IAAIDecoding Huginn's Latent CoT
Latent Chain-of-Thought? Decoding the Depth-Recurrent Transformer
Probes Huginn-3.5B on arithmetic with the Logit Lens and Coda Lens to see whether interpretable latent chain-of-thought emerges across recurrent blocks.
Inside the method
Prelude → [ Core ] × r (probed each step) → Coda
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Core recurrence
- Depth control
- Analysis
- KV / state strategy
- See paper
Reading note
Finds limited evidence of interpretable latent CoT and inconsistent probing across blocks; more recurrence gives only marginal gains, well short of explicit CoT on these tasks.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{huginnlatentcot2025,
title = {Latent Chain-of-Thought? Decoding the Depth-Recurrent Transformer},
author = {Wenquan Lu and Yuechuan Yang and Kyle Lee and Yanshu Li and Enqi Liu},
year = {2025},
eprint = {2507.02199},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2507.02199}
}