    2019

  13. [13]

    Vision research9(9), 1079–1094 (1969)

    Gogel, W.C.: The sensing of retinal size. Vision research9(9), 1079–1094 (1969)

    1969

  14. [14]

    com/deepmind- media/Model-Cards/Gemini-3-1-Flash- Lite- Model- Card.pdf , accessed: 2026-05-11

    Google: Gemini 3.1 flash-lite model card (2026),https://storage.googleapis. com/deepmind- media/Model-Cards/Gemini-3-1-Flash- Lite- Model- Card.pdf , accessed: 2026-05-11

    2026

  15. [15]

    In: Findings of the Association for Computational Linguistics: ACL 2025

    Huang, K.H., Qin, C., Qiu, H., Laban, P., Joty, S., Xiong, C., Wu, C.S.: Why vision language models struggle with visual arithmetic? towards enhanced chart and geometry understanding. In: Findings of the Association for Computational Linguistics: ACL 2025. pp. 4830–4843 (2025)

    2025

  16. [16]

    GPT-4o System Card

    Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024) VLMs Fail at Slant-from-Texture Perception 17

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Gemma 3 Technical Report

    Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Matejovicova, T., Ramé, A., Rivière, M., Rouillard, L., et al.: Gemma 3 technical report. arXiv preprint arXiv:2503.197864(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Artificial Intelligence38(1), 1–48 (1989)

    Kanatani, K.i., Chou, T.C.: Shape from texture: General principle. Artificial Intelligence38(1), 1–48 (1989)

    1989

  19. [19]

    Kemp, J.T., Vishwanath, D., Domini, F.: Sensory uncertainty does not drive perceptual discriminability in 3d vision (2024)

    2024

  20. [20]

    Vision research38(11), 1655–1682 (1998)

    Knill, D.C.: Surface orientation from texture: ideal observers, generic observers and the information content of texture cues. Vision research38(11), 1655–1682 (1998)

    1998

  21. [21]

    LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models

    Li, F., Zhang, R., Zhang, H., Zhang, Y., Li, B., Li, W., Ma, Z., Li, C.: Llava-next- interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Ministral 3

    Liu, A.H., Khandelwal, K., Subramanian, S., Jouault, V., Rastogi, A., Sadé, A., Jeffares, A., Jiang, A., Cahill, A., Gavaudan, A., et al.: Ministral 3. arXiv preprint arXiv:2601.08584 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    Transactions of the Association for Computational Linguistics11, 635–651 (2023)

    Liu, F., Emerson, G., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics11, 635–651 (2023)

    2023

  24. [24]

    Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? In: European conference on computer vision. pp. 216–233. Springer (2024)

    2024

  25. [25]

    Journal of Computational Social Science9(1), 11 (2026)

    Lou, J., Sun, Y.: Anchoring bias in large language models: An experimental study. Journal of Computational Social Science9(1), 11 (2026)

    2026

  26. [26]

    In: International Conference on Learning Representations (ICLR) (2024)

    Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi, H., Cheng, H., Chang, K.W., Galley, M., Gao, J.: Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  27. [27]

    https://github.com/huggingface/peft(2022)

    Mangrulkar, S., Gugger, S., Debut, L., Belkada, Y., Paul, S., Bossan, B., Tietz, M.: PEFT: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft(2022)

    2022

  28. [28]

    Meta AI blog post, https : //ai.meta.com/blog/llama-4-multimodal-intelligence/, accessed: 2026-03-04

    Meta: Llama 4 multimodal intelligence. Meta AI blog post, https : //ai.meta.com/blog/llama-4-multimodal-intelligence/, accessed: 2026-03-04

    2026

  29. [29]

    GitHub repository, https://github.com/vikhyat/ moondream, accessed: 2026-03-04

    Moondream: Moondream. GitHub repository, https://github.com/vikhyat/ moondream, accessed: 2026-03-04

    2026

  30. [30]

    Website,https://docs.ollama

    Ollama: Ollama (software and documentation). Website,https://docs.ollama. com/, accessed: 2026-03-04

