Looped language models deliver better reasoning with less training data
Closing the Loop: Practical Training Recipes for Looped Language Models
Computation and LanguageArtificial Intelligence
Summary
Training language models to reason better usually requires huge amounts of data and compute. The authors show a way to train a special kind of model, called a looped model, more efficiently, using far less data while still improving performance on math and reading tests. They also demonstrate how to upgrade already trained models into looped ones easily, gaining better results without training from scratch. This makes these models cheaper and more practical to use.
What this means in practice
- •For ai model trainers: Train language models that reason better using less data and compute by adopting efficient looped training recipes.
- •For platform engineers: Upgrade existing language model checkpoints into looped models for improved reasoning performance without full retraining.
- •For financial analysts: Use more parameter-efficient language models with improved math reasoning to automate complex calculations and report generation.$Commercial implications: Enables selling advanced automated financial tools leveraging improved, smaller models for precise numerical reasoning.
Authors
Andrei Marchenko, Viacheslav Bezrukov, Oleg Kashurin, Inessa Fedorova, Dmitry Bocharov, Yuliana Shakhvalieva, Maria Tikhonova, Valerii Ternovskii
Abstract
Looped language models increase effective depth by repeatedly applying a shared block of layers, but existing large-scale recipes require multi-stage training over trillions of tokens, while the benefits of recurrence remain difficult to separate from differences in data and training. In this work, we establish practical training recipes for looped language models, with three main results. (1) We develop a compute-efficient from-scratch pipeline that reduces the training budget from 7.7T tokens in Ouro to 310B tokens while retaining strong reasoning performance. Pretraining followed by high-quality mid-training, together with learning-rate warmup and stronger exit-gate regularization, enables stable recurrent training without prior multi-stage schedules. (2) Under controlled comparisons, our 1.4B LoopLM outperforms a parameter-matched dense model trained on the same data and token budget on all 12 evaluated benchmarks, including +14 points on GSM8K, +10 on MATH, and +22 on DROP. At matched inference compute, it approaches a 3.9B dense model on mathematical reasoning and reading comprehension while using only 36\% as many parameters. (3) We introduce a minimal recipe for converting pretrained dense models into looped ones: a single learned input-mixing scalar and a smoothed exit loss, with no step-specific parameters. Applied to Qwen3-1.7B-Base, Looped Qwen improves over an identically continued dense baseline on every evaluated benchmark across two data regimes, with statistically clear gains on GSM8K, MATH, and MMLU-Pro on the curated mixture. Together, these results make looped language models substantially cheaper to train from scratch and practical to introduce into existing pretrained checkpoints, while isolating the gains due to recurrence itself.