Neural Polar Decoder Overview

• Neural Polar Decoder (NPD) is a family of neural network-based decoders for polar codes that retain polar structure while learning to map soft channel observations directly to information bits.
• They integrate neuralized SC/SCL, graph-structured modules, partitioned decoders, and neural BP techniques to overcome latency and scalability issues inherent in traditional decoding.
• NPD methods enable efficient decoding across various channel conditions, including AWGN, deletion, and synchronization-error channels, and support advanced applications like DNA storage and pilotless wireless reception.

Neural Polar Decoder (NPD) denotes a family of neural-network-based decoders for polar codes. In one widely used sense, an NPD is a one-shot, highly parallel decoder that maps soft channel observations directly to the information bits; in another, it is a neuralized version of successive-cancellation (SC) or successive-cancellation list (SCL) decoding in which analytic channel embeddings and factor-graph operations are replaced by trainable neural modules (Cao et al., 2019, Aharoni et al., 18 Jun 2025). Across these formulations, the defining property is not a single architecture but the retention of polar-code structure—generator matrix, frozen-bit pattern, SC/BP recursion, or Plotkin tree—while learning the decoding map, the message updates, or both. The resulting literature spans residual denoise-then-decode receivers, learned SC/SCL message passing, neural BP on sparse Tanner graphs, partitioned and concatenated sub-decoders, matched decoders for learned nonlinear kernels, mutual-information estimators for unknown channels, and synchronization-error decoders for deletion and IDS channels (Cammerer et al., 2017, Hebbar et al., 2024, Aharoni et al., 20 Jun 2025, Aharoni et al., 16 Jul 2025).

1. Definition and decoding foundations

Polar coding starts from the Arıkan kernel

$$F=\begin{bmatrix}1 & 0\ 1 & 1\end{bmatrix},$$

and, for blocklength $N=2^n$, the generator matrix

$G_N = B_N F^{\otimes n}.$

A source vector $u$ contains information bits on a reliable index set and frozen bits fixed to zero elsewhere, and encoding is $x=uG_N$. Under BPSK and AWGN, received symbols produce per-symbol log-likelihood ratios, and classical SC decoding combines these through the standard $f$ and $g$ recursions while traversing the polar factor graph sequentially (Cao et al., 2019, Aharoni et al., 18 Jun 2025).

This sequential structure is the immediate motivation for NPDs. SC has low $O(N\log N)$ complexity, but its strictly sequential nature limits throughput and increases latency at finite blocklengths. SCL and CRC-aided SCL improve finite-length performance by maintaining multiple candidate paths, but they preserve the same basic sequential dependence and path-management overhead. NPDs were introduced precisely to alter this trade-off: either by learning a direct parallel map from channel observations to information bits, or by preserving the SC/SCL graph while replacing closed-form operations with neural components that are trained from samples (Cao et al., 2019, Hirsch et al., 3 Oct 2025).

The literature therefore uses “Neural Polar Decoder” in a structurally broad but technically coherent way. One-shot NPDs treat decoding as classification or regression from soft inputs to information bits. Graph-structured NPDs preserve the recursive decomposition of the polar transform and learn channel embeddings, check-node updates, bit-node updates, or leaf decision maps. BP-derived NPDs unfold iterative decoding into a trainable network. More recent work extends the same idea to channels with memory and synchronization errors, where analytic LLR rules are unavailable or computationally prohibitive (Xu et al., 2018, Aharoni et al., 20 Jun 2025).

2. Major architectural families

One-shot feed-forward decoders map noisy channel outputs directly to the $K$ information bits in a single forward pass. A representative example is the residual neural network decoder (RNND), which inserts a residual denoiser in front of an MLP, CNN, or LSTM decoder. The denoiser learns a residual mapping $r(y;\theta)$ and outputs $N=2^n$0; the decoder then predicts only the $N=2^n$1 information bits, with frozen-bit positions enforced implicitly by the output dimension. The RNND was evaluated for a short polar code with $N=2^n$2, $N=2^n$3, using MLP, 1-D CNN, and LSTM back-ends (Cao et al., 2019).

Graph-structured SC/SCL NPDs replace the analytic SC/SCL primitives by shared neural modules. In this formulation, the decoder uses a channel embedding network $N=2^n$4, a check-node network $N=2^n$5, a bit-node network $N=2^n$6, and an embedding-to-LLR network $N=2^n$7, all reused across the polar graph. At each decision point, $N=2^n$8 produces an LLR, and the same outputs can drive SCL path metrics and pruning. This architecture was used not only for decoding but also for mutual-information estimation and input-distribution optimization over black-box channels, including Honda–Yamamoto shaping and SCL decoding under non-uniform inputs (Aharoni et al., 18 Jun 2025).

Partitioned and concatenated decoders attack the scalability problem of monolithic neural decoding. One approach partitions the polar factor graph into smaller sub-blocks, trains a neural decoder for each sub-block, and connects them through remaining conventional BP stages. This reduces training complexity from a dependence on the full $N=2^n$9 message space to a sum over smaller $G_N = B_N F^{\otimes n}.$0 subproblems and yields a non-iterative, highly parallel decoder (Cammerer et al., 2017). A related approach recursively concatenates short trained subcode decoders into a larger network by adding at most three small layers that perform left-LLR propagation, re-encoding through XOR, and right-LLR propagation. In that construction, the larger network starts from SC-level BER without end-to-end retraining and can then be fine-tuned toward ML behavior (Stupachenko, 2022).

BP-derived NPDs unfold message passing into a trainable neural network. The sparse neural network decoder (SNND) first converts the polar factor graph to a sparse LDPC-like Tanner graph, unfolds $G_N = B_N F^{\otimes n}.$1 BP iterations into $G_N = B_N F^{\otimes n}.$2 layers, and replaces SPA by min-sum with learnable normalization. A single learned scalar $G_N = B_N F^{\otimes n}.$3 is sufficient in the most compressed version, and the resulting network achieves near-SPA performance with one trainable parameter (Xu et al., 2018). A distinct recurrent neural BP decoder ties the edge-scaling parameters across iterations and then applies codebook-based quantization to the learned weights, reducing memory and arithmetic cost while retaining fast convergence (Teng et al., 2018). Another line, the gated hypernetwork decoder, generates per-edge update functions conditioned on current message magnitudes and blends them with classical BP through a learnable damping factor (Nachmani et al., 2019).

Assistive and matched neural decoders extend the notion of NPD beyond direct replacement of SC/BP primitives. CNN-assisted bit flipping leaves the BP core intact but uses a CNN on the full collection of BP metadata to predict high-probability flip positions, thereby guiding CRC-checked restarts (Teng et al., 2019). DeepPolar goes further by learning nonlinear large-kernel polar transforms and pairing them with a matched neural SC decoder that mirrors the learned kernel recursion; the decoder is therefore “polar” not because the kernel remains linear, but because the recursive Plotkin-tree structure is preserved (Hebbar et al., 2024).

3. Learning objectives and supervision regimes

The earliest and still common regime is supervised learning from synthetic channel data. RNND uses a multi-task objective

$G_N = B_N F^{\otimes n}.$4

where the denoiser minimizes MSE to the transmitted BPSK symbols and the decoder minimizes MSE on the $G_N = B_N F^{\otimes n}.$5 information bits. For the $G_N = B_N F^{\otimes n}.$6 code, all $G_N = B_N F^{\otimes n}.$7 codewords were used in training, with train-SNR fixed at $G_N = B_N F^{\otimes n}.$8 dB, batch size $G_N = B_N F^{\otimes n}.$9, Adam, learning rate $u$0, and $u$1 epochs over the codebook (Cao et al., 2019). In practical-system NPDs for OFDM and single-carrier links, stage-wise cross-entropy is accumulated over all graph depths through NSCLoss, and the model jointly trains the received-signal embedding and a learned constant embedding for punctured positions (Hirsch et al., 3 Oct 2025). DeepPolar uses BCE over the $u$2 information bits and alternates decoder and encoder optimization, with a two-stage curriculum aligned to the polar hierarchy (Hebbar et al., 2024).

A second regime is unsupervised or weakly supervised training from code constraints. For neural BP on polar factor graphs, modified syndrome losses exploit the fact that frozen-bit constraints induce a polar parity-check matrix $u$3. Soft syndromes are computed with min-sum semantics and penalized through a hinge-like loss, allowing training without labels (Teng et al., 2019). A CRC-enabled extension applies the same principle to the outer CRC constraints and, in experiments, yielded lower BLER than supervised BCE under the same RNN-BP architecture (Teng et al., 2020).

A third regime is self-supervised one-shot decoding. Instead of using ground-truth information vectors as labels, the decoder outputs soft information-bit values, re-encodes them through the known polar generator matrix, and minimizes the distance between the re-encoded vector and the received channel observation. In the proposed scheme, the re-encoder makes the network act as a bounded-distance decoder, and simulations for $u$4 showed BER and BLER approaching MAP for very short packets together with markedly stronger generalization than a conventional supervised one-shot baseline (Song et al., 2023).

A fourth regime is sample-based mutual-information estimation and input optimization. In the black-box-channel setting, one NPD estimates $u$5 from $u$6 and a second “constant-output” NPD estimates $u$7 from $u$8. Their empirical loss difference yields an MI estimator, and alternating estimation and improvement updates the parameters of an input-distribution model—Bernoulli for memoryless channels, LSTM for FSCs—through a REINFORCE-style gradient. The same trained NPDs then support code design, Honda–Yamamoto shaping, and SCL decoding under the optimized non-uniform input law (Aharoni et al., 18 Jun 2025).

