Pith. sign in

REVIEW 5 minor 75 references

Averaging over many random token orders lets autoregressive models beat diffusion as generative image classifiers, while running far faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 22:09 UTC pith:C5KFI37O

load-bearing objection Order-marginalized any-order AR classifiers beat diffusion GCs and approach DINOv2, with large efficiency gains and solid ablations.

arxiv 2603.19122 v2 pith:C5KFI37O submitted 2026-03-19 cs.CV

Revisiting Autoregressive Models for Generative Image Classification

Ilia Sudakov , Artem Babenko , Dmitry Baranchuk This is my paper
classification cs.CV
keywords generative classifiersautoregressive modelsany-order ARtoken orderorder marginalizationdiffusion classifiersImageNet robustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Class-conditional generative models can act as classifiers by turning likelihoods into class posteriors via Bayes' rule. Until now, diffusion models have dominated this setting; fixed-order autoregressive models lagged because each single left-to-right (or other fixed) scan only sees partial cues. This paper shows that any-order autoregressive models change the picture: by sampling many random token permutations and averaging the resulting log-likelihoods (the Jensen lower bound on the order-marginal likelihood), the classifier obtains a more complete, shape-aware signal. The resulting order-marginalized AR classifier consistently outperforms strong diffusion classifiers on ImageNet and several out-of-distribution shifts, while needing far fewer network evaluations. It also reaches accuracy competitive with leading self-supervised discriminative models—an unusual result for a purely generative approach.

Core claim

The restrictive inductive bias of a fixed token order is the main reason prior autoregressive generative classifiers underperformed diffusion models. Marginalizing over many random orders with an any-order AR model removes that bias and unlocks higher, more robust classification accuracy at much lower compute cost.

What carries the argument

Order-marginalized log-likelihood (Eq. 5): the Monte-Carlo average of order-conditional log-likelihoods E_π[log p(x|π,c)] estimated with K random permutations of the image tokens; this lower bound, not the direct log-E estimate, is used as the class score.

Load-bearing premise

That averaging only about twenty random token orders is already a faithful enough stand-in for the true order-marginal likelihood, with no systematic bias under distribution shift.

What would settle it

Train an identical any-order AR model, then re-evaluate ImageNet-Val and ImageNet-A/R/S accuracy while sweeping K from 1 to several hundred; if accuracy plateaus or declines well before K equals 20, or if the lower-bound estimator systematically prefers the wrong class under controlled token-order noise, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

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

0 major / 5 minor

Summary. The paper revisits class-conditional autoregressive models as generative classifiers and argues that fixed token order is a restrictive inductive bias. Using RandAR (any-order AR), it estimates order-marginalized class-conditional log-likelihoods via the Jensen lower bound E_π[log p(x|π,c)] (Eq. 5) with K random permutations, then applies Bayes’ rule. Controlled experiments on ImageNet-1K (same resolution, token count, transformer family) show that order-marginalized RandAR outperforms fixed-order AR baselines (LlamaGen, VAR, raster RandAR), diffusion classifiers (DiT, SiT, REPA), and the joint model DAT on in-domain and OOD sets (IN-R/S/A/C), while remaining up to 25 imes faster than DiT. It is competitive with DINOv2 linear probes and shows strong robustness on WILDS (Camelyon17, CelebA, FMoW). Ablations cover K, lower-bound vs. log-E estimation, tokenizer choice, latent noise augmentation, and ELBO weightings for diffusion baselines.

Significance. If the empirical claims hold, the work re-opens autoregressive models as a competitive and far more efficient alternative to diffusion classifiers, and shows that generative classifiers can approach state-of-the-art self-supervised discriminative performance (DINOv2). The central technical insight—that marginalizing over token orders removes a restrictive inductive bias—is cleanly demonstrated by per-token likelihood and accuracy analyses (Figs. 3–5) and by the large gap between K=1 and K≥5. Strengths include extensive controlled comparisons, public-code promise, and ablations of the likelihood estimator and tokenizer. The result is of clear interest to the generative-modeling and robust-classification communities.