    2026

  31. [31]

    OpenAI: Chatgpt (march 2026 version) [large language model] (2026), https://chat.openai.com, accessed: 2026-05-11

    2026

  32. [32]

    Vision research43(23), 2451–2468 (2003)

    Oruç, I., Maloney, L.T., Landy, M.S.: Weighted linear cue combination with possibly correlated error. Vision research43(23), 2451–2468 (2003)

    2003

  33. [33]

    Qwen2.5-VL Technical Report

    Qwen Team: Qwen2.5-vl technical report. arXiv (2025).https://doi.org/10. 48550/arXiv.2502.13923

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Qwen Team: Qwen3.5: Towards native multimodal agents (February 2026), https://qwen.ai/blog?id=qwen3.5

    2026

  35. [35]

    Qwen Team: Qwen3.6-35B-A3B: Agentic coding power, now open to all (April 2026), https://qwen.ai/blog?id=qwen3.6-35b-a3b

    2026

  36. [36]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021) 18 Q. Zhang et al

    2021

  37. [37]

    In: Proceedings of the Asian Conference on Computer Vision

    Rahmanzadehgervi, P., Bolton, L., Taesiri, M.R., Nguyen, A.T.: Vision language models are blind. In: Proceedings of the Asian Conference on Computer Vision. pp. 18–34 (2024)

    2024

  38. [38]

    Vision Research44(13), 1511–1535 (2004)

    Rosas, P., Wichmann, F.A., Wagemans, J.: Some observations on the effects of slant and texture type on slant-from-texture. Vision Research44(13), 1511–1535 (2004)

    2004

  39. [39]

    Mechanisms of Prompt-Induced Hallucination in Vision-Language Models

    Rudman, W., Golovanevsky, M., Arad, D., Belinkov, Y., Singh, R., Eickhoff, C., Mahowald, K.: Mechanisms of prompt-induced hallucination in vision-language models. arXiv preprint arXiv:2601.05201 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    In: Findings of the Association for Computational Linguistics: ACL 2025

    Rudman,W.,Golovanevsky,M.,Bar,A.,Palit,V.,LeCun,Y.,Eickhoff,C.,Singh,R.: Forgotten polygons: Multimodal large language models are shape-blind. In: Findings of the Association for Computational Linguistics: ACL 2025. pp. 11983–11998 (2025)

    2025

  41. [41]

    Chameleon: Mixed-Modal Early-Fusion Foundation Models

    Team, C.: Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818 (2024).https://doi.org/10.48550/arXiv.2405.09818, https://github.com/facebookresearch/chameleon

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.09818 2024

  42. [42]

    Journal of vision10(5), 17–17 (2010)

    Todd, J.T., Thaler, L.: The perception of 3d shape from texture based on directional width gradients. Journal of vision10(5), 17–17 (2010)

    2010

  43. [43]

    Vision Research45(12), 1501–1517 (2005)

    Todd, J.T., Thaler, L., Dijkstra, T.M.: The effects of field of view on the perception of 3d slant from texture. Vision Research45(12), 1501–1517 (2005)

    2005

  44. [44]

    Journal of vision7(12), 9–9 (2007)

    Todd, J.T., Thaler, L., Dijkstra, T.M., Koenderink, J.J., Kappers, A.M.: The effects of viewing angle, camera angle, and sign of surface curvature on the perception of three-dimensional shape from texture. Journal of vision7(12), 9–9 (2007)

    2007

  45. [45]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes wide shut? exploring the visual shortcomings of multimodal llms. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9568–9578 (2024)

    2024

  46. [46]

    arXiv preprint arXiv:2510.25776 (2025)

    Tseng, C.Y., Roy, S., Thasin, M., Zhang, D., Effiong, B.: Streetmath: Study of llms’ approximation behaviors. arXiv preprint arXiv:2510.25776 (2025)

    work page arXiv 2025

  47. [47]

