Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

QSpark: Towards Reliable Qiskit Code Generation

Created by
  • Haebom

Author

Kiana Kheiri, Aamna Aamir, Andriy Miranskyy, Chen Ding

Outline

This paper explores fine-tuning a large-scale language model (LLM) using reinforcement learning (RL) techniques to improve the error tolerance of quantum circuits. To address the problem that existing LLMs, such as Granite-20B-Code and StarCoder, often generate erroneous Qiskit code, we fine-tuned the Qwen2.5-Coder-32B model on a richly annotated synthetic dataset using two RL methods, GRPO and ORPO. Experimental results show that ORPO achieves a Pass@1 performance of 56.29% on the Qiskit HumanEval benchmark, approximately 10% better than Granite-8B-QK, while GRPO achieves 49%. While these models outperform general-purpose baseline models, they still have limitations when it comes to high-difficulty tasks.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that the quantum programming performance of LLM can be improved using reinforcement learning techniques.
◦
GRPO and ORPO algorithms outperform classical LLM on quantum programming tasks.
◦
In particular, GRPO demonstrates excellent performance in basic and intermediate level tasks.
•
Limitations:
◦
Still struggling to solve difficult quantum programming problems.
◦
Since the training was performed using a synthetic dataset, further verification is required for performance in a real quantum programming environment.
👍
Made with Slashpage