Pith. sign in

REVIEW 3 major objections 6 minor 4 cited by

HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read A transformer pretrained with a joint embedding predictive architecture on 100 million simulated jets learns transferable jet representations that boost few-shot classification accuracy by four to six points and transfer to top tagging…

desk verdict A genuine first JetClass-scale JEPA pretraining for jets, with a promising few-shot signal that is currently supported only by single-run validation peaks, so treat the headline gain as provisional. read the letter →

arxiv 2502.03933 v1 pith:XPC7UWH2 submitted 2025-02-06 cs.LG hep-exhep-ph

classification cs.LGhep-exhep-ph
keywords foundationmodeljettaggingself-supervisedlearningjointembeddingpredictivearchitecturetransformercolliderphysicsfew-shotquark-gluon
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a transformer pretrained with a joint embedding predictive architecture (JEPA) on 100 million simulated particle jets learns a reusable representation for collider physics. The core evidence is few-shot learning: with only 0.05% of JetClass labels (5,000 samples) the pretrained model reaches 56.4% macro accuracy versus 50.5% for the same architecture trained from scratch, and at 0.5% labels it reaches 62.4% versus 58.6%. The same backbone transfers to top tagging and quark-gluon tagging, roughly matching from-scratch supervised training on those datasets. The paper's motivating idea is that predicting missing jet constituents in latent space, without reconstructing the input, captures abstract physical structure that is cheaper and more transferable than pixel-level or point-level reconstruction.

What carries the argument

The central mechanism is prediction in embedding space. A jet is tokenized by sampling center particles via farthest point sampling, grouping k nearest neighbors in the (η, φ) plane, and encoding each group with a small PointNet-style MLP+max-pool network to produce patch tokens. The token sequence is split into context and target blocks; a transformer context encoder and an EMA teacher target encoder produce embeddings, and a predictor network, conditioned on mask tokens, predicts target embeddings from context embeddings. The training loss is smooth L1 between predicted and teacher target embeddings; no decoder reconstructs the input particles. Two additions matter: a physics bias matrix in the attention (pairwise token interactions computed from group four-momenta) and register tokens, each contributing roughly 2% accuracy in ablations.

What would settle it

Re-running the 0.05%-label JetClass experiment with many seeds and comparing accuracy distributions would settle it: the paper's claim predicts a persistent >4-point gap, and overlapping error bars would refute it.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a Joint Embedding Predictive Architecture can be instantiated for collider jets and that its pretrained embeddings transfer to classification tasks. Pre-training on the 100M-jet JetClass dataset with a JEPA objective—encoding visible token patches and predicting the embeddings of contiguous masked target patches, with an EMA teacher providing targets—produces a backbone that, after fine-tuning on 5,000 labels, achieves 56.4% accuracy (vs. 50.5% from scratch) and, after 50,000 labels, 62.4% (vs. 58.6%). On top tagging (TQTR, 100% labels) the fine-tuned pretrained model reaches 92.9% accuracy, slightly below a supervised JetClass model fine-tuned on the same data (93.8%) and below dedicated taggers ParticleNet (94.0%) and ParT (94.4%); on quark-gluon tagging the frozen pretrained model reaches 82.1%, comparable to a supervised frozen model (82.3%) and from-scratch (81.9%). The paper interprets the pattern as showing that self-supervised JEPA pretraining is most valuable when labels are scarce.

Load-bearing premise

The load-bearing premise is that JEPA pretraining on masked jet patches captures physically meaningful structure that transfers to classification, and that the few-shot accuracy gaps are real rather than run-to-run noise, which the paper does not quantify.

Editorial extensions

If this is right

  • Few-shot jet classification becomes practical: 5,000 labeled jets suffice for meaningful accuracy, which could accelerate analyses with rare or expensively labeled signals.
  • A single pretrained backbone can serve multiple downstream tasks; the same model is evaluated for jet classification, top tagging, and quark-gluon discrimination without per-task pretraining.
  • Because the objective avoids input-space reconstruction, the approach may be more efficient than masked or generative pretraining and may focus computation on semantics rather than detail.
  • The gains concentrate in the low-label regime; with full labels the pretrained and from-scratch models converge (68.5% vs 68.5% at 100% JetClass), suggesting self-supervised pretraining here is not a substitute for task-specific supervised training at scale.
  • The ablations identify concrete design choices—contiguous one-target masking, physics bias, registers—that future jet pretraining models can adopt.

