A cheaper neural decoder matches the old accuracy, but its speed edge is simulated
Three Yonsei University researchers rebuilt an AlphaQubit-style decoder and swapped its attention engine for a lighter design called Mamba. On real chip data it matches the accuracy of their own attention version. Its claimed speed advantage at today’s code sizes comes from a simulation with an assumed delay, not from a measurement.
A race against the chip
A quantum computer that corrects its own errors measures a set of checks over and over, and a classical program called a decoder has to turn each batch of results into a guess about which errors happened. If the decoder falls behind, the backlog grows and the protection breaks down. On Google’s Willow chip one round of checks takes 1.1 microseconds. In its 2024 paper, Google decoded a distance 5 code in real time with an average decoder latency of 63 microseconds 2. That is the pace to keep.
Trapped ion machines run their check cycles far more slowly, so a decoder there has much more time per round.
Who did it and what they changed
The paper comes from Changwon Lee, Tak Hur and Daniel K. Park at Yonsei University in Seoul. It is academic work, published after peer review in npj Quantum Information on October 28, 2026 1.
Their starting point is AlphaQubit, the neural decoder from Google DeepMind. Its core uses attention, which compares every check result with every other one. Since a surface code of distance d has about d squared checks, the cost of those pairwise comparisons grows with d to the fourth power. The Yonsei team replaced that part with Mamba, a type of network that reads the checks in sequence and keeps a compressed memory, so its cost grows with d squared. The goal was to keep the accuracy of the attention design while making the cost grow more slowly.
Accuracy held
The team tested on data Google recorded on its Sycamore chip in 2022. The Mamba decoder reached a logical error per roundHow often the protected qubit fails during one cycle of error checks. of about 2.98% at distance 3 and 3.03% at distance 5, matching the attention version they built for comparison. Google’s own decoder scored 3.03% and 2.91% on the same data in 2022 3. Both new decoders trailed a slower tensor network decoder at distance 5. The authors trained on up to 100 million samples, against up to a billion for AlphaQubit, and they blame that smaller dataset for the gap.
Same recorded chip data, similar accuracy, and no gain from the larger code for the new decoder. Source: Lee, Hur, Park, npj Quantum Information (2026); Google Quantum AI, arXiv:2207.06431.
What is missing
The authors timed one Mamba block and one attention block on an RTX 4090 graphics card. In their own plot, both take roughly 10 microseconds at small distances, and the Mamba block is not clearly faster until about distance 21 1. At distance 39 the gap is large, about 0.6 milliseconds against 0.19, but those are code sizes far beyond today’s chips.
For distances 3 to 9, the real-time advantage comes from a simulation. The authors took AlphaQubit’s reported decoding time of about 40 microseconds at distance 9, turned it into an assumed extra error, and scaled that extra error with d to the fourth for attention and d squared for Mamba. That assumption sets how much worse the attention decoder gets in the simulation, before any hardware measurement. A single block at about 10 microseconds is also still around ten times slower than a 1.1 microsecond cycle, before counting the rest of the network.
What I take from it
For superconducting players, decoding speed is still an engineering cost they have to pay. For ions, the slower clock pays most of it, though they pay for that clock in total run time.
The fair comparison is with what already runs. Google’s Willow decoded distance 5 in real time with a conventional decoder, at an average delay of 63 µs for the whole decoder. The Mamba team timed only a single building block, about 10 µs on a GPU at today’s sizes, and the full network is several of those. Even at distance 39, where its lead over attention is large, one block takes about 0.19 ms, while a Willow cycle lasts 1.1 µs.
For this design to improve anything real on a superconducting chip, three things have to happen. It has to run on fast dedicated hardware beside the chip, it has to beat a conventional decoder’s accuracy at the same speed, and it has to show both on a live device at distance 7 or more.
