To advance Polar code design for 6G applications, we develop a learning-based universal sequence design framework that is scalable and adaptable to various channel conditions and decoding techniques. Most importantly, our method achieves a code length of up to 2048, making it suitable for deployment. In all of them configuration supported in 5G, our method achieves competitive performance relative to the NR sequences obtained in 5G and produces up to 0.2 dB gain over the beta broadening baseline in . We also highlight the main features that have enabled learning at scale: (i) the inclusion of natural law delay learning based on the universal location of Polar codes, (ii) the exploitation of the weak long-term effect of decisions to limit the look-ahead test, and (iii) the joint development of multiple integrations to increase the learning efficiency.
