Code for ICML 2026 paper experiments studying how deep sequence models tend to memorize through a learned geometric structure.
transformers representation-learning path-finding memorization icml sequence-models graph-learning large-language-models mechanistic-interpretability spectral-bias icml-2026 geometric-memory
-
Updated
May 19, 2026 - Python