---

Table of Contents

1. News
2. Abstract
3. Installation
4. Usage
5. Training
6. Citation
7. Acknowledgments

🔥 News

• [2025.12] Our demo is now running on huggingface/spaces.
• [2025.12] We add an interactive demo to visualize the assigned positions. Please find it in ./visual!
• [2025.12] We have released the training code and evaluation scripts!
• [2025.12] Pre-trained models (based on OLMo-2 1B) are now available on Hugging Face.
• [2025.12] The paper "RePo: Language Models with Context Re-Positioning" is released on arXiv.

🧩 Abstract

In-context learning is fundamental to modern Large Language Models (LLMs); however, prevailing architectures impose a rigid and fixed contextual structure by assigning linear or constant positional indices. Drawing on Cognitive Load Theory (CLT), we argue that this uninformative structure increases extraneous cognitive load, consuming finite working memory capacity that should be allocated to deep reasoning and attention allocation. To address this, we propose RePo, a novel mechanism that reduces extraneous load via context re-positioning. Unlike standard approaches, RePo utilizes a differentiable module, $f_\phi$, to assign token positions that capture contextual dependencies, rather than replying on pre-defined integer range. By continually pre-training on the OLMo-2 1B backbone, we demonstrate that RePo significantly enhances performance on tasks involving noisy contexts, structured data, and longer context length, while maintaining competitive performance on general short-context tasks. Detailed analysis reveals that RePo successfully allocate higher attention to distant but relevant information, assign positions in dense and non-linear space, and capture the intrinsic structure of the input context.

This is the initial repository for the research project RePo. Please feel free to open issues if you have any questions or find any mistakes.

🛠️ Installation

1. Clone the repository

   git clone https://github.com/SakanaAI/repo
   cd repo
   

2. Setup for Evaluation

   # We tested this setup on H100 and 6000Ada
   # in ./repo
   conda create -n olmes python=3.11
   
   ### Important: enable only if you have CUDA > 12.4, this is critical for the compile of vLLM
   # conda install -c nvidia/label/cuda-12.4.0 cuda-toolkit
   # pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
   
   ### install torch
   pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0
   
   ### install vLLM with RePo
   cd vllm
   python use_existing_torch.py
   pip install -r requirements/build.txt
   mkdir -p vllm/vllm_flash_attn
   pip install -e . --no-build-isolation
   
   ### install transformers with RePo
   cd ../transformers
   pip install -e '.[torch]' --no-build-isolation
   
   ### install test suites
   cd ../olmes
   pip install -e . --no-build-isolation
   

3. Setup for Train

   # We tested this setup on H100
   # in ./repo
   cd OLMo
   
   ### install OLMo with RePo
   conda env create -f environment.yml
   conda activate olmo
   pip install flash-attn==2.7.4.post1
   pip install -e .[all]
   

💻 Usage

Quick Inference

Please download the 1B model from huggingface in adavance:

cd olmes
python scripts/hf_download_1b.py # only download for the first use
bash run_eval_1b.sh

The scripts for the evaluation of 7B model are also provided under olmes.

We will upload the checkpoint of 7B model soon.

🏋️ Training

Please take a look at the script OLMo/batch_run_stage2_1b.sh, you need to replace the placeholder to the state-2 data by your real data path, following the instruction of OLMo.

cd OLMo
SLURM_ARRAY_TASK_ID=2 bash batch_run_stage2_1b.sh -d $YOUR_DATA_DIR

If you want to train 7B model, please check OLMo/batch_run_stage2_7b.sh and pre-download the stage-1 checkpoint of 7B model, because online downloading is super slow.

📜 Citation

If you find this project useful, please cite our paper:

@article{sakana2025repo,
  title={RePo: Language Models with Context Re-Positioning},
  author={Huayang Li, Tianyu Zhao, Deng Cai, and Richard Sproat},
  year={2025},
  eprint={2512.14391},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2512.14391},
}

🙏 Acknowledgments

• We utilized code from OLMo, olmes, vLLM, and transformers.