LOTUS uses looped Transformers to make latent reasoning faster and more CoT-like

burny_tech · x · 2026-07-22

The paper “Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers” proposes LOTUS, a looped Transformer that updates multiple hidden thoughts in parallel and supervises them against chain-of-thought steps.

The authors argue this narrows the gap between latent reasoning and explicit CoT at scale. In their results, a 3B model nearly matches explicit CoT on GSM8K while running 2.5× faster, and is 6.9× faster on long reasoning tasks. They also say the latent states can be read back into interpretable reasoning steps, suggesting the representation is compressed rather than opaque.

Original post →

More from Research

Research channel →