Preprint / scholarly article

A Model-Agnostic, Performance-Pushforward Theory of Scaling Laws

K. Takahashi

Published
DOI
10.5281/zenodo.17520859

Abstract

The central idea is to view performance as the pushforward of an EVI λ gradient flow under a Lipschitz evaluation map, while scale is measured by the geometry of preimages of performance targets.

Keywords

  • large language models
  • AI
  • machine learning
  • optimization
  • information theory
  • numerical analysis
  • computer science
  • scaling laws
  • gradient flows
  • ambrosio-gigli-savaré
  • evi

Identifiers and source records