    In: ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition

    Verbin, D., Zickler, T.: Toward a universal model for shape from texture. In: ProceedingsoftheIEEE/CVFconferenceoncomputervisionandpatternrecognition. pp. 422–430 (2020)

    2020

  48. [48]

    Vision Language Models are Biased

    Vo, A., Nguyen, K.N., Taesiri, M.R., Dang, V.T., Nguyen, A.T., Kim, D.: Vision language models are biased. arXiv preprint arXiv:2505.23941 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  50. [50]

    ACM Transactions on Applied Perception20(4), 1–18 (2023)

    Wang, Y., Zhang, Q., Aubuchon, C., Kemp, J., Domini, F., Tompkin, J.: On human-like biases in convolutional neural networks for the perception of slant from texture. ACM Transactions on Applied Perception20(4), 1–18 (2023)

    2023

  51. [51]

    Artificial intelligence17(1-3), 17–45 (1981)

    Witkin, A.P.: Recovering surface shape and orientation from texture. Artificial intelligence17(1-3), 17–45 (1981)

    1981

  52. [52]

    Corpus Linguistics and Linguistic Theory20(1), 123–152 (2024)

    Woodin, G., Winter, B., Littlemore, J., Perlman, M., Grieve, J.: Large-scale patterns of number use in spoken and written english. Corpus Linguistics and Linguistic Theory20(1), 123–152 (2024)

    2024

  53. [53]

    Yang,L.,Kang,B.,Huang,Z.,Xu,X.,Feng,J.,Zhao,H.:Depthanything:Unleashing thepoweroflarge-scaleunlabeleddata.In:ProceedingsoftheIEEE/CVFconference on computer vision and pattern recognition. pp. 10371–10381 (2024)

    2024

  54. [54]

    arXiv preprint arXiv:2509.18905 (2025)

    Yu,S.,Chen,Y.,Ju,H.,Jia,L.,Zhang,F.,Huang,S.,Wu,Y.,Cui,R.,Ran,B.,Zhang, Z., et al.: How far are vlms from visual spatial intelligence? a benchmark-driven perspective. arXiv preprint arXiv:2509.18905 (2025)

    work page arXiv 2025

  55. [55]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yue, X., Ni, Y., Zhang, K., Zheng, T., Liu, R., Zhang, G., Stevens, S., Jiang, D., Ren, W.,Sun,Y.,etal.:Mmmu:Amassivemulti-disciplinemultimodalunderstandingand VLMs Fail at Slant-from-Texture Perception 19 reasoning benchmark for expert agi. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9556–9567 (2024)

    2024

  56. [56]

    2022 ieee

    Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. 2022 ieee. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18102–18112 (2021)

    2022

  57. [57]

    IEEE transactions on pattern analysis and machine intelligence46(8), 5625–5644 (2024)

    Zhang, J., Huang, J., Jin, S., Lu, S.: Vision-language models for vision tasks: A survey. IEEE transactions on pattern analysis and machine intelligence46(8), 5625–5644 (2024)

    2024

  58. [58]

    Zhang, R., Jiang, D., Zhang, Y., Lin, H., Guo, Z., Qiu, P., Zhou, A., Lu, P., Chang, K.W., Qiao, Y., et al.: Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In: European Conference on Computer Vision. pp. 169–186. Springer (2024)

    2024

  59. [59]

    EMNLP 2023 (2023)

    Zhang,Y.,Pan,J.,Zhou,Y.,Pan,R.,Chai,J.:Groundingvisualillusionsinlanguage: Do vision-language models perceive illusions like humans? In: Proceedings of Con- ference of Empirical Methods in Natural Language Processing. EMNLP 2023 (2023)

    2023

  60. [60]

    judged slant

    Zhang, Y., Unell, A., Wang, X., Ghosh, D., Su, Y., Schmidt, L., Yeung-Levy, S.: Why are visually-grounded language models bad at image classification? Advances in Neural Information Processing Systems37, 51727–51753 (2024) VLMs Fail at Slant-from-Texture Perception 1 1 Recent models We also found anchoring in 12 frontier VLMs, including closed source ones...

    2024