Jul 2026
Conditional Optimal Bridge for Riemannian Activation Steering
This formulation yields the first principled derivation of the log-density-ratio steering objective from a well-posed optimization problem, and consistently outperforms prior activation steering baselines while avoiding the OOD degradation commonly observed in existing methods.
Seyed Arshan Dalili, Ajay Narayanan Sridhar, V. Narayanan et al.
· arXiv.org · 1 citation