Max Planck Institute for the Science of Light: Researchers Use AI to Design Improved Quantum Error Correction Codes

A new method for building quantum error-correcting codes, key for unlocking the potential of quantum computation, has been devised by Zidu Liu and Florian Marquardt at the Max Planck Institute for the Science of Light, in collaboration with Friedrich-Alexander University and 1Max Planck Institute. Their work presents structured concept evolution (SCE), a framework using large language models to discover families of quantum low-density parity-check (qLDPC) codes. The method moves beyond traditional, challenging discrete design problems by evolving algebraic specifications alongside executable programs, resulting in the identification of competitive code families, including those utilising non-abelian groups, and demonstrating the potential of artificial intelligence in advancing quantum information science.

Discovery of efficient quantum error correction via language model guided code search

A reduction in the cost per logical qubit by roughly an order of magnitude has been achieved compared to the surface code, overcoming a key obstacle to scalable quantum computation. Previously, the surface code’s quadratic overhead in physical qubits limited progress as quantum platforms approached the thousand-qubit threshold; this new advancement enables constant-rate codes, offering a pathway to fault-tolerant computation with sharply reduced resource requirements. At the Institute for the Science of Light, in collaboration with Friedrich-Alexander University, scientists developed structured concept evolution (SCE), a framework combining large language models with algebraic mutation grammar to discover lifted-product code families, a type of quantum low-density parity-check (qLDPC) code. Quantum error correction is crucial because qubits, the fundamental units of quantum information, are inherently susceptible to noise and decoherence, leading to errors in computation. Without effective error correction, even small error rates would quickly overwhelm quantum algorithms, rendering them useless. qLDPC codes, with their sparse parity checks, offer a promising approach to mitigating these errors while minimising the overhead in physical qubits required to protect a single logical qubit.

Lightweight language models, specifically GPT-5.4-mini and GPT-5.4-nano, facilitated the discovery of diverse code families. These codes extend beyond standard designs like bivariate-bicycle codes, incorporating constructions based on non-abelian groups, broadening the possibilities for quantum error correction. The SCE framework successfully navigated an exponentially large search space, identifying codes benchmarked under depolarizing noise with Belief Propagation plus Ordered Statistics Decoding (BP+OSD). These codes also exhibit finite encoding rates and growing distance, important for scalability, and were evaluated with a code capacity relevant to current hardware limitations. The ‘distance’ of a code refers to its ability to correct errors; a larger distance implies a greater capacity to detect and correct more errors. Finite encoding rates are essential for practical implementation, ensuring that the overhead in physical qubits remains manageable. The use of lightweight language models is particularly significant, as it demonstrates that complex code discovery doesn’t necessarily require massive computational resources, opening the door to wider accessibility and experimentation.

Refining quantum error correction via language model guided concept evolution

Structured concept evolution, the core of this work, employs a large language model to refine the concepts that generate quantum error-correcting codes, rather than designing them directly. This technique treats code construction as an evolutionary process, with the language model acting as a mutation operator systematically altering algebraic specifications paired with executable programs. These programs then produce the complex parity-check matrices defining a quantum low-density parity-check (qLDPC) code, a system of checks and balances for quantum bits ensuring data integrity. Parity-check matrices are central to error correction; they define the relationships between qubits that allow the detection and correction of errors. A sparse parity-check matrix, as used in qLDPC codes, reduces the complexity of the decoding process and minimises the hardware requirements.

Traditional methods are limited in scope, but this approach allows exploration of a far wider range of code designs, circumventing the limitations of manually crafted or randomly generated solutions. The team focused on lifted-product codes, a specific type of qLDPC code built from two base protographs with dimensions mA x nA and mB x nB, utilising a finite group of order q to scale the block length linearly. Implementation was achieved within an open-source framework called OpenEvolve, enabling the process with lightweight language models. Lifted-product codes offer a structured approach to code construction, allowing for systematic exploration of different code parameters and properties. The use of finite groups provides a mathematical framework for defining the relationships between qubits and constructing the parity-check matrices. OpenEvolve facilitates reproducibility and collaboration, allowing other researchers to build upon and extend this work. The dimensions mA x nA and mB x nB define the structure of the base protographs, which are building blocks for the larger qLDPC code. The order of the finite group, q, determines the scaling of the block length, influencing the code’s performance and complexity.

Generating diverse quantum low-density parity-check (qLDPC) codes offers a potential pathway to more efficient quantum computation. However, the current evaluation focuses solely on ‘code-capacity depolarizing noise’ and utilises Belief Propagation plus Ordered Statistics Decoding, raising a key question regarding performance under alternative noise models prevalent in different quantum hardware platforms. Diverse technologies underpin quantum computer construction, each with unique error characteristics; evaluating codes against a single noise model provides a valuable, albeit incomplete, picture of their potential. Superconducting qubits, for example, are susceptible to different types of noise than trapped ions or photonic qubits. Therefore, a comprehensive evaluation of qLDPC codes should consider a range of noise models to ensure their robustness across different hardware architectures. Belief Propagation plus Ordered Statistics Decoding (BP+OSD) is a decoding algorithm used to infer the original quantum state from the noisy measurements. While effective for certain noise models, its performance may vary under different conditions. This automated approach refines the concepts underpinning code construction, allowing exploration of a wider range of possibilities, and scientists discovered code families based on complex, non-abelian mathematical groups, exceeding the limitations of more conventional designs. Non-abelian groups offer a richer mathematical structure that can potentially lead to more powerful and efficient error-correcting codes, although their implementation can be more challenging.

The researchers successfully developed a new framework, structured concept evolution, to discover diverse families of quantum low-density parity-check codes. This matters because these codes are essential for correcting errors in quantum computers, a crucial step towards building more reliable machines. The discovered codes, based on complex mathematical groups, demonstrate performance under code-capacity depolarizing noise when decoded with Belief Propagation plus Ordered Statistics Decoding. The authors made their approach and resulting codes openly available to encourage further research and collaboration in the field.