Uniform Scaling Limits in AdamW-Trained Transformers

William Gibson, Christoph Reisinger · 2026

Read Uniform Scaling Limits in AdamW-Trained Transformers by William Gibson, Christoph Reisinger (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-11059.txt.

Description

William Gibson and Christoph Reisinger study uniform scaling limits in AdamW-trained transformers. The team models hidden-state dynamics as an interacting particle system coupled by attention. Joint dynamics of hidden states and backpropagated variables converge in L2, uniformly over starts, to a forward-backward ODE system. The rate is order L to the minus 1 plus L to the minus one-third times H to the minus one-half, with L as depth and H as head count.

The limiting ODE system matches a McKean-Vlasov ODE when attention heads lack causal masking. Using flow maps of this McKean-Vlasov ODE and measure concentration, the team bounds discrete-continuous model differences uniformly over compact initial-condition sets. Without a covering argument, constants are independent of token count. Under an AdamW adaptation, bounds also ignore token embedding dimension. Transformer theorists get uniform scaling limits with token-independent constants.

Catalog record

CREATORWilliam Gibson, Christoph Reisinger
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE258504
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.11059

Browse the full library · About the Tower of Babel