minor comments (5)
  1. Table 2 and §6.1: results for 1000-class sets (IN-Val, IN-S, IN-C) are reported on 2K subsets following prior work; a short note confirming that full-set numbers (or larger subsets) do not change rankings would strengthen the claim.
  2. No error bars or multi-seed statistics are given for the main ImageNet tables; even a single additional seed for RandAR-L/XL would help quantify variance.
  3. Fig. 6 efficiency curves: clarify whether wall-clock times include VQ-VAE encoding and whether batching across classes is used; this affects the claimed 25 imes factor.
  4. Eq. (5) and Table 5: the superiority of the Jensen lower bound is well ablated, but a one-sentence discussion of possible residual Monte-Carlo bias under strong distribution shift would be useful for readers.
  5. Minor typos: “seacoast binoculars” caption in Fig. 1; “Сn” vs. “Cn” in Fig. 2; occasional missing spaces after periods.

Circularity Check

0 steps flagged

No significant circularity; the paper is a self-contained empirical comparison of order-marginalized AR likelihoods against external public baselines.

full rationale

The central claim is empirical superiority of Monte-Carlo order-marginalized log-likelihoods (Eq. 5, Jensen lower bound estimated with K random permutations of RandAR) over diffusion ELBO classifiers and competitive performance versus DINOv2/ViT on ImageNet and OOD sets. Likelihood estimation follows the standard generative-classifier recipe (Bayes + class-conditional log p(x|c)); the lower-bound choice is justified by training alignment and directly ablated against the direct log-E estimator (Table 5). All baselines (DiT, SiT, LlamaGen, VAR, A-VARC, DINOv2, ResNets, WILDS ERM/RWY) are either public checkpoints or re-trained from official code under matched architectures/resolutions; no parameters are fitted on the evaluation sets and then re-used as “predictions.” Self-citations are limited to the RandAR architecture paper and ordinary related-work references; none supply a uniqueness theorem or load-bearing premise that forces the reported accuracy numbers. Residual Monte-Carlo bias of the K≈20 estimator is acknowledged and measured but does not reduce the accuracy claims to a definitional identity. Consequently the derivation chain contains no self-definitional steps, fitted-input-as-prediction, or self-citation circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The work is empirical; it inherits standard likelihood-based generative classification (Bayes rule + class-conditional density estimation) and the RandAR architecture. The only non-standard modeling choice is the use of the Jensen lower bound on the order-marginal likelihood and the Monte-Carlo sample size K.

free parameters (2)
  • K (number of random orders) = 20 (main tables)
    Monte-Carlo sample size for order marginalization; chosen by validation (K=20 near-optimal). Directly controls both accuracy and compute.
  • latent noise augmentation strength = ~27% flipped tokens
    Fraction of tokens flipped via flow-matching noise during training with LlamaGen tokenizer; selected by reconstruction-FID and classification ablations.
axioms (3)
  • domain assumption Class-conditional generative models yield valid posteriors via Bayes rule under a uniform class prior.
    Standard premise of all generative classifiers (Eq. 1); inherited from prior DC/AR literature.
  • ad hoc to paper The Jensen lower bound E[log p(x|π,c)] is a better classification objective than the true log-E for models trained with that bound.
    Justified by training alignment and Table 5 ablation; not derived from first principles.
  • domain assumption RandAR’s position-instruction tokens correctly realize arbitrary-order conditionals p(x|π,c).
    Taken from the RandAR paper; verified only by generation quality, not by formal likelihood correctness.

pith-pipeline@v1.1.0-grok45 · 21832 in / 2289 out tokens · 22882 ms · 2026-07-13T22:09:51.376636+00:00 · methodology

0 comments
read the original abstract

Class-conditional generative models have emerged as accurate and robust classifiers, with diffusion models demonstrating clear advantages over other visual generative paradigms, including autoregressive (AR) models. In this work, we revisit visual AR-based generative classifiers and identify an important limitation of prior approaches: their reliance on a fixed token order, which imposes a restrictive inductive bias for image understanding. We observe that single-order predictions rely more on partial discriminative cues, while averaging over multiple token orders provides a more comprehensive signal. Based on this insight, we leverage recent any-order AR models to estimate order-marginalized predictions, unlocking the high classification potential of AR models. Our approach consistently outperforms diffusion-based classifiers across diverse image classification benchmarks, while being up to 25x more efficient. Compared to state-of-the-art self-supervised discriminative models, our method delivers competitive classification performance - a notable achievement for generative classifiers.

