IAAILoopRPT
LoopRPT: Reinforcement Pre-Training for Looped Language Models
Reinforcement pre-training for looped language models that reframes next-token prediction as next-token reasoning and assigns rewards directly to latent steps using an EMA teacher reference and noisy latent rollouts.
Inside the method
[ Shared stack ] × R with RL signals on latent steps
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Compresses reasoning into fewer iterations
- KV / state strategy
- See paper
Reading note
Instantiated on Ouro at several scales; improves per-step representation quality, gives Pareto gains in accuracy versus computation, with large gains on hard tokens.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{looprpt2026,
title = {LoopRPT: Reinforcement Pre-Training for Looped Language Models},
author = {Guo Tang and Shixin Jiang and Heng Chang and Zihan Zhang and Nuo Chen and Yuhan Li and Huiming Fan and Jia Li and Ming Liu and Bing Qin},
year = {2026},
eprint = {2603.19714},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2603.19714}
}