Kurate.org
Rankings
Back to Rankings

Distilling Knowledge from Search: Decoder Model Factory for Quantum LDPC Codes

Anil Bilgin, Wanda Hou, Pranav Srinivas, Yi-Zhuang You

Oct 30, 2026arXiv:2610.41398v1
quant-phcond-mat.dis-nncs.AI
Share
Scorecard· 16/16
7.0/10 impact

Coherent new paradigm (search-trajectory distillation for qLDPC decoders) with careful ablations and transfer results, but simulation-only, no circuit-level noise, and not uniformly better than Relay-BP.

Abstract

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 Impact Assessments

(1 model)

Core Contribution

The paper introduces GraphMP, a small (about 10.7k parameters) learned message-passing decoder that operates on detector-error-model (DEM) Tanner graphs. A single frozen checkpoint is applied across qLDPC families (BB, GB, mitten, lifted-product), code sizes, physical error rates, and noise models (code-capacity and phenomenological). GraphMP is first distilled from a BP teacher and then embedded in a beam-search decoder, following Ye et al., with the BP soft decoder replaced by GraphMP. The central conceptual contribution is the "decoder model factory", an expert-iteration-style loop. Beam search guides the model, trajectories that pass logical verification are collected, and the next model is trained on every intermediate pinned state of those trajectories. This addresses the mismatch between single-pass training and use inside a search. The loop is also shown to adapt the decoder when the noise changes from depolarizing to XZ. The paper frames this as a quantum–classical architecture with a fast decoding cycle and a slow learning cycle.

Methodological Rigor

The design is largely careful.

  • The root-only versus trajectory-trained ablation is a good control. A matched-source-problem control isolates conditional-state training from changes in how often each problem is sampled.
  • Five independent factory runs are reported, with ranges, Wilson intervals, and paired Z scores.
  • The benchmark comparison with Holm-adjusted paired tests is unusually statistical for this literature.
  • The comparison includes strong baselines (BP+OSD/LSD, Relay-BP, BP beam search, Tesseract).
  • Baseline configurations are fully specified.
  • The paper reports limitations candidly: weak transfer to the largest codes (30–38% logical error versus 5–8% for the BP teacher at the single-pass bootstrap stage), and phenomenological Mitten 150 and the XZ updated-prior comparison, where classical decoders remain better.
  • The gradient-SNR analysis justifying bit-wise over logical-coset targets is a nice side result. It is an initialization-only diagnostic, so it is suggestive rather than conclusive.

    • Weaknesses:

  • Everything is simulated. Logical verification comes from the sampled error, so the "operating experience" claim is prospective.
  • The evaluation is code-capacity and phenomenological only. There is no circuit-level noise, which is the practically relevant regime.
  • Training uses only three small codes, and "transfer" is demonstrated on sizes up to roughly 5–15× larger only after search.
  • Only 2048 shots per task are used in the main comparison, so rare-event regimes are poorly resolved. Some logical-rate curves are fitted extrapolations.
  • The noise-adaptation test is a toy: an abrupt switch to XZ noise, with a candidate selector that retains the original prior.
  • The comparison of GPU GraphMP with single-core CPU baselines is not apples-to-apples, and the authors acknowledge this.
  • The "recursive self-improvement" is bounded, with sharply diminishing returns after G1.
  • The paper never shows the factory beating the combination of an updated prior with a classical decoder in the drift scenario.
  • Some prior-work overlap exists: expert iteration, GNN-assisted BP with harder-example collection, and BP beam search all exist. The contribution is the integration plus the conditional-state distillation.

    • Potential Impact

    If the approach scales to circuit-level noise and hardware logical feedback, it offers a path to decoders that continuously adapt to drifting noise without hand-tuned priors. That is a real problem for qLDPC and neutral-atom architectures. The companion-paper hook (QDPU/FPGA) signals industrial ambitions, since the authors are a company. The conceptual transfer of the ExIt/STaR/ReST paradigm to QEC decoding is attractive, and the observation that search-trajectory distillation fixes the training/deployment mismatch is a reusable lesson. The practical near-term impact is limited, however. The decoder is not yet competitive with Relay-BP in absolute performance across all tasks. Real-time FPGA deployment of a 24-round GRU message-passing network plus beam search is unproven. Hardware logical feedback is nontrivial and consumes experimental time.

    Timeliness

    Highly timely. It addresses qLDPC decoding, learned decoders and foundation decoders, noise drift, and GPU/QPU coupling (NVQLink). It also draws on the current enthusiasm for self-improvement loops from LLM reasoning work.

    Strengths

  • A coherent, well-motivated framework with a clean ablation of the key mechanism.
  • Cross-family, no-fine-tuning transfer using a DEM interface.
  • Honest reporting and substantial appendix documentation, including statistical tests and decoding-time methodology.
  • A compact model with favorable per-shot time against classical decoders on the tested tasks.

    • Limitations

  • Simulation-only verification, with no circuit-level noise.
  • Gains diminish quickly.
  • Transfer to the largest codes depends on search.
  • The adaptation experiment is a toy, with unequal baselines.
  • The grand vision (quantum immune system) outruns the evidence.
  • The 2048-shot comparisons are small relative to practical logical error targets.
  • Reproducibility is moderate: the architecture, hyperparameters, and training recipe are detailed, but no code release is stated.

    • Overall

    A solid and well-executed methods paper with a memorable framing and meaningful ablations. Its value rests on whether the factory paradigm scales to circuit-level noise and real hardware, which is not yet shown. It is likely to be cited in the learned-QEC-decoder community. It is unlikely to be transformative on current evidence.

    Rating:7/ 10
    Significance 7Rigor 7.5Novelty 7Clarity 7.5

    Generated Nov 2, 2026

    Comparison History (0)

    No comparisons yet.