LEAP: Layer-skipping Efficiency via Adaptive Progression for Vision Transformer Distillation

Jiaqi Zhang1, Ashton Lee2, Anthony Wong1, John Zou1, Sami BuGhanem1, Randall Balestriero1
1Brown University   2Rice University
LEAP curriculum overview figure.
Training dataset
Student
Teacher
CKA: 0.30 L2 loss: 1.00
Target: teacher block 1 / 12  ·  Advance threshold: CKA ≥ 0.85

Overview of LEAP. Rather than supervising the student against a fixed teacher block, our curriculum advances the supervisory target through the teacher's feature maps shallow-to-deep based on online CKA alignment.

Abstract

Vision Foundation Models (VFMs) with Vision Transformer (ViT) backbones, such as DINOv2, have become essential for downstream tasks like object recognition and semantic segmentation. The immense computational requirements of backbones often necessitate distillation into smaller architectures for edge deployment. Feature-based knowledge distillation (KD) often suffers from the teacher–student gap; the student struggles to imitate the teacher's complex feature map due to its limited capacity.

To mitigate this bottleneck, we propose LEAP: Layer-skipping Efficiency via Adaptive Progression, a training curriculum for ViT feature-based knowledge distillation. By utilizing the teacher's intermediate feature maps as a sequence of progressively more difficult targets, our curriculum allows the student to build a foundational representation before tackling higher-level abstractions. Our results demonstrate that this paradigm significantly accelerates convergence through adaptive difficulty selection across various student model sizes and dataset scales.

With our curriculum, the LEAP-distilled ViT-S achieves 90.1% accuracy on ImageNet-100, a +12.24% improvement over the baseline. On ImageNet-1K, LEAP achieves +3.84% and +7.75% improvement for the instance retrieval task on the Oxford and Paris datasets, respectively. Furthermore, the curriculum enables 25.1% savings in training FLOPs and 21% savings in training time on ImageNet-100 by implementing early-stopping for teacher inference during the initial stages of training.

Left: linear-probing accuracy of LEAP vs. baseline over training epochs. Right: visualization of the curriculum's progress through teacher blocks.

Faster convergence. Left: LEAP's linear-probing accuracy outpaces the baseline from the very first epochs. Right: the curriculum advances quickly through shallow blocks and dwells on deeper, more semantic ones.

Key Results

Table 1. Linear probing & training efficiency comparison.

Method Lin. Probe (patch) ↑ Lin. Probe (CLS) ↑ mini-IN-C ↑ FLOPs Saving Train Time Saving
Baseline (ViT-S) 77.86% 77.12% 47.80%
Baseline (ViT-Tiny) 75.90% 76.04% 46.90%
LEAP (ViT-S) 90.10% 89.66% 66.69% 25.1% 21%
LEAP (ViT-Tiny) 81.76% 82.02% 51.74% 28.82% 22.5%

Table 2. Instance recognition performance (mAP) on the Oxford and Paris datasets.

Method roxford5k rparis6k
E M H Mean E M H Mean
Baseline (ViT-S) 10.28 8.80 2.15 7.08 24.83 21.29 7.25 17.79
Baseline (ViT-Tiny) 8.04 7.72 1.84 5.87 21.35 19.42 6.59 15.79
LEAP (ViT-S) 22.56 17.30 4.81 14.89 53.84 40.99 15.97 36.93
LEAP (ViT-Tiny) 14.70 12.15 2.84 9.90 43.39 33.36 12.38 29.71

Table 3. Semantic segmentation performance on ADE20K.

Method Linear Seg. (mIoU) ↑ EOMT val (mIoU) ↑ MS (mIoU) ↑
Baseline (ViT-S) 12.15% 24.49% 24.62%
Baseline (ViT-Tiny) 9.51% 15.41% 14.88%
LEAP (ViT-S) 20.53% 38.10% 39.36%
LEAP (ViT-Tiny) 14.71% 21.67% 21.82%

Table 4. Comparison of distillation alignment strategies.

Alignment Strategy Linear Probing ↑ Projectors (Params)
Baseline (last-layer only) 81.80% ×1 (0.07M)
One-to-one Alignment 83.38% ×12 (0.89M)
LEAP 83.36% ×1 (0.07M)

BibTeX

@article{zhang2026leap,
  title         = {LEAP: Layer-skipping Efficiency via Adaptive Progression for Vision Transformer Distillation},
  author        = {Zhang, Jiaqi and Lee, Ashton and Wong, Anthony and Zou, John and BuGhanem, Sami and Balestriero, Randall},
  journal       = {arXiv preprint arXiv:2606.19483},
  year          = {2026}
}