Researchers at the Max Planck Institute for the Science of Light have developed a method that uses large language models to discover families of quantum error-correcting codes, cutting the physical qubit overhead per logical qubit by roughly an order of magnitude compared to the surface code. The advance tackles one of the hardest obstacles to building scalable, fault-tolerant quantum computers.
How structured concept evolution works
Zidu Liu and Florian Marquardt, working with Friedrich-Alexander University, designed structured concept evolution (SCE). The framework treats code construction as an evolutionary process, with a large language model acting as a mutation operator. It systematically alters algebraic specifications paired with executable programs that generate the parity-check matrices for quantum low-density parity-check (qLDPC) codes.
Lightweight models - specifically GPT-5.4-mini and GPT-5.4-nano - guided the search, showing that complex code discovery does not demand massive compute. SCE navigated an exponentially large design space and found code families that go beyond conventional bivariate-bicycle designs, including constructions based on non-abelian groups. The team implemented the approach in an open-source framework called OpenEvolve.
Performance under depolarizing noise
The discovered codes were benchmarked under code-capacity depolarizing noise with Belief Propagation plus Ordered Statistics Decoding (BP+OSD). They exhibit finite encoding rates and growing distance, both essential for practical scalability. The paper is available on ArXiv.
The evaluation, however, covers only one noise model. Different quantum hardware - superconducting qubits, trapped ions, photonic systems - each have distinct error profiles. A full assessment across multiple noise models will be necessary to confirm the codes' robustness on real devices. The researchers have released the code and framework openly to encourage that work.
Why this matters for Science and Research professionals
Quantum error correction remains a central bottleneck for reliable quantum computation. The SCE method offers a way to discover code families that outperform manually crafted designs, using language models small enough for many labs to run. Because the framework and resulting codes are open source, researchers can immediately test them against other noise models and hardware constraints. The approach may also transfer to other discrete design problems in quantum information science where search spaces overwhelm human intuition. This work is part of a growing use of AI for Science & Research to address long-standing technical challenges.
Your membership also unlocks:
