| name | adaptive-confidence-gated-qec-decoding |
| description | Two-stage adaptive confidence-gated neural decoding framework for quantum error correction — lightweight neural fast-path with high-confidence fallback to classical refinement. Use when designing real-time QEC decoding, neural-classical hybrid inference, confidence-gated routing, latency-constrained quantum control, or hardware-aware QEC co-design. Activation: confidence gating, two-stage decoding, neural decoder, QEC latency, surface code, MWPM refinement, accuracy-latency tradeoff, rotated surface code, adaptive inference, hardware-aware |
| metadata | {"arxiv_id":"2607.05814","published":"2026-07-07","authors":"Sumit Chongder","category":"quant-ph, cs.ET, cs.LG","tags":["quantum-error-correction","neural-decoding","confidence-gating","surface-code","hardware-aware","two-stage-inference","latency-optimization"]} |
Adaptive Confidence-Gated QEC Decoding
Two-stage inference framework that routes syndrome measurements through a lightweight neural fast-path, escalating only low-confidence predictions to classical MWPM refinement.
Core Architecture
Syndrome Measurement → Neural Fast-Path (FFNN)
├─ High-confidence (>threshold) → Accept
└─ Low-confidence (<threshold) → MWPM Refinement
Key Results (arXiv:2607.05814)
- Routing efficiency: Only 3.3%–6.2% of syndromes escalated to refinement at confidence threshold 0.95
- Accuracy improvement: 99.21% (neural-only) → 99.81% (with refinement)
- Throughput: ~4.6×10⁵ samples/s on commodity CPU (batch size 512)
- Scalability: Neural path not the bottleneck beyond d=7
Methodology
- Train lightweight FFNN on syndrome-to-correction mapping for target code distance
- Calibrate confidence threshold on validation set — sweep accuracy vs latency trade-off
- Deploy two-stage decoder: neural fast-path for majority, MWPM fallback for edge cases
- Benchmark: logical accuracy, per-shot latency, throughput, resource scaling across distances d∈{3,5,7,9,11}
Design Patterns
Confidence-Gated Routing
- Route majority of inputs through cheap model
- Escalate only uncertain predictions to expensive model
- Trade-off: threshold controls accuracy vs latency
- Applicable beyond QEC: any two-stage inference pipeline
Hardware-Aware Co-Design
- Match decoder complexity to hardware throughput constraints
- Identify actual bottleneck (neural vs graph computation)
- Size batch for hardware saturation point
Validated vs Roadmap Contributions
- Clearly distinguish experimentally validated results from future directions
- Release complete benchmarking pipeline, models, and data
- Explicit scope boundaries in publications
Reusable Workflow
1. Define QEC code (surface code, distance d)
2. Generate training data (Stim simulator, circuit-level noise)
3. Train FFNN decoder (syndrome → correction)
4. Calibrate confidence threshold (accuracy-latency sweep)
5. Integrate MWPM fallback for low-confidence predictions
6. Benchmark across distances, noise models, batch sizes
7. Identify bottleneck → optimize accordingly
Pitfalls
- Confidence threshold sensitivity: Too high → excessive MWPM calls; too low → accuracy loss
- Neural saturation point: Beyond d=7, neural path is NOT the throughput bottleneck — optimize graph stage instead
- Noise model mismatch: Decoder trained on one noise model may degrade under different noise
References