LLR-based attention-weighted projection decoding for sparse vector-coded short packet transmissions