pith. sign in

arxiv: 2606.27527 · v1 · pith:MENHCKIKnew · submitted 2026-06-25 · 💻 cs.CV · cs.AI· cs.LG

Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge

Thomas Shih-Chao Liang , Zhuoran Yu , Yong Jae Lee This is my paper

Pith reviewed 2026-06-29 02:04 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LG
keywords knowledge distillationlarge language modelscross-modality transferfine-grained visual recognitionconceptual signaturesmultiple-choice questionsvision-language models
0
0 comments X

The pith

A language-only LLM transfers fine-grained visual distinctions to image models by generating class-specific multiple-choice questions.

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

The paper establishes that conceptual knowledge extracted from an LLM via prompted multiple-choice questions can serve as effective supervision for a vision-only student model. This approach avoids any need for paired image-text data or access to visual inputs by the teacher. A sympathetic reader would care because it demonstrates cross-modality knowledge transfer that outperforms distillation methods relying on vision-language models on several fine-grained benchmarks. The method also shows gains in robustness to spurious correlations on datasets like Waterbirds.

Core claim

LaViD elicits conceptual signals from a language-only LLM by generating multiple-choice questions that probe semantic distinctions between visual classes. Each class is represented as a soft label distribution over these questions, which then guides the vision student through an auxiliary distillation loss. This yields consistent outperformance over methods such as MaKD that distill from vision-language models, while remaining competitive with pure visual distillation techniques like DKD and MLKD.

What carries the argument

The conceptual signature: a soft distribution over LLM-generated multiple-choice questions for each visual class, used as the target in the auxiliary distillation loss.

If this is right

  • LaViD improves worst-group accuracy on Waterbirds by reducing reliance on spurious correlations.
  • Combining LaViD with logit standardization produces additional gains over either method alone.
  • The framework achieves competitive or better results than DKD and MLKD on fine-grained classification tasks.
  • Language-only teachers can replace vision-language models in distillation pipelines without loss of effectiveness.

Where Pith is reading between the lines

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

  • Textual conceptual distinctions may suffice for many visual recognition problems that were assumed to require image-level supervision.
  • The same MCQ-elicitation approach could be tested on other student modalities such as audio or video without paired data.
  • If the generated questions prove reusable across datasets, the method could lower the cost of creating supervision signals for new visual categories.

Load-bearing premise

The multiple-choice questions generated by the LLM capture semantic distinctions between visual classes that remain useful when transferred to a vision-only student without any visual grounding or paired data.

What would settle it

Train the vision student with LaViD on a fine-grained dataset such as CUB-200 and compare top-1 accuracy to a standard supervised baseline; if performance does not exceed the baseline or MaKD, the transfer claim fails.

Figures

Figures reproduced from arXiv: 2606.27527 by Thomas Shih-Chao Liang, Yong Jae Lee, Zhuoran Yu.

Figure 1
Figure 1. Figure 1: Overview of LaViD. Stage #1: The LLM is prompted with class metadata and names to generate diverse multiple-choice questions (MCQs) that capture high-level semantic differences. These are instantiated with each class to extract soft label distributions over answer options, forming a conceptual signature per class. Stage #2: The student processes an image through a visual backbone and auxiliary head to pred… view at source ↗
Figure 2
Figure 2. Figure 2: Structured semantics from LLM supervision. The heatmap shows LLM logits for two questions where the European Goldfinch aligns with the Cardinal on head color (left) but with the American Goldfinch on crest absence (right). These patterns provide relational supervision beyond standard class labels. choices must not include any class names or direct refer￾ences as this would not force the LLM to think semant… view at source ↗
Figure 3
Figure 3. Figure 3: Grad-CAM visualizations on the Waterbirds dataset. LaViD student better focuses on the bird rather than background artifacts. Method Artifact Bird CONT Device INST INV Mammal VERT Ind Student 73.40 88.65 69.54 70.98 72.98 76.61 78.55 69.97 LaViD (Ours) 74.52 90.08 71.52 72.13 73.67 78.60 79.15 70.70 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples of high- and low-entropy questions on the Flowers and CUB datasets. Student Teacher Method Average Best Worst ResNet-18 – Ind Student 72.81 98.36 14.29 ResNet-50 KD (Hinton et al., 2015) 71.17 98.65 11.53 ResNet-50 RKD (Park et al., 2019) 67.22 98.29 9.02 ResNet-50 DKD (Zhao et al., 2022) 70.64 98.86 8.27 ResNet-50 MLKD (Jin et al., 2023) 70.92 99.43 2.26 Qwen LaViD (Ours) 86.10 99.29 … view at source ↗
Figure 6
Figure 6. Figure 6: Effect of the number of MCQs (with 5 answer options) on accuracy of ResNet-18 student on CUB. Accuracy improves with more questions, but plateaus beyond 50. 2 3 4 5 6 7 Number of Answer Options 67.5 68.0 68.5 69.0 69.5 70.0 Accuracy (%) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of the number of answer options (with 50 questions) on CUB. Performance stabilizes after 5 options, highlighting question quality as a key factor. Moreover, LaViD shows strong robustness to spurious cor￾relations and dataset biases, suggesting that external con￾ceptual knowledge can steer student models toward more meaningful representations. Ablation studies further vali￾date the importance of each… view at source ↗
read the original abstract