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

75 extracted references · 24 linked inside Pith

  1. [1]

    IEEE transactions on medical imaging38(2), 550–560 (2018)

    Bandi, P., Geessink, O., Manson, Q., Van Dijk, M., Balkenhol, M., Hermsen, M., Bejnordi, B.E., Lee, B., Paeng, K., Zhong, A., et al.: From detection of individual metastases to classification of lymph node status at the patient level: the came- lyon17 challenge. IEEE transactions on medical imaging38(2), 550–560 (2018)

    2018

  2. [2]

    arXiv preprint arXiv:2112.03126 (2021)

    Baranchuk, D., Rubachev, I., Voynov, A., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126 (2021)

    Pith/arXiv arXiv 2021

  3. [3]

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

    Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked genera- tive image transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11315–11325 (2022)

    2022

  4. [4]

    Advances in Neural Information Processing Systems37, 50062–50097 (2024)

    Chen, H., Dong, Y., Shao, S., Zhongkai, H., Yang, X., Su, H., Zhu, J.: Diffusion models are certifiably robust classifiers. Advances in Neural Information Processing Systems37, 50062–50097 (2024)

    2024

  5. [5]

    arXiv preprint arXiv:2305.15241 (2023)

    Chen, H., Dong, Y., Wang, Z., Yang, X., Duan, C., Su, H., Zhu, J.: Robust classi- fication via a single diffusion model. arXiv preprint arXiv:2305.15241 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    arXiv preprint arXiv:2401.14404 (2024)

    Chen, X., Liu, Z., Xie, S., He, K.: Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    In: Proceedings of the IEEE/CVF international conference on com- puter vision

    Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on com- puter vision. pp. 9640–9649 (2021)

    2021

  8. [8]

    arXiv preprint arXiv:2510.12060 (2025)

    Chen, Y.C., Inouye, D.I., Gao, J.: Your var model is secretly an efficient and explainable generative classifier. arXiv preprint arXiv:2510.12060 (2025)

    arXiv 2025

  9. [9]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition

    Christie, G., Fendley, N., Wilson, J., Mukherjee, R.: Functional map of the world. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recog- nition. pp. 6172–6180 (2018)

    2018

  10. [10]

    Dieleman, S.: Diffusion is spectral autoregression (2024),https://sander.ai/ 2024/09/02/spectral-autoregression.html

    2024

  11. [11]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  12. [12]

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

    Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873–12883 (2021)

    2021

  13. [13]

    Communications of the ACM63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM63(11), 139–144 (2020)

    2020

  14. [14]

    arXiv preprint arXiv:1912.03263 (2019)

    Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263 (2019)

    Pith/arXiv arXiv 1912

  15. [15]

    arXiv preprint arXiv:2304.02012 (2023)

    Guo, Q., Ma, C., Jiang, Y., Yuan, Z., Yu, Y., Luo, P.: Egc: Image generation and classification via a diffusion energy-based model. arXiv preprint arXiv:2304.02012 (2023)

    Pith/arXiv arXiv 2023

  16. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Hang, T., Gu, S., Li, C., Bao, J., Chen, D., Hu, H., Geng, X., Guo, B.: Efficient diffusion training via min-snr weighting strategy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7441–7451 (October 2023)

    2023

  17. [17]

    In: Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recogni- tion. In: Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. pp. 770–778. CVPR ’16, IEEE (Jun 2016) 16 I. Sudakov et al

    2016

  18. [18]

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

    Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8340–8349 (2021)

    2021

  19. [19]

    arXiv preprint arXiv:1903.12261 (2019)

    Hendrycks,D.,Dietterich,T.:Benchmarkingneuralnetworkrobustnesstocommon corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)

    Pith/arXiv arXiv 1903

  20. [20]

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

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D.: Natural adversarial examples. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15262–15271 (2021)

    2021

  21. [21]

