We introduce GraphMP, a learning-based message-passing decoder for quantum low-density parity-check codes that transfers without fine-tuning across code families, code sizes, physical error rates, and noise models. GraphMP uses local message passing to keep computation scalable as the decoding graph grows. Guided by its predictions, beam search progressively refines candidate error patterns, trading additional decoding time for lower logical error rates. We develop the decoder model factory as a paradigm for training successive models on experience generated by earlier models during decoding. Model-guided search supplies verified trajectories for knowledge distillation, folding computation spent on decoding into reusable training data. Repeated collection and distillation improve the decoder across generations and adapt its strategy as the noise changes. These results motivate a coupled quantum--classical architecture with two interacting cycles: rapid decoding protects quantum information, while slower learning renews the models from operating experience. The long-term vision is a scalable decoder model factory that produces high-performance, low-latency decoders and continually adapts them to evolving quantum hardware.
AI Summary
The paper asks whether a learned qLDPC decoder can transfer across decoding tasks and improve by training on its own search experience. GraphMP uses shared local message-passing rules on detector–error graphs to predict per-error probabilities. It is initialized by distillation from belief propagation (BP), then embedded in beam search that pins uncertain variables and repeatedly predicts the remainder. Search trajectories whose final corrections are syndrome-consistent and logically correct supply conditional-completion training examples for successive models; verification uses simulated errors.
A single model is trained on three code-capacity tasks and evaluated without fine-tuning on nine additional tasks spanning code sizes, families, and phenomenological noise. Across five independent training runs, three collection–distillation generations improve decoding, with diminishing gains. In one final-generation run, pooled logical error across the three training tasks falls from $3.39\%$ in single-pass decoding to $0.65\%$ with width-$8$, depth-$10$ search, and $0.44\%$ with width $32$. Under the reported configurations, final search-enhanced GraphMP has lower measured logical error than BP+OSD on all nine code-capacity tasks and is broadly competitive with Relay-BP. Controlled comparisons support training on intermediate search states rather than only search inputs and outputs.
Further training recovers accuracy after a toy depolarizing-to-XZ noise switch, including on an unseen task. Limitations include task-dependent accuracy gaps, diminishing generational improvements, and no demonstrated circuit-level or hardware learning. Hardware collection would require dedicated logical-feedback experiments. Reported GPU-versus-single-core-CPU throughput comparisons do not establish real-time decoding latency; sustained noise drift and deadline-constrained deployment remain untested.
Little sign of AI-generated text
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
We introduce GraphMP, a learning-based message-passing decoder for quantum low-density parity-check codes that transfers without fine-tuning across code families, code sizes, physical error rates, and noise models. GraphMP uses local message passing to keep computation scalable as the decoding graph grows. Guided by its predictions, beam search progressively refines candidate error patterns, trading additional decoding time for lower logical error rates. We develop the decoder model factory as a paradigm for training successive models on experience generated by earlier models during decoding. Model-guided search supplies verified trajectories for knowledge distillation, folding computation spent on decoding into reusable training data. Repeated collection and distillation improve the decoder across generations and adapt its strategy as the noise changes. These results motivate a coupled quantum--classical architecture with two interacting cycles: rapid decoding protects quantum information, while slower learning renews the models from operating experience. The long-term vision is a scalable decoder model factory that produces high-performance, low-latency decoders and continually adapts them to evolving quantum hardware.
The paper asks whether a learned qLDPC decoder can transfer across decoding tasks and improve by training on its own search experience. GraphMP uses shared local message-passing rules on detector–error graphs to predict per-error probabilities. It is initialized by distillation from belief propagation (BP), then embedded in beam search that pins uncertain variables and repeatedly predicts the remainder. Search trajectories whose final corrections are syndrome-consistent and logically correct supply conditional-completion training examples for successive models; verification uses simulated errors.
