DeepLoop Explores Depth Scaling for Looped Transformers

DeepLoop investigates depth scaling for Looped Transformers, revealing that residual branch scaling must account for the number of parameter visits rather than just nominal layers. This aligns with rumors that some frontier models operate similarly to looped architectures.

2026-07-17 ~ 2026-07-18 · 2 related posts

• DeepLoop: Depth Scaling for Looped Transformers — Shuzhen Li · 2026-07-17
• Looped Transformer Scaling Research — BlackHC · 2026-07-18