The transition from 5G to 6G wireless systems marks a paradigm shift from “connected things” to “connected intelligence,” driven by the necessity to manage hyper-heterogeneous networks and overcome the Shannon capacity limit. This Systematic Literature Review (SLR) analyzes 118 primary studies to evaluate the transformative impact of Generative AI (GenAI) and Large Language Models (LLMs) on 6G architecture. We categorize the integration of GenAI into five semantic clusters: Architecture, Management, Security, Semantics, and Edge AI. The synthesis reveals that 6G is evolving toward an “AI-Native” ecosystem where LLMs show strong promise for augmenting network orchestration through Intent-Based Networking (IBN) and generative models demonstrate significant potential to augment or transcend traditional physical layer algorithms. Furthermore, the review identifies a fundamental transition from bit-oriented to semantic-oriented communication, utilizing GenAI to reconstruct meaning from minimal data. However, critical challenges remain, particularly the “energy–intelligence paradox” and the risks of model hallucinations in critical infrastructure. We conclude that while GenAI provides the necessary cognitive flexibility for 6G, its successful deployment depends on solving the “inference gap” through split learning and extreme model quantization at the edge.

This study investigates the classification of 5G network slice types using the public ”Network Slicing in 5G”dataset. Weaddressacritical scientific pitfall by identifying and mitigating evaluation leakage caused by near-deterministic rule-encoded binary indicators and heavy data duplication. By excluding these artificial context flags and utilizing only a minimal set of QoS-observable telemetry — specifically packet delay, packet loss rate, time, and equipment category — we establish a rigorous, leakage-aware evaluation protocol. Five representative classifier families were evaluated using a group-safe splitting strategy to ensure results reflect real-world operational conditions. Our experimental results demonstrate that tree-based ensembles significantly outperform linear models, with the strongest ensembles reaching about 95% accuracy (Histogram-based Gradient Boosting at 94.74% and Extra Trees marginally higher at 95.14%). This research underscores that while measurable network telemetry provides sufficient signal for high-accuracy slice recognition, non linear models are necessary to navigate the complex, stochastic overlaps inherent in real wireless environments.