Reading between the lines

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

  • Because the few-shot numbers come from a single run without reported uncertainties, a natural extension is to repeat the 0.05% and 0.5% experiments over many seeds; the paper's thesis predicts a consistent gap, while overlapping error bars would undercut it.
  • The EMA-teacher JEPA objective may be more robust to detector simulation mismatches than reconstruction-based SSL; this is testable by pretraining on real or mixed real/simulated jets and comparing downstream transfer, which the paper does not do.
  • The contiguous masking via the point sequencer suggests that physics-inspired masking (e.g., masking subjets or splitting the jet by energy scales) could improve learning; the paper only tests random vs. contiguous block masks.
  • Since JEPA is backbone-agnostic, pairing it with a stronger tokenizer or a graph transformer might close the remaining gap to dedicated full-label taggers such as ParT on top tagging.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript introduces HEP-JEPA, a transformer-based foundation model for collider jet physics. The model tokenizes jets into groups via farthest-point sampling and k-nearest neighbors, then uses a JEPA-style objective: a context encoder processes a fraction of the tokens and a predictor estimates target token embeddings in latent space, trained on 100M jets from JetClass. The authors evaluate few-shot classification on JetClass at label fractions from 0.05% to 100%, transfer to top tagging (TQTR) and quark-gluon tagging (Komiske et al.), and perform ablations over masking strategy, number of target tokens, context ratio, physics bias, registers, and augmentations. The central claim is that fine-tuned HEP-JEPA outperforms a from-scratch supervised model by 4–6 points at the lowest label fractions.

Significance. If the few-shot gains are reproducible, HEP-JEPA would be a useful demonstration that latent-space predictive pretraining can reduce labeling requirements for jet classification, complementing existing masked-particle and generative HEP foundation models. The paper deserves credit for a large-scale 100M-jet pretraining run, detailed ablations, and honest comparison with task-specific state of the art, where it acknowledges that ParT/ParticleNet perform better. The significance of the work hinges entirely on the robustness of the few-shot numbers in Table 1, which the current manuscript does not establish because the results are single-run peak validation accuracies with no uncertainty quantification and no test-set evaluation.