    Advances in neural information processing systems30(2017)

    Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems30(2017)

    2017

  22. [22]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  23. [23]

    In: Conference on Causal Learning and Reasoning

    Idrissi, B.Y., Arjovsky, M., Pezeshki, M., Lopez-Paz, D.: Simple data balancing achieves competitive worst-group-accuracy. In: Conference on Causal Learning and Reasoning. pp. 336–351. PMLR (2022)

    2022

  24. [24]

    arXiv preprint arXiv:2309.16779 (2023)

    Jaini, P., Clark, K., Geirhos, R.: Intriguing properties of generative classifiers. arXiv preprint arXiv:2309.16779 (2023)

    Pith/arXiv arXiv 2023

  25. [25]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Kasa, S.R., Gupta, K., Roychowdhury, S., Kumar, A., Biruduraju, Y., Kasa, S.K., Priyatam, P.N., Bhattacharya, A., Agarwal, S., Huddar, V.: Generative or discrim- inative? revisiting text classification in the era of transformers. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 9615–9637 (2025)

    2025

  26. [26]

    Kim, D., Thomas, X., Ghadiyaram, D.: Revelio: Interpreting and leveraging seman- ticinformationindiffusionmodels.In:ProceedingsoftheIEEE/CVFInternational Conference on Computer Vision. pp. 4659–4669 (2025)

    2025

  27. [27]

    Advances in neural information processing systems34, 21696–21707 (2021)

    Kingma, D., Salimans, T., Poole, B., Ho, J.: Variational diffusion models. Advances in neural information processing systems34, 21696–21707 (2021)

    2021

  28. [28]

    In: Thirty-seventh Conference on Neural Information Processing Systems (2023),https://openreview.net/forum?id=NnMEadcdyD

    Kingma, D.P., Gao, R.: Understanding diffusion objectives as the ELBO with simple data augmentation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023),https://openreview.net/forum?id=NnMEadcdyD

    2023

  29. [29]

    arXiv preprint arXiv:1312.6114 (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    Pith/arXiv arXiv 2013

  30. [30]

    In: International conference on machine learning

    Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in- the-wild distribution shifts. In: International conference on machine learning. pp. 5637–5664. PMLR (2021)

    2021

  31. [31]

    arXiv preprint arXiv:2510.21890 (2025)

    Lai, C.H., Song, Y., Kim, D., Mitsufuji, Y., Ermon, S.: The principles of diffusion models. arXiv preprint arXiv:2510.21890 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    In: Proceedings of the fourteenth international conference on artificial intelligence and statistics

    Larochelle, H., Murray, I.: The neural autoregressive distribution estimator. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 29–37. JMLR Workshop and Conference Proceedings (2011)

    2011

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2206–2217 (2023)

    2023

  34. [34]

    In: The Thirteenth International Conference on Learning Representations (2024)

    Li, A.C., Kumar, A., Pathak, D.: Generative classifiers avoid shortcut solutions. In: The Thirteenth International Conference on Learning Representations (2024)

    2024

  35. [35]

    arXiv preprint arXiv:2503.10568 (2025) Revisiting AR Models for Generative Image Classification 17

    Li, H., Yang, J., Li, G., Wang, H.: Autoregressive image generation with random- ized parallel decoding. arXiv preprint arXiv:2503.10568 (2025) Revisiting AR Models for Generative Image Classification 17

    arXiv 2025

  36. [36]

    arXiv preprint arXiv:2502.17437 (2025)

    Li, T., Sun, Q., Fan, L., He, K.: Fractal generative models. arXiv preprint arXiv:2502.17437 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Advances in Neural Information Processing Systems37, 56424–56445 (2024)

    Li, T., Tian, Y., Li, H., Deng, M., He, K.: Autoregressive image generation with- out vector quantization. Advances in Neural Information Processing Systems37, 56424–56445 (2024)

    2024

  38. [38]

    arXiv preprint arXiv:2210.02747 (2022)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

    Pith/arXiv arXiv 2022

  39. [39]

    In: Proceedings of the IEEE international conference on computer vision

    Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730– 3738 (2015)

    2015

  40. [40]

    arXiv preprint arXiv:2409.04410 (2024)

