Research map / Image generation / Fast diffusion sampling

Structured Masked Diffusion for Joint Multiuser Decoding

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Masked diffusion language model decoding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CIDER, a learned masked-diffusion multiuser decoder with demixing and parity-aware propagation for unsourced random access, matching FFT-BP accuracy while being 6x to over 100x faster.

Problem

Joint multiuser decoding in unsourced random access: recovering an unordered set of K codewords from a noisy superposition given slot-wise soft evidence, where classical SIC, joint BP, and list-recovery decoders are brittle or slow as ambiguity grows.

Approach

Masked discrete diffusion refinement over a K x L grid conditioned on AMP soft evidence, with a demixing module that softly allocates each slot-symbol candidate across competing rows (preventing duplicate-row collapse) and a parity-aware module that passes extrinsic Tanner-graph check messages to keep rows code-consistent; optional quality-guided remasking at high load.

Key results

• At L=12, CIDER reaches SER 0.0011 / CER 0.0073 in 1.26 ms vs FFT-BP SER 0.0015 / CER 0.0078 in 8.34 ms.
• At L=48, CIDER achieves SER 0.0045 in 7.66 ms while FFT-BP needs 767.51 ms and SIC-BP 8604.10 ms, with FFT-BP/SIC-BP SER 0.1144.
• Generic masked diffusion (MDD) and one-shot neural baselines collapse to SER around 0.4 at K=2 in the module ablation, while the full CIDER reaches SER 0.0011.

Surprising findings

• Generic masked diffusion (MDD) and one-shot neural decoders collapse to near-total failure (CER near 1) at K=2, showing neural decoders alone do not solve the problem.
• Iterating the A+B modules without masked diffusion (N_iter=12) degrades performance sharply (SER 0.3863 vs 0.1013 one-shot).

Limitations

• Evaluated only under a simulated AMP-based detector and AWGN-style URA model; joint two-stage training and richer channels are future work.
• Scaling to larger loads requires stochastic binning and PRISM-style remasking; CER rises to about 0.26 at K=8 without remasking.
• Focuses on sparse-graph codes (LDPC, tree-code) because the parity module relies on Tanner-graph propagation.

Experimental setup

Benchmarks: simulated URA GF(64) LDPC decoding (K=2, L in {12,18,24,48}) stochastic-binning protocol up to K_tot=100

Metrics: symbol error rate codeword error rate wall-clock time (ms/sample)

Baselines: SIC-BP FFT-BP Top-J search MLP CNN Transformer GNN NBP Tanner-Attention MDD Stitch

Builds on: masked discrete diffusion (MDM) PRISM belief propagation AMP-MMSE detector

Closest papers by meaning

• Learning Linear Block Error Correction Codes (ICML 2024 · similarity 0.82)
• DMax: Aggressive Parallel Decoding for dLLMs (NeurIPS 2026 · similarity 0.81)
• Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.81)
• Friendly Attacks to Improve Channel Coding Reliability (AAAI 2024 · similarity 0.80)
• Efficient Minimum Bayes Risk Decoding using Low-Rank Matrix Completion Algorithms (NeurIPS 2024 · similarity 0.80)
• Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.80)