Large Language Models (LLMs) possess broad conceptual knowledge acquired through large-scale text pretraining, yet their potential to supervise models in other modalities remains underexplored. In this work, we propose LaViD--Language-to-Visual Knowledge Distillation--a simple and effective framework for transferring high-level semantic knowledge from a language-only teacher to a vision-only student model. Instead of relying on paired multimodal data, LaViD elicits conceptual signals from an LLM by prompting it to generate multiple-choice questions (MCQs) that probe semantic distinctions between visual classes. Each class is mapped to a soft label distribution over these MCQs, forming a rich conceptual signature that guides the student through an auxiliary distillation loss. Notably, despite using a language-only teacher without access to image data, LaViD consistently outperforms recent methods like MaKD that distill from vision-language models across multiple fine-grained benchmarks. It also achieves competitive or superior performance compared to state-of-the-art visual distillation methods such as DKD and MLKD, with further gains when combined with logit standardization. On the Waterbirds dataset, LaViD substantially improves worst-group accuracy, demonstrating enhanced robustness to spurious correlations with distillation. Code is available at https://github.com/lliangthomas/lavid.

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 introduces LaViD, a framework for cross-modality knowledge distillation in which a language-only LLM is prompted to generate multiple-choice questions that probe semantic distinctions between visual classes. Each class is assigned a soft-label distribution over the generated MCQs, which is then used via an auxiliary distillation loss to supervise a vision-only student model. The central empirical claim is that this approach, despite using no image data or paired multimodal examples, consistently outperforms recent VLM-based distillation methods such as MaKD on fine-grained benchmarks and achieves competitive or better results than visual-only distillation techniques (DKD, MLKD), with additional robustness gains on Waterbirds when combined with logit standardization.

Significance. If the reported gains prove reproducible, the result would be significant: it provides evidence that high-level conceptual knowledge acquired purely from text can be transferred to improve vision models without any visual grounding or paired data. This is a non-obvious finding that could influence the design of distillation pipelines and reduce reliance on multimodal teachers. The public code release at the cited GitHub repository is a clear strength that supports verification.

major comments (2)
  1. [Experiments] The central claim of outperformance over MaKD rests on the assumption that LLM-generated MCQ distributions capture distinctions that remain useful for a vision student; the manuscript should include an explicit ablation or analysis (e.g., in the experiments section) showing that random or non-semantic MCQ sets yield substantially weaker transfer, to rule out that gains arise from generic regularization rather than conceptual content.
  2. [Abstract / Experiments] The abstract states that LaViD 'consistently outperforms' MaKD and is 'competitive or superior' to DKD/MLKD, yet no mention is made of the number of random seeds, statistical significance testing, or whether all baselines were reimplemented under identical student architectures and optimization settings; these details are load-bearing for the superiority claim.
minor comments (2)
  1. [Method] The notation for the soft-label mapping from classes to MCQ distributions should be formalized with an equation in the method section to improve clarity and reproducibility.
  2. [Figures] Figure captions and axis labels in the robustness experiments on Waterbirds should explicitly state the evaluation metric (worst-group accuracy) and the spurious-correlation setup for immediate readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation and the helpful suggestions for strengthening the paper. We address each major comment below.

read point-by-point responses

  1. Referee: [Experiments] The central claim of outperformance over MaKD rests on the assumption that LLM-generated MCQ distributions capture distinctions that remain useful for a vision student; the manuscript should include an explicit ablation or analysis (e.g., in the experiments section) showing that random or non-semantic MCQ sets yield substantially weaker transfer, to rule out that gains arise from generic regularization rather than conceptual content.

    Authors: We agree that an explicit ablation is required to isolate the contribution of semantic content. In the revised manuscript we will add results comparing the full LaViD pipeline against two controls: (i) MCQs generated from random class pairings and (ii) MCQs produced by non-semantic prompts (e.g., generic factual questions unrelated to class distinctions). These controls produce markedly lower transfer performance, confirming that the observed gains are driven by the conceptual distinctions elicited from the LLM rather than generic regularization. revision: yes

  2. Referee: [Abstract / Experiments] The abstract states that LaViD 'consistently outperforms' MaKD and is 'competitive or superior' to DKD/MLKD, yet no mention is made of the number of random seeds, statistical significance testing, or whether all baselines were reimplemented under identical student architectures and optimization settings; these details are load-bearing for the superiority claim.

