Test-time interventions help some batches but hurt others—PORTAL solves this by learning when to apply them through smart label acquisition, turning risky interventions into reliable tools that adapt to changing conditions.
PORTAL is a controller that decides when to apply test-time interventions (like updating model statistics) to handle changing data. Instead of always using an intervention or never using it, PORTAL intelligently selects which batches benefit from it by strategically acquiring labels and comparing candidate prediction rules.