[2603.22184] Revisiting Quantum Code Generation: Where Should Domain Knowledge Live?

Computer Science > Machine Learning arXiv:2603.22184 (cs) [Submitted on 23 Mar 2026] Title:Revisiting Quantum Code Generation: Where Should Domain Knowledge Live? Authors:Oscar Novo, Oscar Bastidas-Jossa, Alberto Calvo, Antonio Peris, Carlos Kuchkovsky View a PDF of the paper titled Revisiting Quantum Code Generation: Where Should Domain Knowledge Live?, by Oscar Novo and 4 other authors View PDF Abstract:Recent advances in large language models (LLMs) have enabled the automation of an increasing number of programming tasks, including code generation for scientific and engineering domains. In rapidly evolving software ecosystems such as quantum software development, where frameworks expose complex abstractions, a central question is how best to incorporate domain knowledge into LLM-based assistants while preserving maintainability as libraries evolve. In this work, we study specialization strategies for Qiskit code generation using the Qiskit-HumanEval benchmark. We compare a parameter-specialized fine-tuned baseline introduced in prior work against a range of recent general-purpose LLMs enhanced with retrieval-augmented generation (RAG) and agent-based inference with execution feedback. Our results show that modern general-purpose LLMs consistently outperform the parameter-specialized baseline. While the fine-tuned model achieves approximately 47% pass@1 on Qiskit-HumanEval, recent general-purpose models reach 60-65% under zero-shot and retrieval-augmented settings, and u...