    Luo, Z., Shi, F., Ge, Y., Yang, Y., Wang, L., Shan, Y.: Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    In: European Conference on Computer Vision

    Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: European Conference on Computer Vision. pp. 23–40. Springer (2024)

    2024

  42. [42]

    Advances in neural informa- tion processing systems29(2016)

    Van den Oord, A., Kalchbrenner, N., Espeholt, L., Vinyals, O., Graves, A., et al.: Conditional image generation with pixelcnn decoders. Advances in neural informa- tion processing systems29(2016)

    2016

  43. [43]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    Pith/arXiv arXiv 2023

  44. [44]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Pang, Z., Zhang, T., Luan, F., Man, Y., Tan, H., Zhang, K., Freeman, W.T., Wang, Y.X.: Randar: Decoder-only autoregressive visual generation in random orders. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 45– 55 (2025)

    2025

  45. [45]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    2023

  46. [46]

    Advances in Neural Information Processing Systems36, 71874–71910 (2023)

    Puli, A., Zhang, L., Wald, Y., Ranganath, R.: Don’t blame dataset shift! short- cut learning due to gradients and cross entropy. Advances in Neural Information Processing Systems36, 71874–71910 (2023)

    2023

  47. [47]

    Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving lan- guage understanding by generative pre-training (2018)

    2018

  48. [48]

    arXiv preprint arXiv:2502.20388 (2025)

    Ren, S., Yu, Q., He, J., Shen, X., Yuille, A., Chen, L.C.: Beyond next-token: Next- x prediction for autoregressive visual generation. arXiv preprint arXiv:2502.20388 (2025)

    Pith/arXiv arXiv 2025

  49. [49]

    In: International Conference on Learning Representations (ICLR) (2023)

    Rissanen, S., Heinonen, M., Solin, A.: Generative modelling with inverse heat dis- sipation. In: International Conference on Learning Representations (ICLR) (2023)

    2023

  50. [50]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684– 10695 (June 2022)

    2022

  51. [51]

    International journal of computer vision115(3), 211–252 (2015)

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recog- nition challenge. International journal of computer vision115(3), 211–252 (2015)

    2015

  52. [52]

    arXiv preprint arXiv:1911.08731 (2019) 18 I

    Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neu- ral networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) 18 I. Sudakov et al

    Pith/arXiv arXiv 1911

  53. [53]

    Singh, J., Leng, X., Wu, Z., Zheng, L., Zhang, R., Shechtman, E., Xie, S.: What matters for representation alignment: Global information or spatial structure? arXiv preprint arXiv:2512.10794 (2025)

    arXiv 2025

  54. [54]

    arXiv preprint arXiv:2011.13456 (2020)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

    Pith/arXiv arXiv 2011

  55. [55]

    arXiv preprint arXiv:2406.06525 (2024)

    Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., Yuan, Z.: Autoregres- sive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525 (2024)

    Pith/arXiv arXiv 2024

  56. [56]

    Tian, K., Jiang, Y., Yuan, Z., Peng, B., Wang, L.: Visual autoregressive modeling: Scalableimagegenerationvianext-scaleprediction.Advancesinneuralinformation processing systems37, 84839–84865 (2024)

    2024

  57. [57]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    Pith/arXiv arXiv 2023

  58. [58]

    In: International conference on machine learning

    Van Den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. In: International conference on machine learning. pp. 1747–1756. PMLR (2016)

    2016

  59. [59]

    Advances in neural information processing systems30(2017)

    Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems30(2017)

    2017

  60. [60]

    Springer science & business media (2013)

    Vapnik, V.: The nature of statistical learning theory. Springer science & business media (2013)

    2013

  61. [61]

    arXiv preprint arXiv:2412.01819 (2024)

    Voronov,A.,Kuznedelev,D.,Khoroshikh,M.,Khrulkov,V.,Baranchuk,D.:Switti: Designing scale-wise transformers for text-to-image synthesis. arXiv preprint arXiv:2412.01819 (2024)

    Pith/arXiv arXiv 2024

  62. [62]

    Advances in neural information processing systems32(2019)

    Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in neural information processing systems32(2019)

    2019

  63. [63]

    arXiv preprint arXiv:2508.11330 (2025)

    Wang, Y., Chen, L.: Noise matters: Optimizing matching noise for diffusion clas- sifiers. arXiv preprint arXiv:2508.11330 (2025)

    arXiv 2025

  64. [64]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Y., Ren, S., Lin, Z., Han, Y., Guo, H., Yang, Z., Zou, D., Feng, J., Liu, X.: Parallelized autoregressive visual generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 12955–12965 (2025)

    2025

  65. [65]

    Wu, G., Zhang, S., Shi, R., Gao, S., Chen, Z., Wang, L., Chen, Z., Gao, H., Tang, Y., Yang, J., et al.: Representation entanglement for generation: Training diffusion transformersismucheasierthanyouthink.arXivpreprintarXiv:2507.01467(2025)

    arXiv 2025

  66. [66]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yang, X., Wang, X.: Diffusion model as representation learner. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18938–18949 (2023)

    2023

  67. [67]

    arXiv preprint arXiv:2305.10435 (2023)

    Yenduri, G., Srivastava, G., Maddikunta, P.K.R., Jhaveri, R.H., Wang, W., Vasi- lakos, A.V., Gadekallu, T.R., et al.: Generative pre-trained transformer: A com- prehensive review on enabling technologies, potential applications, emerging chal- lenges, and future directions. arXiv preprint arXiv:2305.10435 (2023)

    Pith/arXiv arXiv 2023

  68. [68]

    arXiv preprint arXiv:2510.13872 (2025)

    Yin, X., Zhang, C., Steele, J., Shavit, N., Wang, T.T.: Joint discriminative- generative modeling via dual adversarial training. arXiv preprint arXiv:2510.13872 (2025)

    arXiv 2025

  69. [69]

    arXiv preprint arXiv:2110.04627 (2021) Revisiting AR Models for Generative Image Classification 19

    Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627 (2021) Revisiting AR Models for Generative Image Classification 19

    Pith/arXiv arXiv 2021

  70. [70]

    Yu, J., Xu, Y., Koh, J.Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.K., Hutchinson, B., Han, W., Parekh, Z., Li, X., Zhang, H., Baldridge, J., Wu, Y.: Scaling autoregressive models for content-rich text-to-image generation (2022),https://arxiv.org/abs/2206.10789

    Pith/arXiv arXiv 2022

  71. [71]

    In: The Twelfth International Conference on Learning Representations (2024),https://openreview.net/forum?id=gzqrANCF4g

    Yu, L., Lezama, J., Gundavarapu, N.B., Versari, L., Sohn, K., Minnen, D., Cheng, Y., Gupta, A., Gu, X., Hauptmann, A.G., Gong, B., Yang, M.H., Essa, I., Ross, D.A., Jiang, L.: Language model beats diffusion - tokenizer is key to visual gen- eration. In: The Twelfth International Conference on Learning Representations (2024),https://openreview.net/forum?id...

    2024

  72. [72]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yu, Q., He, J., Deng, X., Shen, X., Chen, L.C.: Randomized autoregressive vi- sual generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18431–18441 (2025)

    2025

  73. [73]

    Advances in Neural Information Processing Systems37, 128940–128966 (2024)

    Yu, Q., Weber, M., Deng, X., Shen, X., Cremers, D., Chen, L.C.: An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems37, 128940–128966 (2024)

    2024

  74. [74]

    arXiv preprint arXiv:2410.06940 (2024)

    Yu, S., Kwak, S., Jang, H., Jeong, J., Huang, J., Shin, J., Xie, S.: Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940 (2024)

    Pith/arXiv arXiv 2024

  75. [75]

    discriminative

    Yue, X., Wang, Z., Wang, Y., Zhang, W., Liu, X., Ouyang, W., Bai, L., Zhou, L.: Understand before you generate: Self-guided training for autoregressive image generation. arXiv preprint arXiv:2509.15185 (2025) Revisiting AR Models for Generative Image Classification 1 A Model settings ImageNet – RandAR: In our main experiments, we train RandAR from scratch...

    arXiv 2025