A geometric account of why post-alignment fine-tuning can pull model behavior back toward earlier training-history manifolds.
Jun 26, 2026