Structured Masked Diffusion for Shared-Codebook Unsourced Random Access Decoding
Taekyun Lee*, Jiyoung Yun*, Jeffrey G. Andrews, Hyeji Kim (* equal contribution)
NeurIPS 2026
CIDER combines masked-diffusion refinement, cross-row demixing, and parity-aware propagation to decode shared-codebook unsourced random access messages. In the evaluated setting, it lowers error rates against SIC-BP and FFT-BP and runs 8.2–167× faster than the optimized FFT-BP implementation.