Scaling Latent Reasoning via Looped Language Models

2025-10-29

AI Summary

Current large language models (LLMs) rely on post-training explicit textual reasoning—e.g., chain-of-thought (CoT)—and fail to fully harness latent reasoning capabilities acquired during pretraining. This work proposes LoopLM, the first LLM architecture to incorporate recurrent computation into pretraining, enabling iterative reasoning directly in latent space. To optimize computational resource allocation across layers, we introduce an entropy-regularized deep allocation strategy. Building upon this framework, we develop the Ouro model family (1.4B/2.6B parameters), pretrained on 7.7 trillion tokens. Experiments demonstrate that Ouro matches or exceeds state-of-the-art 12B models on diverse reasoning benchmarks while producing more consistent reasoning trajectories. Crucially, gains stem from enhanced internalized knowledge manipulation—not from increased parameter count or knowledge capacity.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Modern LLMs are trained to "think" primarily via explicit text generation, such as chain-of-thought (CoT), which defers reasoning to post-training and under-leverages pre-training data. We present and open-source Ouro, named after the recursive Ouroboros, a family of pre-trained Looped Language Models (LoopLM) that instead build reasoning into the pre-training phase through (i) iterative computation in latent space, (ii) an entropy-regularized objective for learned depth allocation, and (iii) scaling to 7.7T tokens. Ouro 1.4B and 2.6B models enjoy superior performance that match the results of up to 12B SOTA LLMs across a wide range of benchmarks. Through controlled experiments, we show this advantage stems not from increased knowledge capacity, but from superior knowledge manipulation capabilities. We also show that LoopLM yields reasoning traces more aligned with final outputs than explicit CoT. We hope our results show the potential of LoopLM as a novel scaling direction in the reasoning era. Our model could be found in: http://ouro-llm.github.io.

Problem

• Enhancing latent reasoning through iterative computation
• Improving knowledge manipulation over capacity expansion
• Aligning reasoning traces more closely with outputs

Innovation

• Iterative latent space computation for reasoning
• Entropy-regularized objective for depth allocation
• Scaled pre-training with 7.7 trillion tokens