4. Complexity, scalability, and implementation

One-shot NPDs were originally motivated by latency. RNND performs inference in a single forward pass through small networks with parameter counts on the order of $u$9k–$x=uG_N$0k: MLP-NND $x=uG_N$1, MLP-RNND $x=uG_N$2, CNN-NND $x=uG_N$3, CNN-RNND $x=uG_N$4, RNN-NND $x=uG_N$5, and RNN-RNND $x=uG_N$6. Under the tested setup, MLP-RNND ran more than $x=uG_N$7 faster than SC, while remaining only slightly slower than plain NND because of the denoising front-end (Cao et al., 2019).

The principal obstacle to naive neural decoding is training complexity, which scales exponentially with the number of information bits. Partitioned and concatenated NPDs address this directly. Partitioning trains sub-block decoders near MAP and couples them through deterministic BP stages; concatenation builds a decoder for any $x=uG_N$8 by recursively gluing shorter trained networks with ReLU-based propagation and XOR layers. In the concatenation framework, the total number of layers is $x=uG_N$9 at each merge, and the constructed decoder for $f$0 began at SC-level BER before fine-tuning (Cammerer et al., 2017, Stupachenko, 2022).

BP-derived NPDs expose a different complexity profile. SNND replaces the deep polar factor graph by a sparse Tanner graph with fully parallel flooding, fixed per-iteration latency of $f$1, and approximately $f$2 fewer operations per iteration than a dense factor-graph neural BP baseline for $f$3 and $f$4 (Xu et al., 2018). The recurrent neural BP decoder with codebook quantization reduces parameter count by about $f$5 relative to DNN-BP, and for the $f$6 code with $f$7, $f$8, and $f$9, memory drops from about $g$0 bits to about $g$1 bits, a $g$2 reduction, while multiplications are eliminated entirely (Teng et al., 2018).

For channels with synchronization errors, complexity reduction is more dramatic. The deletion-channel NPD modifies only the embedding network so that padded, positionally encoded, variable-length observations can feed a fixed polar graph. Trellis-based exact SC decoding on deletion channels has $g$3 complexity, whereas the NPD has $g$4 complexity, where $g$5 is a user-selected computational budget independent of the channel. At $g$6, the paper reports about $g$7 blocks/s for NPD, compared with about $g$8 blocks/s for trellis SC at $g$9 and about $O(N\log N)$0 blocks/s at $O(N\log N)$1 (Aharoni et al., 16 Jul 2025).

In end-to-end communication systems, graph-structured NPDs retain the same asymptotic polar-decoding structure while making complexity independent of channel memory size. For SC, complexity is $O(N\log N)$2; for list decoding, $O(N\log N)$3. This avoids the $O(N\log N)$4 dependence of successive-cancellation-trellis decoding on the channel-state size and is the reason NPDs can operate directly on ISI channels, OFDM links, and CP-less waveforms without explicit channel estimation or equalization (Hirsch et al., 3 Oct 2025).

5. Empirical performance and application domains

In the short-block regime most closely associated with early NPD work, the main reference point is SC or SCL. RNND improves BER over plain one-shot NNDs at comparable latency; for example, MLP-RNND achieves roughly $O(N\log N)$5 dB improvement over MLP-NND at BER $O(N\log N)$6 and comes very close to SC for the short $O(N\log N)$7 code. The residual denoiser increases the effective test SNR by about $O(N\log N)$8 dB at low test SNR, with larger gains as test SNR increases (Cao et al., 2019). CNN-assisted bit flipping, using an RNN-BP core with $O(N\log N)$9 iterations on a $K$0 polar code with CRC length $K$1, achieves lower BLER than SCL with list size $K$2 while requiring only half the flipping attempts of the critical-set bit-flipping baseline at matched BLER (Teng et al., 2019).

A distinct application area is code design over unknown or asymmetric channels. For black-box channels, NPD-guided MI optimization recovers the known optimum $K$3 on binary-input AWGN. On the Ising channel, estimated per-symbol MI increases from about $K$4 for uniform i.i.d. inputs to about $K$5, approaching the known lower bound $K$6 and upper bound $K$7. On the Trapdoor channel, per-symbol MI increases from about $K$8 to about $K$9, approaching known bounds $r(y;\theta)$0. At comparable rates around $r(y;\theta)$1, optimized inputs with SCL-NPD yield lower BER across list sizes $r(y;\theta)$2, and at $r(y;\theta)$3 improvements up to one order of magnitude over uniform i.i.d. inputs are reported (Aharoni et al., 18 Jun 2025).