major comments (3)
  1. [§5.1.1, Table 1] The central few-shot claim is reported as peak validation accuracies from single runs, with no seeds, no error bars, no test-set accuracies, and no fixed stopping rule. Because peak validation accuracy is an optimistic estimator that rewards run-to-run luck, the claimed 4–6 percentage point advantage over from-scratch training may be inflated by checkpoint selection. Please report mean ± std over multiple seeds and evaluate on the held-out JetClass test split. In addition, the label-fraction column is internally inconsistent: 0.05% of 100M training samples is 50K, not 5K, and 0.5% is 500K, not 50K; Figure 2 also calls 0.5% labels '50000 training samples.' The sample counts must be corrected or the base dataset size clarified.
  2. [§5.2.2, Table 3] The text states that fine-tuned HEP-JEPA outperforms from-scratch training but falls short of a fine-tuned supervised model, yet Table 3 does not contain a HEP-JEPA fine-tuned row or a supervised fine-tuned row. The only HEP-JEPA entry differs from from-scratch by 0.002 (0.821 vs 0.819), which is within likely noise and does not by itself demonstrate transfer. Add the missing rows, or revise the text and qualify the transfer claim.
  3. [§6, Tables 4–7] The ablation protocol selects hyperparameters using peak validation accuracy from a one-epoch pretraining evaluated with an SVM linear probe on 50K JetClass validation samples, while the final model is trained for four epochs and fine-tuned on the same JetClass validation distribution for few-shot evaluation. This creates a risk of selection bias and optimistic few-shot numbers; please clarify whether the final configuration was fixed before the few-shot experiments and provide test-set results. Also, Table 4's caption swaps the context and target sample ratios relative to Section 3.2.1, and the 0.002 advantage of contiguous-1 over random-4 in Table 4 is too small to support the text's conclusion without uncertainty estimates.
minor comments (6)
  1. [§3.1] The text states p_i ∈ R^7 but lists eight feature components (η, φ, m, ln pT, ln E, ln pT/pTJ, ln E/EJ, ∆R_iJ); correct the dimension or the feature list.
  2. [Table 3 caption] The caption refers to 'two state-of-the-art models on top tagging' for the quark-gluon tagging task; update it to refer to quark-gluon tagging.
  3. [Figure 2] The '3X faster' claim is based on training-step validation-loss curves; specify whether this is wall-clock time or steps and report variance, since the text does not provide timing measurements.
  4. [Table 4] The column header 'FREQUENCY' should be renamed to 'number of target tokens' to match the text, and the caption's context/target ratio ranges need correcting.
  5. [§5.1.1] The few-shot comparison would be more informative if it also included existing self-supervised HEP foundation models such as OmniJet-α and Masked Particle Modelling in the same setting; at present only from-scratch supervised training is used as a baseline in Table 1.
  6. [Abstract / Project site] Consider releasing the pretrained weights and fine-tuning code on the project site, since no code or model artifacts are linked in the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the JEPA pretraining objective is an auxiliary latent-space prediction task, and the downstream few-shot and transfer evaluations are external benchmarks whose outcomes are not encoded in the pretraining loss.

full rationale

The paper's claimed derivation chain is empirical rather than definitional. HEP-JEPA is pretrained to predict masked patch embeddings in latent space (Section 3.2.2), which is an auxiliary task, not an equivalent restatement of the downstream jet-classification objective. The few-shot evaluation on JetClass uses the same data distribution as pretraining, but the class labels are not used during pretraining, so the reported accuracy gains are not forced by construction. The transfer tasks (top tagging on TQTR and quark-gluon tagging on the Komiske et al. dataset) are external benchmarks, and the paper reports comparisons against from-scratch and supervised baselines. No load-bearing step reduces to a fitted parameter renamed as a prediction: the ablations select hyperparameters based on validation accuracy, which is standard model selection rather than circularity. Citations to JEPA (Assran et al., 2023), Point-JEPA (Saito & Poovvancheri, 2024), and the physics bias from Particle Transformer (Qu et al., 2022a) are external prior work, not self-citations, and no uniqueness theorem is invoked to forbid alternatives. Concerns that the few-shot numbers are peak validation accuracies from single runs without seeds or error bars, and that Table 3 omits the fine-tuned HEP-JEPA row referenced in Section 5.2.2, are statistical reporting and internal-consistency issues, not circularity. The derivation is therefore self-contained with respect to the circularity criteria.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper is an empirical ML study, not a first-principles derivation. No physical constants are fitted. The central claim rests on the inductive power of the JEPA objective and the specific tokeniser and masking scheme. The masking ratios and target count were selected via ablations on the same dataset, which is a form of data-dependent tuning. No new physical entities are introduced.

free parameters (4)
  • context_sample_ratio = [0.85, 1.0]
    Chosen after ablation II (Table 5) to maximize accuracy; final model uses this range.
  • target_sample_ratio = [0.15, 0.2]
    Used in pretraining as described in Section 3.2.1; not ablated separately despite Table 4 ambiguity.
  • number_of_target_tokens = 1
    Selected from ablation I (Table 4) where one target performed best.
  • masking_strategy = contiguous
    Selected from ablation I (Table 4) over random masking.
assumptions (3)
  • domain assumption JEPA's EMA teacher design prevents representational collapse.
    The authors rely on this from prior work (Assran et al., 2023) without proof in this paper; if collapse occurred the pretraining would learn nothing.
  • domain assumption The particle group tokeniser with FPS and kNN preserves physics information required for downstream classification.
    The tokeniser selects only c centres and groups k neighbours; the paper does not prove this retains enough information.
  • domain assumption The smooth L1 loss in embedding space is a suitable learning signal for transfer.
    This is the core JEPA objective; the paper assumes it produces semantically meaningful embeddings without direct validation against other self-supervised losses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture." pith.science (2026). https://pith.science/paper/XPC7UWH2

