[2508.20134] QAgent: An LLM-based Multi-Agent System for Autonomous OpenQASM programming

View a PDF of the paper titled QAgent: An LLM-based Multi-Agent System for Autonomous OpenQASM programming, by Zhenxiao Fu and 4 different authors

View PDF
HTML (experimental)

Summary:Programming quantum circuits on the OpenQASM degree is crucial for attaining hardware-aware optimization and dependable execution on noisy intermediate-scale quantum (NISQ) gadgets, but it stays difficult as a result of want for domain-specific planning, iterative code synthesis, and low-level calibration. On this paper, we current QAgent, the primary autonomous multi-agent framework for end-to-end OpenQASM code technology. QAgent integrates schema-aware process planning, example- and tool-driven code synthesis, and hardware-aware calibration inside a unified planning-synthesis-calibration workflow. The system leverages retrieval-augmented technology (RAG) to entry structured kernel information, examples, and backend constraints, and employs coordinated multi-agent reasoning with iterative execution suggestions to make sure correctness. We consider QAgent on 12 consultant quantum kernels and their compositions throughout 5 giant language fashions (LLMs). Outcomes present that QAgent improves Move@1 accuracy by 47-70% on single-kernel duties and achieves over 88% accuracy on multi-kernel workflows for giant fashions, considerably outperforming present baselines. Moreover, underneath lifelike {hardware} frequency drift, QAgent maintains near-unit execution constancy by means of automated calibration, whereas SDK-based LLM strategies endure vital degradation. These outcomes exhibit that integrating planning, synthesis, and calibration is crucial for dependable quantum program technology. The implementation of QAgent is open-sourced at this https URL