Researchers introduce OmniQEC, an artificial intelligence system that searches for quantum error‑correcting codes that work well on real quantum processors. Quantum error correction (QEC) is necessary to protect fragile quantum information. But a code that looks good on paper can still fail when you build noisy circuits and run decoders. OmniQEC tries to close that gap by directly optimizing the circuit‑level logical error rate (LER), a practical measure of how often a protected logical qubit fails in a real implementation.

OmniQEC treats code discovery as an iterative ‘‘generate–evaluate–refine’’ process under the control of an orchestrator built from large language models (LLMs). The orchestrator translates a user objective (for example, a target code family, qubit budget, or connectivity) into a structured search. It proposes executable code constructions, calls specialized tools to check them, and stores results and reasoning so later proposals improve. The system is described as ‘‘self‑evolving’’ because it reuses past evidence to guide future searches.

To make the search computationally feasible, OmniQEC uses a dual‑loop workflow. A fast loop explores many candidates using cheap, code‑level proxies such as the number of logical qubits, code distance, and a code figure of merit Φcode = k d^2 / n (where k is logical qubits, d is distance, and n is physical qubits). A slower loop compiles selected candidates into noisy syndrome‑extraction circuits, runs decoder‑based evaluations, and measures the circuit‑level LER. The slow loop’s physically grounded feedback is fed back to the orchestrator to steer the search toward designs that perform well in practice.

The authors evaluated OmniQEC on families of quantum low‑density parity‑check (qLDPC) codes. They tested four qLDPC construction families, three LLM backends (Claude, GPT, and DeepSeek), and 14 total‑physical‑qubit budgets per backend. The discovered codes showed steadily better suppression of logical errors as the available physical qubits increased. In their simulated comparisons, some discovered codes outperformed a standard bivariate‑bicycle (BB) benchmark: they beat a [[72,12,6]] BB code when the deployment budget was 98 physical qubits, and they beat a [[144,12,12]] BB code when the budget was 240 physical qubits. The paper also reports that the commonly used code‑level figure of merit Φcode does not always match the circuit‑level LER ranking, which highlights the importance of circuit‑level evaluation.