LOTUS: Looped Transformer with Parallel Latents

burny_tech · x · 2026-07-19

Researchers introduced LOTUS (Looped Transformers with parallel supervision on latents). This method inserts K latent blocks between the model's question and answer, running R loops of computation with the base model to obtain the final latents.

Training uses two simple loss functions: first, a standard cross-entropy loss on the chain-of-thought via the model's own LM head; second, a standard next-token prediction based on these latents to derive the final answer.

Original post →

More from Research

Research channel →