The Signal
Park, Kwak and Kim’s Transformer Posts an 18.1% Code-Capacity Threshold
Seong-Joon Park, Hee-Youl Kwak and Yongjune Kim report an 18.1% decoding threshold for surface codes under a code-capacity noise model. Their qubit-centric transformer also shows scalable error suppression in circuit-level tests, where the
Quantum.world Team · Written with AI · fact-checked ·
“Notably, QCT achieves a high threshold of 18.1% under the code-capacity noise model, which closely approaches the theoretical bound of 18.9% and surpasses both the BP+OSD and the minimum-weight perfect matching (MWPM) thresholds, while also exhibiting a substantially higher error suppression factor Λ.”
The paper’s authors, Seong-Joon Park, Hee-Youl Kwak and Yongjune Kim, write in “Qubit-centric transformer for surface code decoding”
Physical Review A publishes QCT
Seong-Joon Park, Hee-Youl Kwak and Yongjune Kim’s qubit-centric transformer, or QCT, was published in Physical Review A on October 26, 2026. The paper reports an 18.1% decoding threshold for surface codes under a code-capacity noise model, close to the 18.9% theoretical bound cited by the authors.
The result comes from a neural decoder designed to infer logical errors from quantum error-correction syndromes. Rather than feed those syndromes into a transformer as a stabilizer-based sequence, QCT converts them into tokens organized around physical qubits. A graph-based attention mask incorporates the code’s topology and directs the model toward relevant qubit interactions.
What the 18.1% figure measures
The 18.1% figure is tied to the code-capacity model; it is not a reported physical gate-error limit for a working quantum processor. The authors compare QCT’s result with thresholds for belief propagation with ordered statistics decoding, or BP+OSD, and minimum-weight perfect matching, or MWPM. They also report a substantially higher error-suppression factor for QCT, though the abstract does not give the numerical value.
That distinction matters when comparing decoder results. A threshold belongs to the assumptions of the noise model used to evaluate it. The headline percentage should therefore be read as a result for the stated code-capacity setting, not as a blanket measure of how much noise a quantum computer can tolerate.
A separate test adds repeated measurements
The paper also extends QCT to a circuit-level noise model. For that version, the authors add temporal convolution blocks to process repeated syndrome measurements. They report scalable error suppression across the code distances they evaluate and say the decoder consistently outperforms correlated MWPM and BP+OSD under those conditions.
The abstract does not attach a circuit-level threshold percentage to that result. It also does not provide a latency figure, so it cannot establish from the summary alone how quickly QCT returns a decoding decision or how its runtime compares with those other methods.
The model result and the deployment question
QCT’s design links the transformer’s input and attention structure to the surface code’s physical layout, while its circuit-level extension incorporates information from repeated measurements. The reported comparisons make a case for that approach in the evaluated simulations, across both code-capacity and circuit-level settings.
Those findings are distinct from a demonstration inside a live quantum error-correction loop. The published abstract reports model-based decoding performance; its 18.1% threshold specifically belongs to the code-capacity test, while the circuit-level result is described in terms of error suppression.