Research2026-10-01
Researchers introduced LoopCD, a training-free decoding method for looped Transformers. It contrasts final predictions with earlier loop states. LoopCD-Logits and LoopCD-Hidden are its two variants. LoopCD-Logits adds one extra output pass. LoopCD-Hidden adds zero output overhead. It improves performance across four looped Transformer families. It enables halving recurrent loops while matching baseline quality. This reduces inference compute by up to 48.2%. The paper was submitted to arXiv on 1 October 2026.
What changed
Standard decoding discarded earlier loop states in looped Transformers.
What it unlocks
Using earlier loop states as guidance signals to improve decoding quality without extra training.
- AIME 2024 pass@1 61.88% to 73.33%
- HumanEval pass@1 22.56% to 31.71%
- 22.5% to 48.2% forward FLOPs reduction
Sources