Gather-and-Aggregate

This repository accompanies the paper:
"Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism"

We investigate the retrieval capabilities of recurrent models (like Mamba) and how they compare to Transformers. Our core contribution is an analysis of the Gather-and-Aggregate (G&A) mechanism, showing that retrieval behavior is driven by a small subset of heads in both Transformers and State-space models (SSM).

---

📂 Repository Structure

.
├── notebook.ipynb         # Main experiment notebook
├── kv_retrieval.py        # Tools for key-value retrieval analysis
├── utils.py               # Miscellaneous utilities
├── visualize.py           # Plotting and visualization functions
├── models/                # Architecture components (e.g., hybrid, SSM, transformer)
└── assets/                # Static images or data files

---

🚀 Quickstart

1. Clone the repo:

   git clone https://github.com/goombalab/Gather-and-Aggregate.git
   cd Gather-and-Aggregate
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Run the notebook:

   jupyter notebook notebook.ipynb
   

---

🔍 Core Experiments

• Head Importance Analysis:
  Identify which attention heads perform retrieval across models like Llama and Mamba.

• Hybrid Architectures:
  Replace non-retrieval heads with SSM layers to form faster, retrieval-preserving models.

• Hidden-State Alignment:
  Distill representations from transformer models into hybrids to improve performance.

---

📄 Citation

If you find this repository useful, please cite:

@misc{bick2025understandingskillgaprecurrent,
      title={Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism}, 
      author={Aviv Bick and Eric Xing and Albert Gu},
      year={2025},
      eprint={2504.18574},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2504.18574}, 
}