This article presents DISA (Dynamic Immune-Shunt Architecture), a conceptual research proposal for detecting potentially deceptive behavior in large language models through sparse internal monitoring rather than relying solely on observable output. The framework aims to address the risk that capable models might behave differently when they recognize they are being monitored.
DISA consists of four main components:
- ASAM (Asynchronous Sparse Activation Monitoring) samples selected internal states like residual activations and attention representations to reduce computational overhead compared to continuous monitoring.
- MDRE (Multi-Signal Deception Risk Estimation) combines multiple signals, including activation-probe evidence, behavioral inconsistency, and uncertainty, to estimate risk without confusing hallucination with deception.
- RIA (Reversible Intervention Adapter) applies temporary interventions, such as steering directions or low-rank adapters, to reduce targeted behavior while preserving model capabilities.
- IAI (Isolated Audit Infrastructure) supports the evaluation process using a four-category dataset that independently controls triggers and target behaviors to prevent memorization.
The proposal emphasizes that DISA is a falsifiable framework intended to be tested, criticized, or rejected, rather than a definitive solution to deceptive alignment.