Lorsa (Low-Rank Sparse Attention)

Lorsa is a novel attention decomposition method designed to tackle attention superposition, extracting tens of thousands of interpretable attention units from the attention layers of large language models.

📢 Important Notice

The complete implementation of Lorsa has been migrated to the Language-Model-SAEs repository.

This repository is a comprehensive, fully-distributed framework for training, analyzing, and visualizing Sparse Autoencoders (SAEs) and their frontier variants, including:

• Lorsa (Low-Rank Sparse Attention)
• CLT (Cross-layer Transcoder)
• MoLT (Mixture of Linear Transforms)
• CrossCoder
• And many more SAE variants

🔗 Links

• Code Implementation: OpenMOSS/Language-Model-SAEs
• Paper: Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition

📖 About Lorsa

Lorsa employs low-rank sparse decomposition to decompose attention layer outputs into interpretable feature units, effectively addressing the feature superposition problem in attention mechanisms. This enables a deeper understanding of how attention mechanisms work in large language models.

Key Features

• Extract tens of thousands of interpretable attention units from attention layers
• Leverage low-rank sparse decomposition to handle attention superposition
• Support large-scale distributed training
• Provide comprehensive visualization tools

🚀 Quick Start

Please visit the Language-Model-SAEs repository for:

• Installation instructions
• Training examples
• Usage tutorials

📚 Citation

If you use Lorsa in your research, please cite:

@article{He2025Lorsa,
  author    = {Zhengfu He and Junxuan Wang and Rui Lin and Xuyang Ge and 
               Wentao Shu and Qiong Tang and Junping Zhang and Xipeng Qiu},
  title     = {Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition},
  journal   = {CoRR},
  volume    = {abs/2504.20938},
  year      = {2025},
  url       = {https://arxiv.org/abs/2504.20938},
  eprint    = {2504.20938},
  eprinttype = {arXiv}
}

📝 Changelog

• 2025.4.29: Initial release of Lorsa, introducing low-rank sparse attention decomposition
• 2025.11.9: Implementation migrated to Language-Model-SAEs repository for better framework support

---

For any questions or suggestions, please submit an Issue in the Language-Model-SAEs repository!