Thrillcrazyer/Qwen3-4B_LoopUS

<div align="center"> <h1>LoopUS: <br> Recasting Pretrained LLMs into Looped Latent Refinement Models</h1> </div>

<p align="center"> <a href="https://pnubaelab.github.io/"><b>BAELAB</b></a>, Pusan National University, Busan, Korea <br> <a href="https://aidoheekim.github.io/"><b>DOLAB</b></a>, Changwon National University, Changwon, Korea </p>

<p align="center"> <a href="https://thrillcrazyer.github.io/" target="blank"><strong>Taekhyun Park</strong></a>¹, <a href="https://yongzzai.com/" target="blank"><strong>Yongjae Lee</strong></a>¹, <a href="https://aidoheekim.github.io/" target="blank"><strong>Dohee Kim</strong></a>², <a href="https://pnubaelab.github.io/" target="blank"><strong>Hyerim Bae</string></a>^1,&dagger; </p>

<p align="center"> <a href="https://github.com/Thrillcrazyer/LoopUS"><b>🌟 Github</b></a> | <a href="https://thrillcrazyer.github.io/LoopUS"><b>🌐 Project Page</b></a> | <a href="https://arxiv.org/abs/2605.11011"><b>📄 Paper</b></a> </p>

Abstract

Looped computation shows promise in improving the reasoning-oriented performance of LLMs by scaling test-time compute. Looped Depth Up-Scaling (LoopUS) is a post-training framework that converts a standard pretrained LLM into a looped architecture. LoopUS recasts the pretrained LLM into an encoder, a looped reasoning block, and a decoder. It improves reasoning-oriented performance without extending the generated traces or requiring recurrent training from scratch.

QuickStart

To use this model, clone the official repository and run the chat interface:

git clone https://github.com/Thrillcrazyer/LoopUS.git
cd LoopUS
uv sync
uv run chat.py --model-name Thrillcrazyer/Qwen3_1.7B_LoopUS_SFT

Illustration of LoopUS

<div align="center"> <img src="https://raw.githubusercontent.com/Thrillcrazyer/LoopUS/main/assets/Framework.png" width="800"/> </div>

Citation

If you find LoopUS useful in your research, please cite:

@article{park2026loopus,
  title={LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models},
  author={Park, Taekhyun and Lee, Yongjae and Kim, Dohee and Bae, Hyerim},
  journal={arXiv preprint arXiv:2605.11011},
  year={2026}
}