Paper: Transformers Converge to Invariant Algorithmic Cores Across Models

A new paper by Joshua S. Schiffman introduces Algorithmic Core Extraction (ACE), showing that independently trained Transformers converge to compact, invariant subspaces that are necessary and sufficient for a task and can be extracted and manipulated across models.

2026-11-20 ~ 2026-11-20 · 2 related posts