LoopLM (Looped Language Model)

LoopLM (Looped Language Model) is the name for a pretrained language model that builds iterative, weight-shared computation into pretraining itself, rather than adding it afterward. Rui-Jie Zhu and 32 co-authors, including Yoshua Bengio, introduced the term with Ouro, named after the recursive Ouroboros, in "Scaling Latent Reasoning via Looped Language Models" (arXiv:2510.25741, first posted October 2025, revised through July 2026). The underlying mechanism, applying the same block of layers to a hidden state more than once, is the same one behind the vocab's existing Looped Transformer and Recurrent Depth entries. LoopLM is the name that stuck for models trained from scratch this way, rather than converted or fine-tuned into it.

Ouro combines iterative computation in latent space with an entropy-regularized objective that learns how many loop iterations to allocate per token. It was trained on 7.7 trillion tokens. The 1.4B and 2.6B Ouro models match benchmark results from LLMs up to 12B parameters. The paper attributes this to better knowledge manipulation rather than larger knowledge capacity, and reports that LoopLM's internal reasoning traces line up with its final answers more closely than explicit chain-of-thought does.

LoopLM is now the standard label for this architecture class in follow-up work, including LoopMoE, LoopFormer, DeepLoop, and looped state-space models. A September 2026 Q Labs post on scaling "computational depth," by Akshay Vegesna and Samip Dahal, cites looped models as a direction for scaling depth indefinitely without adding stored weights, distinct from the compute-optimal depth techniques the post introduces.