A breakthrough in quantum error correction has achieved up to a one hundred and twenty four-fold increase in decoding speed without sacrificing accuracy. CoDecoder; a small neural network that works alongside existing decoders to predict outcomes and intelligently route difficult cases for processing, improves the accuracy and speed of error correction for quantum computers. The new approach combines an existing decoder with a small neural network which predicts likely outcomes and intelligently directs more complex cases for processing.

Building a fault-tolerant quantum computer demands both accuracy *and* speed in correcting errors; the challenge lies in keeping pace with operations that occur within microseconds. Think of error correction like making copies of an important document: multiple versions allow you to identify and fix typos, the surface code achieves similar redundancy at the qubit level, while other schemes such as bivariate-bicycle (BB) codes use different approaches to achieve the same goal.

A key component is understanding what kind of error occurred, identifying whether data was flipped or shifted, represented mathematically by something called a *logical operator*. But can adding another layer actually accelerate decoding without sacrificing precision.

CoDecoder accelerates quantum error correction via hybrid neural network prediction

A staggering one hundred and twenty four-fold increase in quantum error correction decoding speed has been achieved at University of Colorado Boulder; this surpasses previous limitations where processing speeds struggled to keep pace with qubit operations measured in microseconds. Combining existing decoders with a small neural network, one predicting likely outcomes and intelligently directing complex cases for processing, was previously hindered by maintaining accuracy during acceleration.

This composition reduces failures by between 6.4% and 31.4% compared to correlated matching, alongside reductions of 22%, 50 % versus full-DEM Relay-BP, while also improving upon the convolutional decoder Cascade with fewer errors across all tested settings.

The team at the University of Colorado Boulder reduced decoding failures from 6.4% to 31.4% when comparing their CoDecoder system against correlated matching on surface codes spanning distances of 7 to 17; similarly, they saw improvements between 22% and 50% versus full-DEM Relay-BP across two bivariate-bicycle (BB) codes. This performance coupled with speed increases up to 124-fold over these existing methods, and faster processing times than re-implemented versions of the Cascade neural decoder for both code families tested.

In particular, employing a joint decision process where both decoders assess routed shots together maintained this acceleration while further reducing failures by 14%, 18 % on BB72 codes. On BB codes, an alternative output method failed 3.8 to 5.0 times more often than their combined scoring system under identical training conditions; demonstrating gains in accuracy through composition.

CoDecoder achieves substantial error reduction despite challenges in cross-architecture comparison

While CoDecoder demonstrably improves performance against established techniques like correlated matching and full-DEM Relay-BP, direct comparisons with other neural network approaches are trickier due to differing training datasets, introducing uncertainty when evaluating its relative merits versus systems such as Cascade. This highlights a key tension within quantum error correction: is it more fruitful to refine existing decoders using machine learning ‘frontends’, or pursue entirely new architectures built from the ground up. Still, acknowledging this difficulty, the team sharply cut errors by up to fifty percent while utilising their CoDecoder alongside existing methods; speed also improved thanks to graphics card processing. The team delivered tangible gains regardless. This composition reduces errors by as much as thirty-one percent on surface codes and fifty percent on bivariate-bicycle codes compared to current methods without sacrificing speed, demonstrating that augmenting existing technology can be more effective than wholesale replacement. Efficient processing on graphics cards and clever routing of information contribute significantly to these improvements; this hybrid approach offers a promising path toward scalable quantum computing.

CoDecoder successfully reduced decoding failures when paired with existing error correction techniques. By predicting logical outcomes using a neural network frontend and intelligently routing data, it improved accuracy by up to thirty-one percent for surface codes of distance 7, 17 and up to fifty percent for two tested bivariate-bicycle (BB) codes.

This composition also resulted in faster processing times, up to 124-fold speed increases were observed compared to current methods, while maintaining or improving upon the performance of standalone neural decoders like Cascade. The team demonstrated that combining machine learning with established algorithms can yield substantial improvements in quantum computation efficiency.

👉 More information
🗞 CoDecoder: Make Quantum Decoders Up to 100x Faster and Even More Accurate
✍️ Yicheng Guang, Yuliang Xiong, Xun Gao and Yueqi Chen (University of Colorado Boulder)
🧠 ArXiv: https://arxiv.org/abs/2610.28748

More like this