LoopCD Improves Looped Transformer Decoding Without Extra Training
September 30, 2026
LoopCD is a training-free contrastive decoding framework that boosts Looped Transformer performance by contrasting final predictions with earlier recurrent passes. It operates in either logit space or hidden-state space with zero to minimal inference overhead.
HOW THIS AFFECTS YOU
●
builderYou can increase the performance of parameter-efficient looped models at inference time with almost no cost.
●
researcherThis demonstrates how to exploit inherent recurrence in looped architectures for better token selection.