Semantic Drift in AI-Native 6G Multi-Agent Systems: A Systematic Review and Taxonomy of Mitigation Strategies

AI-native sixth-generation (6G) wireless networks are envisioned to orchestrate massive populations of autonomous agents—including unmanned aerial vehicles (UAVs), connected autonomous vehicles (CAVs), industrial robotics, and distributed Internet of Things (IoT) nodes—that transcend traditional bit-pipe transmissions to exchange rich, task-oriented semantic representations. Extracted via advanced edge artificial intelligence (AI), transmitted across learned air interfaces, fused into shared knowledge bases, and ultimately translated into collaborative joint actions, this semantic processing pipeline is inherently vulnerable to compounding imperfections.Sensors degrade through calibration loss and environmental wear, edge inference models suffer from severe distribution shifts, semantic encoders and decoders experience mutual misalignment, wireless channels distort latent feature spaces, knowledge bases diverge across heterogeneous vendors, and malicious or faulty Byzantine peers inject persistent false beliefs. We formally define the resulting gradual degradation of alignment between representation and reality, as well as inter-agent consensus divergence, as semantic drift.This comprehensive review traces the entire sensing-to-decision pipeline across seven distinct stages(S1→ S7) . We introduce a rigorous stage-wise formalization accompanied by an end-to-end analytical error bound demonstrating how localized errors propagate and amplify downstream. Furthermore, we present a structured detect–prevent–correct–contain taxonomy of mitigation mechanisms for each stage, evaluate numerical models contrasting vulnerable linear consensus with Byzantine-resilient fusion, outline a closed-loop cross-stage management architecture with joint resource allocation formulations, consolidate evaluation metrics, and articulate open research challenges.