Development

Looped transformers improve reported scaling efficiency

The issue connects changes in model architecture with improved scaling behavior.

development2026-09-20models

Claim

The issue describes research combining looped transformers and model growth with reported compute gains of 1.55x at 1e20 FLOPs and 2.7x at 1e25 FLOPs.

Reported metric: 1.55x; 1e20 FLOPs; 2.7x; 1e25 FLOPs

Reporting date: 2026-09-20 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to September 20, 2026 (theinnermostloop.substack.com)
Looped transformers with model growth raise the scaling exponent (arxiv.org)
1.55x compute at 1e20 FLOPs, 2.7x at 1e25 (x.com)

Indexed from the September 20 newsletter. Linked-source claims have not been independently verified.

Knowledge graphopen in browser →

Reported in

Welcome to September 20, 2026

Referenced by

Welcome to September 20, 2026

Cite or correct this record

Citation formatCitation text

Looped transformers improve reported scaling efficiency
Innermost Loop KB — secondary record
Reporting date: 2026-09-20
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-09-20-looped-transformer-scaling
Claim: The issue describes research combining looped transformers and model growth with reported compute gains of 1.55x at 1e20 FLOPs and 2.7x at 1e25 FLOPs.
Attributed sources:
- Welcome to September 20, 2026 — https://theinnermostloop.substack.com/p/welcome-to-september-20-2026
- Looped transformers with model growth raise the scaling exponent — https://arxiv.org/abs/2609.19107
- 1.55x compute at 1e20 FLOPs, 2.7x at 1e25 — https://x.com/industriaalist/status/2100593211110736240
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.