A single model is trained on three code-capacity tasks and evaluated without fine-tuning on nine additional tasks spanning code sizes, families, and phenomenological noise. Across five independent training runs, three collection–distillation generations improve decoding, with diminishing gains. In one final-generation run, pooled logical error across the three training tasks falls from 3.39% in single-pass decoding to 0.65% with width-8, depth-10 search, and 0.44% with width 32. Under the reported configurations, final search-enhanced GraphMP has lower measured logical error than BP+OSD on all nine code-capacity tasks and is broadly competitive with Relay-BP. Controlled comparisons support training on intermediate search states rather than only search inputs and outputs.
Further training recovers accuracy after a toy depolarizing-to-XZ noise switch, including on an unseen task. Limitations include task-dependent accuracy gaps, diminishing generational improvements, and no demonstrated circuit-level or hardware learning. Hardware collection would require dedicated logical-feedback experiments. Reported GPU-versus-single-core-CPU throughput comparisons do not establish real-time decoding latency; sustained noise drift and deadline-constrained deployment remain untested.
We propose a decoder for quantum low density parity check (LDPC) codes based on a beam search heuristic guided by belief propagation (BP). Our beam search decoder applies to all quantum LDPC codes and achieves different speed-accuracy tradeoffs by tuning its parameters such as the beam width. We perform numerical simulations under circuit level noise for the [[144,12,12]] bivariate bicycle (BB) code at noise rate p=10−3 to estimate the logical error rate and the 99.9 percentile runtime and we compare with the BP-OSD decoder which has been the default quantum LDPC decoder for the past six years. A variant of our beam search decoder with a beam width of 64 achieves a 17× reduction in logical error rate. With a beam width of 8, we reach the same logical error rate as BP-OSD with a 26.2× reduction in the 99.9 percentile runtime. We identify the beam search decoder with beam width of 32 as a promising candidate for trapped ion architectures because it achieves a 5.6× reduction in logical error rate with a 99.9 percentile runtime per syndrome extraction round below 1ms at p=5×10−4. Remarkably, this is achieved in software on a single core, without any parallelization or specialized hardware (FPGA, ASIC), suggesting one might only need three 32-core CPUs to decode a trapped ion quantum computer with 1000 logical qubits.
We present Astra, a novel and scalable decoder using graph neural networks. In general, Quantum Low Density Parity Check (QLDPC) decoding is based on Belief Propagation (BP, a variant of message-passing) and requires time intensive post-processing methods such as Ordered Statistics Decoding (OSD). Our decoder works on the Tanner graph, similarly to BP. Without using any post-processing, Astra achieves higher thresholds and better Logical Error Rates (LER) compared to BPOSD, both for surface codes trained up to distance 11 and Bivariate Bicycle (BB) codes trained up to distance 18. Moreover, we can successfully extrapolate the decoding functionality: we decode high distances (surface code up to distance 25 and BB code up to distance 34) by using decoders trained on lower distances. Extrapolated Astra achieves better LER than BPOSD for BB codes. Astra(+OSD) achieves orders of magnitude lower logical error rates for BB codes compared to BP(+OSD).
Quantum low-density parity-check (QLDPC) codes are promising candidates for low-overhead fault-tolerant quantum computation, but their practical use requires fast, low-complexity, and reliable decoders. Belief propagation (BP) is attractive because of its local message-passing structure, yet standard flooding BP often suffers from convergence failures on QLDPC codes due to short cycles, degeneracy, and symmetric decoding trajectories. Learned sequential BP improves convergence by using a reinforcement-learning policy to choose variable-node update orders, but a single learned trajectory can still be sensitive to unfavorable local Pauli decisions. We propose a learned parallel bit-flipping sequential BP decoder with a new quaternary score combining syndrome gain, a quantized Pauli log-likelihood penalty, and Q-table lookahead at hypothetical post-flip states. Our decoder first runs learned sequential BP for a fixed number of iterations. If the syndrome is not satisfied, it constructs quaternary bit-flipping candidates, each corresponding to changing the current Pauli decision of one qubit. The selected candidates initialize independent learned sequential BP continuations from the same decoder state. Since these continuations are independent, they can be executed in parallel, so testing several candidates mainly increases parallel hardware resources rather than sequential decoding latency. Simulations on representative QLDPC codes over the depolarizing channel show that the proposed decoder improves the reliability of learned sequential BP while having a parallel low-latency structure.
