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

Research Question and Approach

This paper addresses the challenge of creating a universal quantum error correction (QEC) decoder that can adapt to various stabilizer codes, noise models, and physical error rates without requiring separate, specialized decoders for each configuration. The authors propose a meta-decoding framework where a single model is trained on a pooled dataset of syndrome-to-recovery mappings. They evaluate two types of decoders: a classical Meta-MLP and a Variational Quantum Circuit (VQC) optimized via hardware-aware quantum architecture search (QAS). The framework is tested across five regimes, including interpolation, transfer to unseen noise or error rates, and few-shot adaptation to new code structures.

The researchers find that while both the Meta-MLP and VQC models achieve high supervised accuracy in imitating teacher decoders, this accuracy does not always translate to low logical failure rates, particularly in larger topological codes like the Planar5x5. In these complex settings, even rare classification errors can lead to significant logical faults. To mitigate this, the authors introduce a confidence-gated hybrid recovery mechanism. By using the learned model's prediction only when it is highly confident and falling back to a traditional teacher decoder otherwise, they significantly reduce logical failure ratios compared to using the learned model alone. The study also identifies a hierarchy of transfer difficulty: adapting to new noise parameters is easier than adapting to new code structures or sizes.

As quantum hardware scales, the ability to maintain fault tolerance across heterogeneous and evolving noise environments becomes critical. This work demonstrates that machine learning can provide a flexible, unified decoding interface, but it also highlights the danger of relying solely on supervised classification metrics. The proposed confidence-gated approach offers a practical path toward integrating learned decoders into existing QEC pipelines, allowing for the benefits of neural-assisted decoding while maintaining the reliability of traditional, well-understood recovery strategies.