NPDs have also been extended to synchronization-error channels and DNA storage. On deletion channels with $r(y;\theta)$4, the deletion-channel NPD slightly outperforms the trellis SC implementation in FER plots, although the authors explicitly attribute the gap to numerical issues in the trellis implementation. The reduced complexity makes SCL feasible; with $r(y;\theta)$5 and list size $r(y;\theta)$6, FER falls to about $r(y;\theta)$7–$r(y;\theta)$8 across $r(y;\theta)$9 (Aharoni et al., 16 Jul 2025). For IDS channels and DNA storage, NPDs provide both decoding and MI estimation. On binary deletion channels at $N=2^n$00, MI estimates such as about $N=2^n$01 at $N=2^n$02, about $N=2^n$03 at $N=2^n$04, and about $N=2^n$05 at $N=2^n$06 after input optimization sit near the corresponding literature bounds. On real Nanopore data, NPD is competitive with DNAformer while using about $N=2^n$07M parameters versus about $N=2^n$08M, and single-strand SCL with $N=2^n$09 reaches BER about $N=2^n$10 at rate $N=2^n$11 bits/base (Aharoni et al., 20 Jun 2025).

A further extension is joint code and decoder invention. DeepPolar learns nonlinear large-kernel encoders and matched decoders on the Plotkin tree. For $N=2^n$12 and $N=2^n$13, DeepPolar with $N=2^n$14 outperforms Polar$N=2^n$15, RM$N=2^n$16, and KO$N=2^n$17 in BER, and a parallel leaf decoder reduces parameters by about $N=2^n$18 while improving BER over DeepPolar-SC. At the same time, BLER remains inferior to Polar SC/SCL in several settings because BCE targets BER rather than block error (Hebbar et al., 2024).

In practical wireless receivers, NPDs have been adapted to complete OFDM and single-carrier systems. A single model per modulation and waveform generalizes across TDL profiles, SNRs, Doppler values, delay spreads, and even nonlinear PA distortion. The reported receiver operates directly on channels with memory, without pilots and without a cyclic prefix, and consistently outperforms the standardized 5G polar decoder in BER, BLER, and throughput, especially for low-rate and short-block configurations typical of control channels (Hirsch et al., 3 Oct 2025).

6. Theory, limitations, and open directions

A recent theoretical line formulates NPD as bitwise over-parameterized learning of synthesized polar channels. In the bitwise ONN decoder, each bit-channel is modeled by a two-layer ReLU network

$N=2^n$19

with fixed Rademacher second-layer weights and trainable first-layer weights. Under over-parameterization, the empirical MSE converges geometrically, the trajectory remains close to initialization, and the per-iteration training gain in dB is

$N=2^n$20

where $N=2^n$21 is the learning rate, $N=2^n$22 is the minimum eigenvalue of the bit-channel Gram matrix, and $N=2^n$23 is the training-set size. The analysis then converts a population MSE bound into per-bit error, BER, and BLER bounds through posterior margins and a Gaussian-approximation characterization of low-margin probability on AWGN channels (Zhu et al., 30 Apr 2026).

Despite rapid diversification, the limitations are consistent across the literature. Monolithic one-shot decoders still face exponential training difficulty as the number of information bits grows, which is why both partitioning and concatenation remain active design strategies (Cammerer et al., 2017, Stupachenko, 2022). Many one-shot results remain concentrated on very short codes such as $N=2^n$24 or $N=2^n$25, and broader robustness to channel mismatch, longer blocklengths, or changing frozen sets is often not evaluated explicitly (Cao et al., 2019, Song et al., 2023). In black-box input optimization, optimized non-uniform inputs reduce polarization and may require shaping and adaptive frozen sets (Aharoni et al., 18 Jun 2025). In DeepPolar, BCE improves BER but leaves BLER inferior to classical Polar SC/SCL in several configurations (Hebbar et al., 2024). In practical 5G-style deployment, CRC integration, standardized reliability sequences, puncturing or shortening, quantization, pruning, and hardware-aware training remain necessary engineering steps (Hirsch et al., 3 Oct 2025).

The overall trajectory of the field is therefore dual. On one side, NPDs increasingly preserve more of the polar structure—SC recursion, SCL path metrics, BP flooding, Plotkin trees, Honda–Yamamoto shaping—while learning only the channel- and inference-specific parts. On the other, they increasingly move beyond the memoryless setting, into asymmetric channels, finite-state channels, deletion and IDS channels, DNA storage, and pilotless wireless receivers. This suggests that the most durable interpretation of the NPD concept is not “a neural network for decoding polar codes” in the narrow sense, but a broader program of replacing analytically specified polar-decoding primitives by trainable ones wherever explicit models are unavailable, too expensive, or too restrictive (Aharoni et al., 18 Jun 2025, Aharoni et al., 20 Jun 2025, Hirsch et al., 3 Oct 2025).