JEPAMatch: Geometric Representation Shaping
for Semi-Supervised Learning

Ali Aghababaei-Harandi · Aude Sportisse · Massih-Reza Amini

Université Grenoble Alpes, CNRS, Computer Science Laboratory LIG, Grenoble, France

Code (GitHub) Paper (arXiv:2604.21046) License: MIT

JEPAMatch combines FlexMatch-style adaptive pseudo-labeling with a LeJEPA-inspired representation-level objective, so that instead of only thresholding softmax outputs, the model also explicitly shapes its latent space into well-separated, isotropic per-class clusters. This fixes two long-standing FixMatch-family bottlenecks — majority-class dominance in pseudo-labeling, and the ~2²⁰ iteration convergence tax — cutting the iteration budget by 8× while matching or beating prior SSL methods on CIFAR-100, STL-10, and Tiny-ImageNet.

JEPAMatch architecture diagram — A shared backbone feeds two levels: the **Curriculum Level** (top) does FlexMatch-style adaptive pseudo-labeling on weak/strong views; the **Representation Level** (bottom) aligns global and local crops via a JEPA prediction loss, regularized by **Adaptive Class-wise SIGReg** — per-class isotropic Gaussians instead of one global one, kept apart by an active repulsion term.

Highlights

8× fewer iterations to converge. JEPAMatch reaches FlexMatch's final CIFAR-100 (400-label) accuracy roughly 50k steps earlier, and finishes training at 2¹⁷ iterations vs. the standard 2²⁰.
State-of-the-art or competitive results on CIFAR-100, STL-10, and Tiny-ImageNet against FixMatch, FlexMatch, FreeMatch, SoftMatch, CrMatch, SimMatch, Suave, and RegMixMatch.
More robust under class imbalance, and works as a drop-in module on top of FlexMatch, FreeMatch, or SoftMatch's curriculum.

Results

Error rate (%), lower is better, 3 seeds.

CIFAR-100 & STL-10
Method	Iter.	CIFAR-100 (400)	CIFAR-100 (2.5K)	CIFAR-100 (10K)	STL-10 (40)	STL-10 (1K)
FixMatch	2²⁰	46.42 ± 0.82	28.03 ± 0.16	22.20 ± 0.12	35.97 ± 4.14	6.25 ± 0.33
FlexMatch	2²⁰	39.94 ± 1.62	26.49 ± 0.20	21.90 ± 0.15	29.15 ± 4.16	5.77 ± 0.18
FreeMatch	2²⁰	37.98 ± 0.42	26.47 ± 0.20	21.68 ± 0.03	15.56 ± 0.55	5.63 ± 0.15
SoftMatch	2²⁰	37.10 ± 0.77	26.66 ± 0.25	22.03 ± 0.03	21.42 ± 3.48	5.73 ± 0.24
CrMatch	2²⁰	39.45 ± 1.69	25.43 ± 0.14	20.40 ± 0.08	–	4.89 ± 0.17
SimMatch	2²⁰	37.81 ± 2.21	25.07 ± 0.32	20.58 ± 0.11	–	–
FlatMatch	2²⁰	38.76 ± 1.62	25.38 ± 0.85	19.01 ± 0.43	16.20 ± 4.34	4.82 ± 1.21
Suave*	2²⁰	35.40	23.00	18.40	–	–
RegMixMatch*	2²⁰	35.27	23.78	19.41	11.74	4.66
JEPAMatch (Ours)	2¹⁷	34.25 ± 1.97	22.59 ± 1.17	18.55 ± 0.85	13.44 ± 3.2	4.28 ± 1.43

*standard deviation not reported in the original paper. Full baseline table (PseudoLabel, MeanTeacher, MixMatch, ReMixMatch, UDA) in the paper.

Tiny-ImageNet
Method	Iter.	1K labels	10K labels
FlexMatch	2¹⁸	41.73	27.89
SoftMatch	2¹⁸	40.09	25.92
JEPAMatch	2¹⁸	38.82	24.50

Drop-in module on other curricula (CIFAR-100, 400)
Method	Iter.	Error
FlexMatch	2²⁰	50.15 ± 1.51
FreeMatch	2²⁰	49.64 ± 1.46
SoftMatch	2²⁰	49.24 ± 2.16
JEPAMatch (Flex)	2¹⁷	45.77 ± 2.77
JEPAMatch (Free)	2¹⁷	45.12 ± 1.98
JEPAMatch (Soft)	2¹⁷	44.65 ± 2.14

Convergence speed & pseudo-labeling quality

Convergence speed vs FlexMatch — JEPAMatch reaches FlexMatch's peak accuracy (CIFAR-100, 4 labels/class) roughly 50k iterations earlier, and keeps climbing to a substantially higher final accuracy.

Data utilization vs FlexMatch — JEPAMatch keeps more pseudo-labels above the confidence threshold, at higher correctness, than FlexMatch.

Class dominance comparison — FlexMatch's majority class can claim up to a quarter of all pseudo-labels in a batch; JEPAMatch keeps this far more balanced.

Code

Training code, all configs, and the full paper are on GitHub: aah94/JEPAMatch.

git clone https://github.com/aah94/JEPAMatch.git
cd JEPAMatch
pip install -r requirements.txt
python train.py --c config/classic_cv/jepamatch/jepamatch_cifar100_400_0.yaml

Citation

@article{aghababaeiharandi2026jepamatch,
  title   = {JEPAMatch: Geometric Representation Shaping for Semi-Supervised Learning},
  author  = {Aghababaei-Harandi, Ali and Sportisse, Aude and Amini, Massih-Reza},
  journal = {arXiv preprint arXiv:2604.21046},
  year    = {2026}
}