Papers with Code spotlights Looped Transformer: recurrent depth decoupled from parameters

NielsRogge · x · 2026-09-02

Niels Rogge highlighted the Looped Transformer topic page on Papers with Code. The method reuses Transformer blocks across recurrent depth, decoupling effective computation depth from parameter count, enabling iterative refinement, adaptive computation, and test-time depth scaling. Collected papers span HRM-Text, Nanbeige4.2-3B (agentic capabilities in a compact model), LoopViT, SMELT scaling laws for MoE looped transformers, Gated Recurrent Transformers, DeepLoop, learned halting gates, looped models improving compositional tool calling, and more.

Original post →

More from Research

Research channel →