@misc{pith2026250203933,
  author       = {Pith},
  title        = {Pith review of: HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XPC7UWH2}},
  note         = {Machine review of arXiv:2502.03933}
}
read the original abstract

We present a transformer architecture-based foundation model for tasks at high-energy particle colliders such as the Large Hadron Collider. We train the model to classify jets using a self-supervised strategy inspired by the Joint Embedding Predictive Architecture. We use the JetClass dataset containing 100M jets of various known particles to pre-train the model with a data-centric approach -- the model uses a fraction of the jet constituents as the context to predict the embeddings of the unseen target constituents. Our pre-trained model fares well with other datasets for standard classification benchmark tasks. We test our model on two additional downstream tasks: top tagging and differentiating light-quark jets from gluon jets. We also evaluate our model with task-specific metrics and baselines and compare it with state-of-the-art models in high-energy physics. Project site: https://hep-jepa.github.io/

Figures

Figures reproduced from arXiv: 2502.03933 by the authors.

Figure 1
Figure 1. Schematic diagram illustrating the working of the HEP-JEPA model. The model has a structure similar to vision transformers. In the first step, the entire jet is divided into patches using a particle jet tokeniser. These tokens are then masked to form the context and target blocks. Each block is fed into the respective encoder to generate the embeddings. The context embedding, along with the special mask tokens, is u… view at source ↗
Figure 2
Figure 2. Validation loss vs. training step for the two benchmark models training in a few-shot learning setting for jet classification on the JetClass dataset with 0.5% labels (i.e., 50000 training sam￾ples). One model is trained from scratch, whereas the pre-trained HEP-JEPA model is fine-tuned. The validation loss falls quickly for the HEP-JEPA model — it achieves the same minimum valida￾tion loss as the model trained from… view at source ↗
Figure 3
Figure 3. The t-SNE plot of the pooled embedding obtained for samples within the JetClass dataset. scratch but slightly falls short of when the latter is further fine-tuned on the dataset. We also show the accuracy scores from the domain state-of-the-art models (Qu & Gouskos, 2020; Qu et al., 2022a) for reference. 5.3. Visualisation We visualise the representation learned by HEP-JEPA on 50k samples of JetClass sampled uniform… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Learning transferable event representations for charmed baryon physics at BESIII

    physics.data-an 2026-07 conditional novelty 6.0 of 10

    A Particle Transformer pre-trained on simulated Lambda_c events transfers across 12 decay channels, improving classification and momentum-direction regression over training from scratch in low-statistics regimes.

  2. Explicit or Implicit? Encoding Physics at the Precision Frontier

    hep-ph 2026-03 conditional novelty 6.0 of 10

    On three precision classification tasks — reweighting-based unfolding, likelihood-ratio estimation, and weakly supervised anomaly detection — a Lorentz-equivariant transformer and a pretrained foundation model perform...

  3. Enhancing next token prediction based pre-training for jet foundation models

    hep-ph 2025-12 conditional novelty 6.0 of 10

    Using continuous particle features as input and combining next-token with masked-token pre-training markedly improves classification accuracy of the OmniJet jet foundation model without visibly hurting its generative quality.

  4. Are We Ready for AI-Driven Discovery? AI Verification Before the Next Fundamental Physics Breakthrough

    physics.data-an 2026-07 accept novelty 4.0 of 10

    Verification of ML in fundamental physics is essential precisely when models enter statistical modeling, inference, or hypothesis testing, and is bounded by unavoidable inductive bias, sample complexity, and experimen...

