Apple's LoopCD boosts looped transformers for free: AIME 2024 61.9→73.3, no training needed
burny_tech · x · 2026-10-02
Apple MLR researchers introduce LoopCD, a training-free contrastive decoding framework for looped transformers. Earlier recurrent passes, which standard decoding discards, serve as aligned weak predictions to contrast against the final one — via logit space (LoopCD-Logits) or hidden-state space (LoopCD-Hidden, zero output overhead).
- Ouro-2.6B-Thinking: AIME 2024 pass@1 rises from 61.88% to 73.33%
- Huginn: HumanEval jumps from 22.6% to 31.7%
- Half loop depth matches full depth, cutting forward FLOPs by up to 48.2%
No training or extra models required. The authors also nod to rumors that frontier models like "GPT-6 Astra" and "Gemini 4" may use looped transformers.
Related event: Apple Releases LoopCD: Training-Free Decoding Boosts Looped Transformers(3 posts)→
More from Infra
- Supabase Launches Compute: Long-Running Servers for Heavy Workloads — dshukertjr · 2026-10-03
- Underdog raises from a16z to put private AI on devices you already own — Thom_Wolf · 2026-10-03
- Details emerge on DeepSeek's custom chip: skip-scale optimization cuts power, boosts frequency — teortaxesTex · 2026-10-03
- SkyPilot raises $20M seed to unify GPU scheduling across clouds and clusters — skypilot_org · 2026-10-03
- The Epochalypse: 32-bit systems will think it's 1901 after Jan 19, 2038, 03:14:07 UTC — iannuttall · 2026-10-03
- Dev builds free 3D interactive guide to explain how unified GPU/CPU systems power AI data centers — ghumare64 · 2026-10-03