Quantum computing requires effective error correction strategies to mitigate noise and decoherence. Quantum Low-Density Parity-Check (QLDPC) codes have emerged as a promising solution for scalable Quantum Error Correction (QEC) applications by supporting constant-rate encoding and a sparse parity-check structure. However, decoding QLDPC codes via traditional approaches such as Belief Propagation (BP) suffers from poor convergence in the presence of short cycles. Machine learning techniques like Graph Neural Networks (GNNs) utilize learned message passing over their node features; however, they are restricted to pairwise interactions on Tanner graphs, which limits their ability to capture higher-order correlations. In this work, we propose HyperNQ, the first Hypergraph Neural Network (HGNN)- based QLDPC decoder that captures higher-order stabilizer constraints by utilizing hyperedges-thus enabling highly expressive and compact decoding. We use a two-stage message passing scheme and evaluate the decoder over the pseudo-threshold region. Below the pseudo-threshold mark, HyperNQ improves the Logical Error Rate (LER) up to 84% over BP and 50% over GNN-based strategies, demonstrating enhanced performance over the existing state-of-the-art decoders.
The decoder is a critical component of a fault-tolerant quantum computer, computing corrections based on parity-check measurements performed throughout the computation. A soft decoder supplements its output with a confidence score which, when used alongside post-selection, can substantially improve logical performance. We introduce a soft beam search decoder for quantum low-density parity-check codes that uses internal decoder data as a confidence metric, removing the need for extra computation. We perform circuit-level simulations of five quantum LDPC codes relevant for superconducting and trapped ion architectures equipped with our global soft decoder and we obtain up to 580× logical-error suppression at a physical error rate of 10−3 while rejecting only 0.1% of shots. Then, we simulate an error detected measurement, which is a core logical operation of the walking cat architecture, using a streaming version of soft beam decoder and we achieve up to 210× error suppression while increasing the rejection probability by only 0.5 percentage points. Finally, we revisit recent quantum LDPC code memory experiments on trapped ions, demonstrating that our soft decoder doubles the logical qubit lifetimes at the price of a mean rejection rate of 2.6%--5.6% per syndrome round, bringing all five codes into the beyond-breakeven regime.
Quantum computers can solve selected problems more efficiently than classical computers, but current devices are limited by high physical error rates. Quantum Error Correction (QEC) mitigates this by encoding many physical qubits into a logical qubit, with the surface code among the most widely studied approaches. Since syndrome measurements are produced continuously, the decoder must process them fast enough to avoid becoming a system bottleneck, making real-time decoding essential for fault-tolerant quantum computing. While most state-of-the-art real-time decoders rely on Minimum-Weight Perfect Matching (MWPM), we instead use a high-accuracy Graph Neural Network (GNN) that trades higher computational cost for lower logical error rates. To make this GNN practical for real-time decoding, we apply algorithm-hardware co-design. We reduce complexity through hardware-guided pruning and retraining, producing two hardware-friendly models that reduce parameter count by 3.1× and 6.5×. These target, respectively, an average decoding latency of one syndrome cycle and a worst-case latency within one syndrome cycle. We further reduce hardware cost using input-graph filtering and post-training quantization. Based on these optimized models, we propose an FPGA-based architecture for low-latency inference and real-time decoding. Evaluated on surface codes up to distance 7 under circuit-level noise at physical error rate p=10−3, our decoder outperforms MWPM in decoding accuracy for both average-latency and max-latency settings. It reduces logical error rate by 40% at 1μs average latency, and by 13% under a strict 1μs deadline.