LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models

Taekhyun Park1, Yongjae Lee2, Dohee Kim3 and Hyerim Bae2†
1Department of Data Science, Pusan National University, Busan, South Korea,
2Department of Industrial Engineering, Pusan National University, Busan, South Korea,
3 Department of of AI Convergence Engineering, Changwon National University, Changwon, South Korea
†Corresponding author
Paper Video Code Hugging Face Hugging Face

Abstract

Looped computation shows promise in improving the reasoning-oriented performance of LLMs by scaling test-time compute. However, existing approaches typically require either training recurrent models from scratch or applying disruptive retrofits, which involve substantial computational costs and may compromise pretrained capabilities. To address these limitations, we introduce **Looped Depth Up-Scaling** (LoopUS), a post-training framework that converts a standard pretrained LLM into a looped architecture. As a key technical contribution, LoopUS recasts the pretrained LLM into an encoder, a looped reasoning block, and a decoder. It operationalizes this latent-refinement architecture through four core components: (1) block decomposition, guided by staged representation dynamics; (2) an input-dependent selective gate to mitigate hidden-state drift; (3) random deep supervision for memory-efficient learning over long recursive horizons; and (4) a confidence head for adaptive early exiting. Collectively, these mechanisms transform a standard non-looped model into a looped form while stabilizing it against both computational bottlenecks and representation collapse. Through stable latent looping, LoopUS improves reasoning-oriented performance without extending the generated traces or requiring recurrent training from scratch.

Framework

Framework Overview

Figure 1: Conceptual view of latent refinement in LoopUS. As the reasoning block is looped, each proposed update is mixed with the previous hidden state by the selective gate, gradually steering the trajectory toward the answer region instead of allowing it to drift.

Framework Overview

Figure 2: Overview of the LoopUS architecture. (a) A pretrained LLM is recast into encoder, reasoning, and decoder blocks, using a selective gate ($\mathcal{G}$) inserted between loop iterations to stabilize the loop dynamics. (b) The looped LLM is trained with random deep supervision using next-token prediction loss ($\mathcal{L}_{\mathrm{LM}}$), monotonicity loss ($\mathcal{L}_{\text{Mono}}$), and confidence loss ($\mathcal{L}_{\text{Q}}$).

Experiments

Method Base
Model		 Train
Tokens		 Setting Task AVG Δ
ARC-E ARC-C HS WG PIQA OBQA
acc_n ↑ acc_n ↑ acc_n ↑ acc ↑ acc_n ↑ acc_n ↑
Ours TinyLlama
1.1B		 3B Original 47.1 25.1 42.2 53.4 66.8 24.2 43.1 --
Adapted 53.0 29.6 55.5 57.9 69.8 30.6 49.4 +6.3
McLeish et al. TinyLlama
1.1B-3T		 52B Original 55.7 31.0 59.1 58.9 73.0 35.0 52.1 --
Adapted 58.6 35.6 45.1 57.6 66.4 32.2 49.3 -2.9
Bae et al. TinyLlama
1.1B		 60B Original 44.7 23.2 42.2 53.4 66.8 29.2 43.3 --
Adapted 49.9 26.2 48.8 54.1 68.6 32.8 46.7 +3.5

LoopUS shows adaptation efficiency under a smaller training-token budget. All methods adapt a TinyLlama-based backbone; w/o and w/ LoopUS denote the checkpoint before and after adaptation, respectively. AVG is the unweighted mean over the six tasks, and Δ reports the change in AVG from Original to Adapted. Results for prior methods are taken from the corresponding papers.

Model Setting Wiki LAMBADA MMLU HS ARC-E ARC-C PIQA WG OBQA AVG Δ
ppl ↓ acc ↑
Qwen 1.7B w/o LoopUS 21 12.21 55.4 46.2 72.5 40.2 72.2 61.3 28 53.7 --
w/ LoopUS 16.9 7.43 56.6 46.3 74.9 43.1 73.3 63.0 29.6 55.3 +1.6
Qwen 4B w/o LoopUS 16.4 7.29 68.3 52.1 80.2 50.4 75.0 66.5 29.4 60.3 --
w/ LoopUS 13.9 5.33 67.7 51.4 81.3 54.0 76.8 68.9 34.4 62.1 +1.8
Qwen 8B w/o LoopUS 12.2 4.58 72.8 57.2 81.5 55.4 76.3 67.9 31.6 63.2 --
w/ LoopUS 10.3 4.32 71.5 56.0 83.9 58.1 78.9 72.4 37.0 65.4 +2.2
Phi-4 14B w/o LoopUS 9.59 4.03 76.9 63.1 81.3 55.8 80.7 77.0 34.0 67.0 --
w/ LoopUS 7.75 3.49 77.5 60.58 83.5 57.7 81.8 77.5 41.8 68.6 +1.7

LoopUS improves pretrained backbones across scales. Results on language modeling and downstream benchmarks. ppl denotes perplexity (lower is better), and acc denotes accuracy (higher is better). AVG is the mean over the seven acc benchmarks, and Δ denotes the change in AVG from the original backbone (w/o LoopUS) to the adapted checkpoint (w/ LoopUS). Bold highlights the better result between the two variants of each backbone. All models are evaluated zero-shot.

Framework Overview

Figure 3: KV caching accelerates LoopUS decoding. With the recursion budget fixed to $B=8$, caching consistently reduces seconds per generated token across Qwen3-1.7B, Qwen3-4B, and Qwen3-8B, with the largest gains appearing at longer generations.

Framework Overview

Figure 4: LoopUS quickly stabilizes latent reasoning trajectories. For representative Qwen4B examples, the PCA projection of hidden states shows a large early transition followed by compact refinement near the final answer, while the selective-gate coefficient and hidden-state distances rapidly contract over thinking iterations.

Framework Overview

Figure 5: Token-level thinking traces reveal staged refinement. Example generations from Qwen4B show how LoopUS organizes reasoning across successive steps, progressively refining intermediate tokens before converging to the final response.

Video

GPU Sponsorship

We are always looking for GPU sponsorship. If you are interested, please contact. pthpark1@pusan.ac.kr