Apple's LoopCD: Training-Free Decoding Boosts Looped Transformers
Apple's MLR team introduced LoopCD, a training-free contrastive decoding framework for looped Transformers. It lifts AIME accuracy from 61.9% to 73.3% while halving FLOPs, boosting performance almost for free.
2026-10-02 ~ 2026-10-03 · 4 related posts
- LoopCD: training-free contrastive decoding lifts looped transformers, AIME 61.9%→73.3% — arankomatsuzaki · 2026-10-02
- Apple's LoopCD boosts looped Transformers 11+ points while cutting FLOPs up to 48% — apple · 2026-10-02
2 near-duplicate retellings: burny_tech · kastnerkyle