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages · cited by 4 Pith papers

  1. [7]

    de Favereau, J., Delaere, C., Demin, P., Giammanco, A., Lemaˆıtre, V ., Mertens, A., and Selvaggi, M

    URL https: //arxiv.org/abs/2309.16588. de Favereau, J., Delaere, C., Demin, P., Giammanco, A., Lemaˆıtre, V ., Mertens, A., and Selvaggi, M. DELPHES 3, A modular framework for fast simulation of a generic collider experiment. JHEP, 02:057,

  2. [10]

    Golling, T., Heinrich, L., Kagan, M., Klein, S., Leigh, M., Osadchy, M., and Raine, J

    URL https: //arxiv.org/abs/2403.00504. Golling, T., Heinrich, L., Kagan, M., Klein, S., Leigh, M., Osadchy, M., and Raine, J. A. Masked particle model- ing on sets: towards self-supervised high energy physics foundation models. Mach. Learn. Sci. Tech., 5(3):035074,

  3. [11]

    9 HEP-JEPA: JEPA-based foundation model for collider physics Harris, P., Kagan, M., Krupa, J., Maier, B., and Woodward, N

    doi: 10.1088/2632-2153/ad64a8. 9 HEP-JEPA: JEPA-based foundation model for collider physics Harris, P., Kagan, M., Krupa, J., Maier, B., and Woodward, N. Re-Simulation-based Self-Supervised Learning for Pre-Training Foundation Models. 3

  4. [12]

    Irwin, R., Dimitriadis, S., He, J., and Bjerrum, E

    URL https://arxiv.org/abs/2111.06377. Irwin, R., Dimitriadis, S., He, J., and Bjerrum, E. J. Chem- former: a pre-trained transformer for computational chemistry. Machine Learning: Science and Technol- ogy, 3(1):015022, jan

  5. [13]

    URL https://dx.doi.org/10.1088/ 2632-2153/ac3ffb

    doi: 10.1088/2632-2153/ ac3ffb. URL https://dx.doi.org/10.1088/ 2632-2153/ac3ffb. Kasieczka, G., Nachman, B., and Shih, D. Official datasets for lhc olympics 2020 anomaly detection chal- lenge, November 2019a. URL https://doi.org/ 10.5281/zenodo.4536624. Kasieczka, G., Plehn, T., Thompson, J., and Rus- sel, M. Top quark tagging reference dataset, March 20...

  6. [14]

    Komiske, P., Metodiev, E., and Thaler, J

    URL https://arxiv.org/abs/2406.09246. Komiske, P., Metodiev, E., and Thaler, J. Pythia8 quark and gluon jets for energy flow, May

  7. [15]

    Leigh, M., Klein, S., Charton, F., Golling, T., Heinrich, L., Kagan, M., Ochoa, I., and Osadchy, M

    URL https: //doi.org/10.5281/zenodo.3164691. Leigh, M., Klein, S., Charton, F., Golling, T., Heinrich, L., Kagan, M., Ochoa, I., and Osadchy, M. Is Tokenization Needed for Masked Particle Modelling? 9

  8. [16]

    From Words to Molecules: A Survey of Large Language Models in Chemistry

    URL https://arxiv.org/abs/ 2402.01439. McCabe, M., Blancard, B. R.-S., Parker, L. H., Ohana, R., Cranmer, M., Bietti, A., Eickenberg, M., Golkar, S., Krawezik, G., Lanusse, F., Pettee, M., Tesileanu, T., Cho, K., and Ho, S. Multiple physics pretraining for physical surrogate models,

Show all 25 references
  1. [17]

    Mikuni, V

    URL https://arxiv.org/ abs/2310.02994. Mikuni, V . and Nachman, B. OmniLearn: A Method to Simultaneously Facilitate All Jet Physics Tasks. 4

  2. [18]

    org/abs/2203.06604

    URL https://arxiv. org/abs/2203.06604. Parker, L., Lanusse, F., Golkar, S., Sarra, L., Cranmer, M., Bietti, A., Eickenberg, M., Krawezik, G., McCabe, M., Morel, R., Ohana, R., Pettee, M., R´egaldo-Saint Blancard, B., Cho, K., and Ho, S. Astroclip: a cross-modal founda- tion mo...

