Ouro LoopLMs: Open LLMs with Latent Reasoning

Описание: In this AI Research Roundup episode, Alex discusses the paper:
'Scaling Latent Reasoning via Looped Language Models'
Ouro introduces open-source Looped Language Models that build iterative latent reasoning into pretraining, reducing reliance on tokenized Chain-of-Thought at inference. The approach applies weight-tied Transformer blocks recurrently with learned exit gates, trained via an entropy-regularized objective and a two-stage gate schedule. Pretrained on 7.7T tokens with up to 4 latent iterations, Ouro 1.4B/2.6B models achieve strong parameter efficiency, matching larger baselines. The paper also evaluates practical inference tricks like KV cache reuse and early-exit criteria.
Paper URL: https://arxiv.org/abs/2510.25741

#AI #MachineLearning #DeepLearning #LLM #LatentReasoning #RecurrentTransformers #OpenSource #InferenceEfficiency