DeepMind's Recirculation Enables Recursive Self-Improvement Without Retraining

omarsar0 · x · 2026-08-23

Google DeepMind proposes 'Recirculation,' a training-free approach to evolve model architectures. By feeding activations back during prefill, the model acts as a dynamical system at inference time, tracking belief states without retraining. This addresses the layer-depth limitation of feedforward transformers in long generations. Generation cost remains flat as serial work is confined to the prefill phase.

Original post →

More from Research

Research channel →