Looped Transformers explained: when is another pass worth the cost?

TheTuringPost · x · 2026-11-08

The Turing Post published a guide on the recent research boom around Looped Transformers — architectures that reuse Transformer layers instead of stacking distinct ones.

Core idea: a standard Transformer uses 24 distinct weighted blocks passed once; a looped version can use 6 blocks cycled 4 times — same total block applications, far fewer parameters.
The lineage traces back to the Universal Transformer with ACT (Graves, 2016) dynamic halting, deciding steps T per position.
The guide covers early algorithms through adaptive depth, and why reusing layers changes computation, reasoning, and efficiency as the AI stack matures.

Original post →

Looped Transformers explained: when is another pass worth the cost?

More from Research