Where models change their minds: Identifying branchpoints for NLA training

Exploring internal patterns behind shortcut behavior in LLMs using branchpoint analysis.