Researchers present LoopCD, a training-free contrastive decoding framework for looped transformers that contrasts a model’s final prediction against an earlier recurrent pass to pick better tokens, without auxiliary models or extra training. Across four looped transformer families, the technique delivers consistent gains: one variant raised Ouro-2.6B-Thinking’s AIME 2024 pass@1 from 61.88% to 73.33%, while another lifted Huginn’s HumanEval pass@1 from 22.56% to 31.71%. The gains are large enough that the number of recurrent loops can be halved while still matching full-depth baselines, cutting forward FLOPs by 22.5% to 48.2%.