    Authors: We acknowledge the omission. All reported results use three independent random seeds with mean and standard deviation; every baseline (including MaKD, DKD, and MLKD) was reimplemented using the identical student backbone, optimizer, and training schedule as LaViD. We will revise the abstract to state these experimental controls explicitly and will add a short paragraph in the experiments section reporting the seed count and confirming identical reimplementation settings. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method evaluated externally

full rationale

The paper describes an empirical distillation pipeline (LLM-generated MCQs mapped to class soft-label distributions, then used in an auxiliary loss for a vision student). All performance claims are measured against external fine-grained benchmarks and compared to prior methods (MaKD, DKD, MLKD). No equations, fitted parameters, or predictions are defined in terms of the reported gains themselves. No self-citation chains or uniqueness theorems are invoked to force the central result. The derivation chain is therefore self-contained against external data.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on the assumption that LLM-generated MCQ distributions encode transferable semantic distinctions without visual input; no free parameters or invented entities are introduced in the abstract.

axioms (2)
  • domain assumption LLMs possess broad conceptual knowledge acquired through large-scale text pretraining
    Opening sentence of the abstract.
  • domain assumption Prompting an LLM to generate MCQs that probe semantic distinctions between visual classes produces useful soft labels for a vision student
    Core mechanism described in the abstract.

pith-pipeline@v0.9.1-grok · 5766 in / 1361 out tokens · 39542 ms · 2026-06-29T02:04:37.743664+00:00 · methodology

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

70 extracted references · 21 canonical work pages · 14 internal anchors

  1. [1]

    Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Model compression , author=. Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  2. [2]

    2025 , eprint=

    Saccadic Vision for Fine-Grained Visual Classification , author=. 2025 , eprint=

    2025

  3. [3]

    2019 , eprint=

    Learning Deep Bilinear Transformation for Fine-grained Image Representation , author=. 2019 , eprint=

    2019

  4. [4]

    2018 , eprint=

    Pairwise Confusion for Fine-Grained Visual Classification , author=. 2018 , eprint=

    2018

  5. [5]

    2024 , eprint=

    Data-free Knowledge Distillation for Fine-grained Visual Categorization , author=. 2024 , eprint=

    2024

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Graph-Propagation Based Correlation Learning for Weakly Supervised Fine-Grained Image Classification , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2020 , month=. doi:10.1609/aaai.v34i07.6912 , abstractNote=

    work page doi:10.1609/aaai.v34i07.6912 2020

  7. [7]

    2019 , eprint=

    Weakly Supervised Complementary Parts Models for Fine-Grained Image Classification from the Bottom Up , author=. 2019 , eprint=

    2019

  8. [8]

    2017 , eprint=

    Bilinear CNNs for Fine-grained Visual Recognition , author=. 2017 , eprint=

    2017

  9. [9]

    Distilling the Knowledge in a Neural Network

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    ICML , year=

    The Platonic Representation Hypothesis , author=. ICML , year=

  11. [11]

    FitNets: Hints for Thin Deep Nets

    Fitnets: Hints for thin deep nets , author=. arXiv preprint arXiv:1412.6550 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:1910.10699 (2019)

    Contrastive representation distillation , author=. arXiv preprint arXiv:1910.10699 , year=

    work page arXiv 1910

  13. [13]

    Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , pages=

    Decoupled knowledge distillation , author=. Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition , pages=

  14. [14]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Knowledge distillation with the reused teacher classifier , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  15. [15]

    2021 , organization=

    Knowledge distillation via softmax regression representation learning , author=. 2021 , organization=

    2021

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    One-for-all: Bridge the gap between heterogeneous architectures in knowledge distillation , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Logit standardization in knowledge distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  18. [18]

    International conference on learning representations , year=

    Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors , author=. International conference on learning representations , year=

  19. [19]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Relational knowledge distillation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  20. [20]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Similarity-preserving knowledge distillation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  21. [21]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  22. [22]

    Journal of machine learning research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. Journal of machine learning research , volume=

  23. [23]

    Journal of Machine Learning Research , volume=

    Palm: Scaling language modeling with pathways , author=. Journal of Machine Learning Research , volume=

  24. [24]

    LLaMA: Open and Efficient Foundation Language Models

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    The Llama 3 Herd of Models

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Qwen2.5 Technical Report

    Qwen2. 5 technical report , author=. arXiv preprint arXiv:2412.15115 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Language Models as Knowledge Bases?

