Predictive Memory Localization (PML) identifies selective operating regimes for activation steering by separating semantic-neighbor movement from capability damage. Testing on 30,000 paths showed geometry-derived directions achieved a 13.1% target-any success rate, outperforming random baselines.
HOW THIS AFFECTS YOU
●
researcherPML provides a more granular framework for understanding how specific internal activations map to semantic control.
●
policyThis method improves our ability to identify and intervene in model behaviors via localized internal signals.