Can learned decoders match minimum-weight perfect matching (MWPM) on quantum error correction codes while staying inside a real-time decoding budget?
I chose the questions and experiments, reviewed the code and outputs, and am studying the repository in parallel with PhD applications.
Benchmarks a ladder of neural decoder families (MLP, CNN, GRU, ConvGRU, 3D-lattice convolution, GNN) against MWPM on the repetition and rotated surface codes, using Stim for circuit simulation, PyMatching for the MWPM baseline, and PyTorch for the learned decoders. Includes a comparison against reported results on public Google Sycamore syndrome data.
Chose the research questions and experiment design, directed the benchmarking methodology, and reviewed the code and results.
AI generated substantial implementation work: the decoder architectures, training loops, and much of the benchmarking code.
Independent reproduction of the repository's headline claims is in progress. The repository's own hardening pass found that an initially favorable real-time surface-code claim did not survive testing under circuit-level noise and single-shot timing, and the repository documents that correction rather than hiding it.
This is not peer-reviewed research and is not evidence of quantum advantage. It is a benchmarking study comparing classical and learned decoding strategies.