MDQEC-QAS: Meta-Decoding for Quantum Error Correction with Hardware-Aware VQC Search and Confidence-Gated Recovery

Technical Summary of MDQEC-QAS

Problem Statement
Quantum error correction (QEC) requires decoders that translate classical syndrome measurements into recovery operations to preserve logical information. Traditional decoders are often specialized to specific code families or noise models, limiting their adaptability. While machine learning (ML) has introduced learned decoders, most remain confined to fixed code/noise regimes. Furthermore, a critical gap exists between teacher-label accuracy (predicting the correct recovery string) and logical-level performance (preventing logical failures), particularly for large topological codes where rare prediction errors can be catastrophic. Additionally, the deployment of Variational Quantum Circuits (VQCs) for decoding is hindered by the lack of systematic methods to select architectures that balance predictive quality with hardware constraints (qubit count, depth, gate count).

Methodology
The authors propose MDQEC-QAS, a unified meta-decoding framework that learns syndrome-to-recovery mappings across multiple stabilizer codes and noise settings using a single shared model. The framework consists of four integrated components:

1. Meta-Dataset Construction: A pooled dataset is generated from four codes (FiveQubit, Steane, Planar3×3, Planar5×5) and four noise families (depolarizing, Z-biased, correlated Z-burst, and depolarizing with syndrome flips). Data is generated at physical error rates $p \in \{0.03, 0.05, 0.07\}$. Syndromes and recovery strings are padded to global dimensions and mapped to a unified label space conditioned on code and noise identifiers.
2. Meta-Decoder Training: Two model types are trained on tuples $(s, c, \eta, p) \to y$:
   • Meta-MLP: A classical multilayer perceptron baseline.
   • Meta-VQC: A hybrid variational quantum classifier where the input is projected to a qubit dimension, embedded via angle encoding, processed through trainable layers with entangling gates, and measured via Pauli-Z expectations.
   • Training employs a weighted cross-entropy loss to prioritize harder regimes (e.g., Planar5×5) and code-label masking to prevent cross-code prediction.
3. Hardware-Aware Quantum Architecture Search (QAS): Instead of fixing a VQC architecture, the authors perform a search over qubit count ($n_q \in \{4,5,6\}$), circuit depth ($L \in \{1,2\}$), and entangling topology (chain, ring, full). Architectures are evaluated using a score that balances validation accuracy against hardware cost proxies (raw/transpiled depth and two-qubit gate counts).
4. Confidence-Gated Hybrid Recovery: To address the sensitivity of large topological codes to prediction errors, a fallback mechanism is introduced. If the learned decoder's prediction confidence exceeds a threshold ($\tau$), its output is used; otherwise, the system falls back to a code-appropriate "teacher" decoder.

Evaluation Regimes
The framework is evaluated across five transfer settings:

• Interpolation: All codes and noise types are seen during training.
• Unseen-$p$ Transfer: Testing on physical error rates not seen in training.
• Unseen-Noise Transfer: Testing on a noise family not seen in training.
• Few-Shot Unseen-Code Adaptation: Testing on a held-out code (Planar5×5) with only 5% of its data in the training set.
• Few-Shot Held-Out-Size Adaptation: Testing on a code size not seen in training.

Key Results

• Teacher-Label Accuracy: The Meta-MLP achieves high accuracy across all regimes (e.g., 0.9993 in interpolation, 0.9118 in unseen-$p$). VQCs selected via hardware-aware search perform competitively, often outperforming accuracy-selected VQCs in transfer settings (e.g., 0.8495 vs. 0.8397 in unseen-$p$), though they generally lag behind the Meta-MLP.
• Transfer Hierarchy: Generalization across noise parameters ($p$ and noise family) is significantly easier than generalization across code structures (unseen code or code size). Few-shot adaptation to an unseen code structure results in the most significant performance degradation.
• Logical-Level Performance: High teacher-label accuracy does not guarantee low logical failure rates, especially for Planar5×5. In the interpolation regime, raw logical failure ratios relative to the teacher were 12.08 (Meta-MLP) and 25.91 (VQC).
• Impact of Confidence Gating: Applying the confidence-gated fallback mechanism drastically reduces logical failure ratios for Planar5×5 to 1.71 (Meta-MLP) and 1.11 (VQC), demonstrating that selective recovery based on confidence is more effective than unconditional replacement of the teacher.

Significance and Claims
The paper claims that its primary contribution is not the isolated use of neural or quantum decoders, but the integration of multi-code meta-decoding, transfer evaluation, hardware-aware VQC search, and confidence-gated hybrid recovery into a single reproducible pipeline.

Key findings emphasized by the authors include:

1. Structural Transfer Difficulty: Adapting to new code structures or sizes is substantially harder than adapting to new noise distributions.
2. Hardware-Aware Selection: Cost-aware VQC selection can yield architectures that remain competitive with accuracy-selected circuits, particularly in transfer scenarios.
3. Reliability vs. Accuracy: For large topological codes, raw supervised accuracy is insufficient as a reliability metric. Confidence-aware mechanisms are necessary to bridge the gap between learned predictions and logical reliability.
4. Hybrid Recovery: The system functions best as a "learned-assisted" recovery system where the learned model acts as a router, delegating uncertain cases to a robust teacher, rather than as a standalone replacement for traditional decoders.