Sparse Internal Control via Inference-Time Node Interventions

The Sparse Internal Control framework enables selective, reversible steering of frozen LLMs by applying interventions to a small set of internal model nodes. The method uses Driver-OMP and Coverage-Driver selectors to identify candidate sites that minimize intervention energy while maximizing target controllability.