Work

Quant-S: a small neural decoder with ternary weights, tested on Sycamore surface-code data

Year: 2026
Type: preprint
Abstract: We report Quant-S, a neural decoder for surface-code memory experiments with 3.6million parameters, of which 3.15 million are ternary weights in the feed-forward blocks. Itreads syndromes round by round with a recurrent state and needs no GPU at inferencetime. We trained it only on simulated syndromes drawn from Google's published detectorerror models and tested it on the held-out half of Google's 2023 Sycamore distance-5experiment (Z basis, 5, 11, 17 and 25 rounds, 25,000 shots each). In both of two trainingseeds, Quant-S has a lower logical error rate than PyMatching at 5, 11 and 17 rounds(absolute reductions 0.011 to 0.023, paired exact McNemar p ≤ 5 × 10⁻⁷) and matches it at25 rounds. Its logical error per round is 3.90 % and 3.96 % for the two seeds, against 4.21% for PyMatching, 3.50 % for correlated matching, 2.94 % for belief matching and 2.82 %for tensor-network decoding on the same shots. A float model of the same size trainedthe same way reaches correlated matching at 17 rounds. On one laptop CPU core thedecision after the last round takes a median of 155 to 181 µs. Quant-S is therefore not themost accurate decoder on this data. Its point is that a decoder this small, with mostlyternary weights, already beats standard matching on real hardware data and runs fastenough for platforms with millisecond cycles, such as trapped ions.
Language: English
Open Access status: green
OpenAlex ID: W7236737828
Cites:
Cited by:
Related to: 10