Scaling Latent Reasoning via Looped Language Models

Overview

Introducing Ouro

Strategy

How a LoopLM decides how deep to think

1. 01

   Train every depth

   Every loop gets an LM loss, weighted by a learned exit distribution with a uniform-prior entropy term.

2. 02

   Teach the exit gate

   A second stage trains only the gate: continue while the next loop still lowers the loss.

3. 03

   Exit with Q-exit

   Stop once the cumulative exit probability reaches q, a deploy-time compute knob.

Early-exit strategies

• The trained gate is best at every budget (~66% MMLU at 2.5 loops).
• Fixed depth: 1 loop ≈ 40%, 4 loops 67.35%.

KV cache sharing

When decoding, keeping only the last loop's cache matches the full cache at ¼ the memory.

Full cache78.9

Last-step only78.8

Averaged78.7

First-step only18.7

Full cache82.4

Last-step only80.4

Averaged78.5

First-step only8.43

Full cache (1× memory)Last-step only (¼)Averaged (¼)First-step only (¼)

View data table

KV cache sharing during decoding (Ouro-1.4B)

	GSM8K	MATH-500	KV memory
Full cache	78.92	82.40	1×
Last-step only	78.85	80.40	¼
Averaged	78.73	78.52	¼
First-step only	18.73	8.43	¼

Recipe

Training pipeline

7.7T tokens in total; the 2.6B model is upcycled from the shared 3T-token checkpoint.

1. Warmup

   Common warmup for both model sizes.

2. 3T

   Stable training I

   Shared pre-training phase at 4K sequence length.

3. Model branching

   Keep the 1.4B model; upcycle a 2.6B model by duplicating layers.

4. 3T

   Stable training II

   Both sizes continue with 4 recurrent steps for stability.

5. 1.4T

   CT annealing

   Continual training on higher-quality math & code data, 16K sequences.

6. 20B

   LongCT

   Long-context training on ProLong at 64K tokens.

7. 300B

   Mid-training

   Advanced capability refinement at 32K tokens.

8. Reasoning SFT

   ~8.3M examples yield the Thinking models.

Decoder-only Transformer with RoPE, SwiGLU and sandwich RMSNorm.

Citation

BibTeX

If you find Ouro useful, please cite:

@article{zhu2025scaling,
  title   = {Scaling Latent Reasoning via Looped Language Models},
  author  = {Zhu, Rui-Jie and Wang, Zixuan and Hua, Kai and Zhang, Tianyu and Li, Ziniu and Que, Haoran and Wei, Boyi and Wen, Zixin and Yin, Fan and Xing, He and Li, Lu and Shi, Jiajun and Ma, Kaijing and Li, Shanda and Kergan, Taylor and Smith, Andrew and Qu, Xingwei and Hui, Mude and Wu, Bohong and Min, Qiyang and Huang, Hongzhi and Zhou, Xun and Ye, Wei and Liu, Jiaheng and Yang, Jian and Shi, Yunfeng and Lin, Chenghua and Zhao, Enduo and Cai, Tianle and Zhang, Ge and Huang, Wenhao and Bengio, Yoshua and Eshraghian, Jason},
  journal = {arXiv preprint arXiv:2510.25741},
  year    = {2025}
}