IAAILLoCoT
From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers
Replaces left-to-right latent thought generation with iterative refinement of a compact latent workspace by a shared transformer. A probabilistic head then samples latent tokens in parallel to condition an autoregressive answer decoder.
Inside the method
Prompt + latent slots → [ Shared transformer ] × k → sampled latents → AR decoder
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Small fixed number of refinement iterations
- KV / state strategy
- See paper
Reading note
On HumanEval and MBPP it matches explicit-CoT SFT accuracy while cutting time to first answer token about 36× and reasoning-phase latency about 42×.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{llocot2026,
title = {From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers},
author = {Gerard Grau García and Arnau Padrés Masdemont and Niccolò Grillo and Jordi Ros-Giralt and Arash Behboodi and Victor Conchello Vendrell},
year = {2026},
eprint = {2610.11472},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.11472}
}