- Who it helps
- Researchers exploring efficient model architectures
- Availability
- November 4 paper is public and links code and a project page. Those artifacts were not executed.
Repeated computation with shared parameters
Looped transformers apply the same blocks more than once. Loopback separates the original input from subsequent updates and learns how to weight those updates.
Memory savings have a compute tradeoff
Fewer stored parameters do not automatically mean less computation. The authors compare models at matched training tokens and executed depth; their reported accuracy and loss remain results for those experiments.
What you can do
- Read the primary description and its stated limits before planning an evaluation. Open reference ↗
Where the evidence stops
Abstract and public explanatory introduction checked November 12. No training or downstream evaluation reproduced.
Sources read. Public source text checked November 12, 2026 UTC. No third-party project or full workflow executed.
Sources & further reading
Paper · paper
Loopback Transformer paper ↗
November 4 abstract read. Shared layers reduce parameter count; reported results remain author evidence.
Xuan Ouyang and coauthors · page checked 2026-11-12 · Published 2026-11-04
Web · analysis
Looped transformer explanation ↗
Public introduction read; later subscriber-only content was not accessed.
Ksenia Se and coauthor · page checked 2026-11-12
Coverage note: Abstract and public explanatory introduction checked November 12. No training or downstream evaluation reproduced.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Loopback Transformer reuses model layers and controls how their earlier updates contribute to later passes. It studies how a smaller parameter set can support repeated computation.