6G CSI-native models cut pilot overhead 93%, boost spectral efficiency 36.6%
New channel-adaptive framework reduces NMSE by 4 dB across multiple tasks...
Researchers from an unnamed institution (likely led by Chenyu Zhang) have published a paper proposing CSI-native foundation models for 6G wireless systems. The key innovation is a channel-adaptive roadmap that treats channel state information (CSI) as propagation-conditioned responses rather than generic tensors. The framework introduces three requirements: scale-aware heterogeneous exposure, physical time-frequency-antenna coordinates, and correlation-bounded token interaction. This aligns pretraining, positional modeling, and attention control with the intrinsic geometry of wireless environments, enabling reusable intelligence across diverse deployment scenarios.
Extensive experiments demonstrate superiority across three dimensions: zero-shot generalization (4 dB NMSE reduction across spatial-temporal-frequency tasks), scale extrapolation (5.4 dB gain under 8× unseen antenna scaling), and inference efficiency (18.8% acceleration for mobility-aware processing). A system-level evaluation using Sionna SYS shows the framework uses only 7.01% of dense-pilot overhead while achieving -18.64 dB average NMSE. It improves average net spectral efficiency by 36.6% over dense LMMSE and 15.5% over WiFo, proving that CSI-native representation learning can support pilot-efficient radio access for 6G. The paper is submitted to IEEE Wireless Communications Magazine.
- Zero-shot NMSE reduction of over 4 dB across spatial, temporal, and frequency tasks
- Scale extrapolation yields 5.4 dB gain at 8× unseen antenna scaling
- Achieves 36.6% better spectral efficiency with only 7.01% pilot overhead vs. dense LMMSE
Why It Matters
Enables pilot-efficient 6G networks that adapt to real propagation geometry, cutting overhead while boosting spectral efficiency.