Towards CSI-Native Foundation Models: A Channel-Adaptive Roadmap for 6G

The paper introduces a unified framework for channel-adaptive foundation models tailored for 6G systems, addressing limitations in current CSI pretraining methods. Key advancements include significant improvements in zero-shot generalization (over 4 dB NMSE reduction), scale extrapolation (up to 5.4 dB gain with 8 times unseen antenna scaling), and enhanced inference efficiency (up to 18.8% acceleration). This framework demonstrates a substantial reduction in pilot overhead to 7.01% while achieving an average NMSE of -18.64 dB and a 36.6% increase in spectral efficiency, highlighting its potential for efficient radio access in future wireless networks.