IAAILoop-wise Contrastive Decoding
LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Models
Links reasoning errors in looped language models to loop instability that produces uncertain hard tokens, and contrasts logits from earlier iterations with the final iteration at decoding time to refine them, with no training and negligible overhead.
Inside the method
logits(final loop) contrasted with logits(earlier loop) → sampling distribution
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Fixed loops of the base model
- KV / state strategy
- See paper
Reading note
Reports improved reasoning across recent representative looped models and tasks. Distinct from the similarly named LoopCD of Liu et al. (2610.02185).
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopcdloopwise2026,
title = {LoopCD: Loop-wise Contrastive Decoding for Improving Reasoning in Looped Language Models},
author = {Byeongho Yu and Junhyuk So and Eunhyeok Park},
year = {2026},
eprint = {2609.24196},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.24196}
}