TrustTwin: World-Model-Enhanced Twins for Trustworthy AIoT Decision Making

Abstract

Recent advances in large language models are pushing IoT systems toward autonomous Artificial Intelligence of Things (AIoT) paradigms, where system behavior depends on protocol semantics, internal state, interaction history, and environmental context. This creates a gap between AI-generated decisions and the behavior of physical entities: a message may be protocol-valid yet unsafe in the current state. We present TrustTwin, a framework for trustworthy decision making based on specification-guided world models and executable twin validation. TrustTwin extracts device actions, states, preconditions, postconditions, and constraints from protocol specifications to evaluate candidate AI actions under the current context. It then validates predicted outcomes against an executable twin and triggers adaptive recovery when discrepancies occur, including revising, delaying, rejecting actions, or updating the model. We instantiate TrustTwin in the Matter ecosystem and show that it improves decision quality, reduces contextually inappropriate actions, and provides traceable feedback for adaptive control.