Attention-Only White-Box Transformer Achieves Competitive Performance

Yang Bai, Linyuan Wang, Haoyang Jiang, Nuolin Sun, Libin Hou, Bin Yan· August 6, 2026 View original

Key takeaways

  • An attention-only Transformer can achieve competitive performance with fewer parameters.
  • LeJEPA-based self-supervised pretraining is effective for optimizing white-box networks.
  • MLP layers in Transformers might be significantly redundant, offering optimization opportunities.
  • This approach could lead to more efficient and deployable AI models.

Who benefits

AI/ML DevelopmentEdge ComputingComputer VisionAutonomous SystemsCloud Computing

Summary

Researchers developed an attention-only white-box Transformer, optimized through LeJEPA-based self-supervised pretraining, which removes MLP layers and reduces parameter count by 31% while maintaining competitive classification accuracy on CIFAR datasets. This work explores the potential redundancy of MLP modules in standard ViT architectures.

A new "attention-only" white-box Transformer architecture has been proposed, which significantly streamlines the traditional Transformer design. This model integrates the derivation of white-box networks with a LeJEPA-based self-supervised learning framework, allowing for a joint optimization perspective. By aligning the LeJEPA framework's assumption of an isotropic Gaussian embedding distribution with the sparse rate reduction objective used in white-box Transformer optimization, the researchers were able to eliminate the need for MLP layers and ISTA structures. The resulting attention-only Transformer achieves classification accuracies of 88.88% on CIFAR-10 and 63.54% on CIFAR-100, closely matching the performance of the original white-box Transformer (CRATE) while reducing the parameter count by approximately 31%. Furthermore, the study suggests that even standard Vision Transformers (ViTs) could potentially remove a substantial portion of their MLP blocks, reducing parameters by about 66% with competitive accuracy, hinting at significant redundancy in current ViT designs.

Why it matters

This research offers a path to more efficient and interpretable Transformer models by reducing complexity and parameter count, which can lead to faster training, lower inference costs, and easier deployment on resource-constrained devices.

How to implement this in your domain

  1. 1Evaluate the feasibility of adopting attention-only Transformer architectures for specific vision tasks.
  2. 2Experiment with LeJEPA-based self-supervised pretraining for model initialization.
  3. 3Investigate replacing MLP blocks with simpler activation functions in existing ViT models to reduce parameters.
  4. 4Benchmark the performance and efficiency gains on proprietary datasets and hardware.

Original post by Yang Bai, Linyuan Wang, Haoyang Jiang, Nuolin Sun, Libin Hou, Bin Yan

"arXiv:2608.04213v1 Announce Type: new Abstract: Existing studies on self-supervised learning for white-box networks typically decouple the derivation of white-box networks via optimization algorithms from self-supervised learning paradigms. In this work, we instead revisit the tw…"

View on X

Originally posted by Yang Bai, Linyuan Wang, Haoyang Jiang, Nuolin Sun, Libin Hou, Bin Yan on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses