Ben-Gurion's SAQ-Decoder Hits 18.6% Threshold, Closing the Quantum Error Gap
A transformer-based decoder from Ben-Gurion University hits near maximum-likelihood accuracy on quantum error correction while scaling linearly with syndrome size.
Takeaways
- SAQ-Decoder hits 10.99% and 18.6% thresholds on toric codes, near the 11.0% and 18.9% ML bounds.
- Dual-stream transformer with asymmetric attention: syndrome tokens attend locally, logical tokens attend globally.
- New differentiable logical loss optimizes logical error rate directly over GF(2) using tanh-based parity expectations.
- Linear-time CPND post-processor guarantees syndrome consistency, replacing cubic-cost OSD refinement.
- Uses 3.5x fewer parameters than QECCT and 5x faster inference at code distance 10.
- Open source code available on GitHub with toric, surface, color, and repetition code support.
Quantum error correction has been stuck between bad options. Classical matching algorithms scale poorly, tensor network methods scale worse, and neural decoders scale nicely but sacrifice too much accuracy to be useful for real fault-tolerant machines. A new paper from Ben-Gurion University argues you can have both, and the numbers are hard to ignore.
The system is called SAQ-Decoder (Stabilizer-Aware Quantum), a transformer paired with a constraint-enforcing post-processing stage that pushes decoding thresholds on toric codes to within a fraction of a percentage point of the theoretical maximum-likelihood ceiling, while running in linear time with respect to syndrome size. Code and weights are on GitHub.
The bottleneck nobody talks about
Quantum computers protect fragile qubits by spreading each logical qubit across many physical qubits, encoded in a stabilizer code. When errors happen, you cannot measure the qubits directly without destroying them, so you measure indirect parity checks called syndromes. A decoder looks at those syndromes, figures out what error most likely occurred, and tells the machine how to undo it.
The catch: quantum error correction demands real-time decoding that processes syndrome measurements and determines corrections within microseconds while maintaining near-optimal accuracy. Miss that deadline and errors pile up faster than corrections can catch them.
The reigning classical champion, Minimum Weight Perfect Matching (MWPM), treats decoding as a graph matching task. It works well under simple noise but scales prohibitively with code distance, and it treats bit-flip and phase-flip errors independently, which is not how real quantum noise behaves. Tensor network decoders achieve better accuracy but run too slowly. Neural approaches like the earlier QEC Transformer (QECCT) are fast but leave accuracy on the table.
The dual-stream trick
SAQ-Decoder's core insight is that syndrome information and logical error information have fundamentally different geometries, so they deserve different processing paths in the same network. The architecture splits into two token streams with asymmetric attention: