Overview

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 operationalizes this latent-refinement architecture through:

1. Block Decomposition: Recasts a pretrained transformer into a reusable latent-refinement architecture.
2. Input-Dependent Selective Gate: Adaptively controls hidden state propagation to mitigate drift.
3. Random Deep Supervision: Enables memory-efficient learning over long recursive horizons.
4. Confidence Head: Allows for adaptive early exiting during inference.

Through stable latent looping, LoopUS improves reasoning-oriented performance without extending the generated traces or requiring recurrent training from scratch.

Illustration of LoopUS

Quick Start

To use this model, please follow the installation instructions in the official repository:

git clone https://github.com/Thrillcrazyer/LoopUS.git
cd LoopUS
uv sync

Chatting Mode

uv run chat.py --model-name Thrillcrazyer/Qwen3_1.7B_LoopUS_SFT

Qualitative Generation

uv run LoopUS-generate \
    --model-name microsoft/phi-4 \
    --decomposed-model Thrillcrazyer/Qwen3_1.7B_LoopUS_SFT \
    --prompt "The meaning of life is" \
    --n-recursion 8

Citation

@misc{park2026loopus,
  title        = {LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models},
  author       = {Taekhyun Park and Yongjae Lee and Dohee Kim and Hyerim Bae},
  year         = {2026},
  eprint       = {2605.11011},
  archivePrefix= {arXiv},
  primaryClass = {cs.LG},
  doi          = {10.48550/arXiv.2605.11011},
  url          = {https://arxiv.org/abs/2605.11011}
}