Paper summary: Scaling Latent Reasoning via Looped Language Models

Problem

Modern LLMs “think” mainly by generating chain-of-thought text, which pushes reasoning into post-training and, the authors argue, under-uses pre-training data.

Method

Ouro (named after the Ouroboros) is a family of Looped Language Models (LoopLM), a close relative of the recurrent-depth idea summarised here in February:

1. iterative computation in latent space: the same layers are applied repeatedly
2. an entropy-regularised objective that learns how much depth to use
3. scaling pre-training to 7.7T tokens

Key results

From the abstract:

• Ouro 1.4B and 2.6B match the results of up to 12B state-of-the-art LLMs across many benchmarks
• controlled experiments suggest the advantage comes from better knowledge manipulation, not greater knowledge capacity
• the looped reasoning traces are more aligned with final outputs than explicit chain-of-thought

Limitations

There’s no dedicated limitations section in the extracted text. Looping adds compute per token at inference, which the paper trades for fewer parameters.