  3. [19]

    doi: 10.1093/mnras/stae1450

    ISSN 1365-2966. doi: 10.1093/mnras/stae1450. URL http: //dx.doi.org/10.1093/mnras/stae1450. Qu, H. and Gouskos, L. ParticleNet: Jet Tagging via Particle Clouds. Phys. Rev. D, 101(5):056019,

  4. [20]

    Qu, H., Li, C., and Qian, S

    doi: 10.1103/ PhysRevD.101.056019. Qu, H., Li, C., and Qian, S. Particle trans- former for jet tagging. ArXiv, abs/2202.03772, 2022a. URL https://api.semanticscholar. org/CorpusID:246652443. Qu, H., Li, C., and Qian, S. Jetclass: A large-scale dataset for deep learning in jet ...

  5. [21]

    2016239118

    doi: 10.1073/pnas. 2016239118. URL https://www.pnas.org/doi/ abs/10.1073/pnas.2016239118. Ross, J., Belgodere, B., Chenthamarakshan, V ., Padhi, I., Mroueh, Y ., and Das, P. Large-scale chemical language representations capture molecular structure and proper- ties,

  6. [22]

    URL https://arxiv.org/abs/2106. 09553. Saito, A. and Poovvancheri, J. Point-jepa: A joint embed- ding predictive architecture for self-supervised learning on point cloud. arXiv:2404.16432,

  7. [25]

    Zhang, R., Guo, Z., Fang, R., Zhao, B., Wang, D., Qiao, Y ., Li, H., and Gao, P

    URL https://arxiv.org/ abs/2302.13971. Zhang, R., Guo, Z., Fang, R., Zhao, B., Wang, D., Qiao, Y ., Li, H., and Gao, P. Point-m2ae: Multi-scale masked au- toencoders for hierarchical point cloud pre-training,

  8. [26]

    URL https://arxiv.org/abs/2205.14401. 11

  9. [2014]

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K

    doi: 10.1007/ JHEP02(2014)057. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional trans- formers for language understanding,

  10. [2015]

    Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M., and Gholami, A

    doi: 10.1016/j.cpc.2015.01.024. Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M., and Gholami, A. Towards foundation models for scientific machine learning: Char- acterizing scaling and transfer behavior,

  11. [2018]

    cite arxiv:1810.04805Comment: 13 pages

    URL http://arxiv.org/abs/1810.04805. cite arxiv:1810.04805Comment: 13 pages. Dillon, B. M., Kasieczka, G., Olischlager, H., Plehn, T., Sorrenson, P., and V ogel, L. Symmetries, safety, and self-supervision. SciPost Phys., 12(6):188, 2022a. doi: 10.21468/SciPostPhys.12.6.188. D...

  12. [2019]

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y ., and Ballas, N

    URL https:// doi.org/10.5281/zenodo.3548091. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y ., and Ballas, N. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on ...

  13. [2020]

    Cacciari, M., Salam, G

    URL https:// arxiv.org/abs/2005.14165. Cacciari, M., Salam, G. P., and Soyez, G. The anti- kt jet clustering algorithm. JHEP, 04:063,

  14. [2021]

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G

    URL https://arxiv.org/ abs/2104.14294. Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual repre- sentations, 2020a. URL https://arxiv.org/abs/ 2002.05709. Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple fram...

  15. [2022]

    Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y ., Assran, M., and Ballas, N

    URL https:// arxiv.org/abs/2106.08254. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y ., Assran, M., and Ballas, N. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471,

  16. [2023]

    Team, G., Anil, R., Borgeaud, S., Wu, Y ., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A

    URL https://arxiv.org/abs/2306.00258. Team, G., Anil, R., Borgeaud, S., Wu, Y ., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805,

  17. [2024]

    doi: 10.1021/acs.jcim. 4c01396. URL https://doi.org/10.1021/acs. jcim.4c01396. PMID: 39231030. Birk, J., Hallin, A., and Kasieczka, G. OmniJet- α: the first cross-task foundation model for particle physics. Mach. Learn. Sci. Tech. , 5(3):035031,

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.