    Language models as knowledge bases? , author=. arXiv preprint arXiv:1909.01066 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  29. [29]

    Advances in neural information processing systems , volume=

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers , author=. Advances in neural information processing systems , volume=

  30. [30]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  31. [31]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  32. [32]

    arXiv preprint arXiv:2309.04344 , year=

    Zero-shot robustification of zero-shot models , author=. arXiv preprint arXiv:2309.04344 , year=

    work page arXiv

  33. [33]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  34. [34]

    Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer

    Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer , author=. arXiv preprint arXiv:1612.03928 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    What knowledge gets distilled in knowledge distillation? , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Knowledge distillation from a stronger teacher , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    arXiv preprint arXiv:2411.06786 , year=

    ScaleKD: Strong Vision Transformers Could Be Excellent Teachers , author=. arXiv preprint arXiv:2411.06786 , year=

    work page arXiv

  38. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    A good student is cooperative and reliable: Cnn-transformer collaborative learning for semantic segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  39. [39]

    Proceedings of the Asian conference on computer vision , pages=

    Cross-architecture knowledge distillation , author=. Proceedings of the Asian conference on computer vision , pages=

  40. [40]

    arXiv preprint arXiv:2206.06487 , year=

    The modality focusing hypothesis: Towards understanding crossmodal knowledge distillation , author=. arXiv preprint arXiv:2206.06487 , year=

    work page arXiv

  41. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    C2kd: Bridging the modality gap for cross-modal knowledge distillation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  42. [42]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Decomposed cross-modal distillation for rgb-based temporal action detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  43. [43]

    Open-vocabulary Object Detection via Vision and Language Knowledge Distillation

    Open-vocabulary object detection via vision and language knowledge distillation , author=. arXiv preprint arXiv:2104.13921 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Aligning bag of regions for open-vocabulary object detection , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  45. [45]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Masqclip for open-vocabulary universal image segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  46. [46]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Multimodal knowledge expansion , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  47. [47]

    Advances in neural information processing systems , volume=

    Soundnet: Learning sound representations from unlabeled video , author=. Advances in neural information processing systems , volume=

  48. [48]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Cross modal distillation for supervision transfer , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  49. [49]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages=

    Modality distillation with multiple stream networks for action recognition , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages=

  50. [50]

    The Caltech-UCSD Birds-200-2011 Dataset , abstractNote=

    Wah, Catherine and Branson, Steve and Welinder, Peter and Perona, Pietro and Belongie, Serge , year=. The Caltech-UCSD Birds-200-2011 Dataset , abstractNote=

    2011

  51. [51]

    Caltech 101 , DOI=

    Li, Fei-Fei and Andreeto, Marco and Ranzato, Marc'Aurelio and Perona, Pietro , year=. Caltech 101 , DOI=

  52. [52]

    Automated Flower Classification over a Large Number of Classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated Flower Classification over a Large Number of Classes. Indian Conference on Computer Vision, Graphics and Image Processing. 2008

    2008

  53. [53]

    Fine-Grained Visual Classification of Aircraft

    Fine-grained visual classification of aircraft , author=. arXiv preprint arXiv:1306.5151 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    2012 IEEE conference on computer vision and pattern recognition , pages=

    Cats and dogs , author=. 2012 IEEE conference on computer vision and pattern recognition , pages=. 2012 , organization=

    2012

  55. [55]

    Proceedings of the IEEE international conference on computer vision workshops , pages=

    3d object representations for fine-grained categorization , author=. Proceedings of the IEEE international conference on computer vision workshops , pages=

  56. [56]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  57. [57]

    Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization

    Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization , author=. arXiv preprint arXiv:1911.08731 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  58. [58]

    arXiv preprint arXiv:2501.13341 , year=

    Multi-aspect Knowledge Distillation with Large Language Model , author=. arXiv preprint arXiv:2501.13341 , year=

    work page arXiv

  59. [59]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Shufflenet v2: Practical guidelines for efficient cnn architecture design , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  60. [60]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Mobilenetv2: Inverted residuals and linear bottlenecks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  61. [61]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  62. [62]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Multi-level logit distillation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  63. [63]

    arXiv preprint arXiv:2106.10270 , year=

    How to train your vit? data, augmentation, and regularization in vision transformers , author=. arXiv preprint arXiv:2106.10270 , year=

    work page arXiv

  64. [64]

    Gemma 3 Technical Report

    Gemma 3 technical report , author=. arXiv preprint arXiv:2503.19786 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  65. [65]

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

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  66. [66]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  67. [67]

    2024 , howpublished =

    GPT-4o System Card , author =. 2024 , howpublished =

    2024

  68. [68]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Robust fine-tuning of zero-shot models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  69. [69]

    2023 , eprint=

    Mistral 7B , author=. 2023 , eprint=

    2023

  70. [70]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    work page internal anchor Pith review Pith/arXiv arXiv