NCP: Verifier-Guided Optimization of Weight-Space Control Policies
Summary
- Researchers introduce the Neural Control Plane (NCP), a closed-loop system that observes language model execution and synthesizes candidate weight-space intervention rules.
- NCP separates an inference data plane from the control policy, utilizing W-ISA for transactional execution and parameter restoration after each request.
- In local experiments with SmolLM-135M-Instruct, NCP generated three candidates that reduced repetition metrics but failed strict admission criteria, leading to archival rather than deployment.
Why it matters
NCP provides an executable framework for studying recursive self-improvement and dynamic interventions at the model weight level while maintaining strict verification gates.