Infinidata Lab · Research benchmark

Benchmarking QEC decoders across codes.

Quantum error correction (QEC) uses measurement data to detect and correct errors in quantum computations. The QEC Model Zoo provides a common benchmark for fair comparisons of AI-based decoders and established baselines across different QEC codes and real-device datasets.

Explore the datasets Explore the decoders

Accepted at NeurIPS 2026. The paper is not yet available.

09Real datasets
11Decoder methods
Cross-codeOne benchmark
The idea

Different approaches. Consistent evaluation.

A QEC decoder interprets measurements to infer likely errors and determine appropriate corrections. Comparing decoders across devices and code families requires a common evaluation setting. The model zoo provides that basis for reproducible, methodologically consistent comparisons.

01

Fair comparisons

Assess AI-based decoders alongside established classical methods using a common benchmark.

02

Real hardware datasets

Evaluate methods on experimental datasets spanning multiple code families and hardware platforms.

03

Reusable research assets

Access model checkpoints and explore pre-decoding techniques for efficient error correction.

01 / The data

Nine datasets.
Real quantum hardware.

Experimental QEC data span surface, color, and repetition codes, plus parity-check experiments, across superconducting, neutral-atom, and spin-qubit platforms.

Code family
Code distance
QEC platform

Showing 9 of 9 datasets

Distance 3

Google Surface D3

Surface code

Distance-3 surface-code memory experiments on Google’s Sycamore processor.

Google · SycamoreSuperconducting
Distance 5

Google Surface D5

Surface code

Distance-5 surface-code memory data for studying performance across code distances.

Google · SycamoreSuperconducting
Distance 3

Lukin D3

Surface code

Repeated surface-code error correction with detection of qubit loss.

Lukin groupNeutral atoms
Distance 5

Lukin D5

Surface code

Distance-5 surface-code experiments with repeated error correction and loss detection.

Lukin groupNeutral atoms
Distance 3

Google Color D3

Color code

Distance-3 color-code experiments on Google’s Willow processor.

Google · WillowSuperconducting
Distance 5

Google Color D5

Color code

Distance-5 color-code data for studying scaling across code distances.

Google · WillowSuperconducting
Distance 3

DiCarlo D3

Repetition code

Distance-3 repetition-code experiments from the DiCarlo research group.

DiCarlo groupSuperconducting
Distance 25

Google Repetition D25

Repetition code

Distance-25 bit-flip repetition-code experiments on Sycamore, using 49 qubits.

Google · SycamoreSuperconducting
No distance

Vandersypen

Parity checks

Spin-qubit parity-check and multi-qubit entanglement experiments.

Vandersypen groupSpin qubits

D3, D5, and D25 indicate code distance. The Vandersypen parity-check dataset does not have a code distance.

02 / The models

A broad range of decoding methods.

The benchmark covers 11 full-decoder methods across gradient-boosted trees, neural networks, tabular foundation models, and classical baselines.

Gradient-boosted trees

GBDT · 3 models

Strong, practical machine-learning methods that learn patterns from structured measurements.

CatBoostLightGBMXGBoost

Neural networks

NN-based · 4 models

Models that discover complex relationships, including patterns across sequences of measurements.

MLPTabMTabRLSTM

Tabular foundation models

TFM · 2 models

Modern models designed to transfer tabular-data learning strategies across problems.

TabPFNv2TabICLv2

Classical baselines

Reference methods · 2 models

Established decoding algorithms that provide an important reference point for AI approaches.

MWPMChromobius
03 / The toolkit

Research tools and model checkpoints.

Beyond the decoder benchmark, the zoo includes dedicated pre-decoding methods and reusable trained model checkpoints.

01 / SPEED

Pre-decoders

Lightweight pre-decoding techniques designed to reduce the computational work of downstream decoding tasks.

02 / COMPARE

Full decoders

Compare machine-learning and classical methods for predicting corrections from QEC measurement data.

03 / REUSE

Model checkpoints

Saved model checkpoints support model reuse and evaluation without requiring training from scratch.

Research with us

Better decoding starts with better comparisons.

Curious about QEC benchmarking, datasets, or collaboration? Connect with Infinidata Lab.

Get in touch
Code availability: The codebase will be available upon publication of the benchmark manuscript.