KQFuzz: Knowledge-Guided Fuzzing for Quantum Libraries via Large Language Models

Quantum computing libraries, such as Qiskit, PennyLane, and Cirq, are essential for developing quantum algorithms. However, these libraries are in a state of rapid, volatile evolution to keep pace with hardware advancements, which frequently introduces hidden software defects. Traditional fuzzing approaches, which rely on manually defined rules or templates, struggle to keep up with these changes and are often limited to low-level circuit gate logic. While Large Language Models (LLMs) offer a more flexible alternative, they suffer from low code validity when generating quantum programs due to a lack of domain-specific awareness and the frequent breaking changes in quantum APIs.

KQFuzz addresses these challenges by grounding LLM-based test generation in comprehensive codebase knowledge. The framework constructs an API corpus that captures four critical dimensions: static metadata, API associations (proximity, type coupling, and call relationships), semantic models, and evolution metrics. This corpus guides a probabilistic selection strategy that steers the LLM to generate valid, version-aligned seed programs.

To further improve the fuzzing process, KQFuzz employs a two-level mutation strategy: parameter-level mutations for numerical corner cases and gate-level structural mutations for entanglement-aware variations. A multi-dimensional fitness function—accounting for gate diversity, entanglement, and call depth—prioritizes high-quality seeds, ensuring the fuzzer explores complex and potentially bug-prone execution paths rather than just simple circuit structures.

By automating the discovery of bugs in high-level quantum orchestration logic, KQFuzz significantly reduces the manual effort required to maintain the reliability of quantum software stacks. The framework’s ability to outperform existing state-of-the-art fuzzers by nearly 20% in coverage demonstrates that integrating domain-specific knowledge is a viable path for testing rapidly evolving, complex software ecosystems. The successful identification and subsequent fixing of 13 bugs across major quantum libraries highlight the practical utility of this approach in ensuring the correctness of scientific and engineering conclusions drawn from quantum computations.