CoDecoder Neural Frontend Cuts Quantum Decoder Failures 6-50% and Speeds Decoding Up to 124x on Surface and BB Codes
CoDecoder demonstrates that a small neural router can measurably accelerate and improve fault-tolerant decoding without replacing mature decoders. The work highlights joint logical scoring and efficient GPU fusion as practical levers for near-term hardware. Larger-scale tests on distance-25+ codes remain necessary to confirm scaling.
The approach adds a logical-operator-aware output head that jointly scores equivalent logical operators on surface codes and all logical qubits on BB codes. Training uses circuit-level depolarizing noise; inference fuses GPU execution with per-shot routing so only low-confidence cases incur full decoder cost. On retained shots the frontend corrects more errors than it introduces even when overall accuracy trails the fallback. Registered tests against correlated matching and Relay-BP show consistent gains; re-implementations of Cascade confirm the composition is both faster and more accurate than the neural baseline alone.
CoDecoder-style frontends will appear in at least one open-source quantum control stack by Q4 2027, achieving >50x average speedup on distance-13 surface codes.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.28748)
- [2]Supporting Source(https://arxiv.org/abs/2309.15354)