Quantum assistant: Specialization of multimodal vision-language models for quantum computing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Code assistants for quantum computing operate exclusively on text, disregarding visual representations --- circuit diagrams, Bloch spheres, and histograms --- prevalent in the domain's documentation. This work presents three contributions: (1) a synthetic generation pipeline that extracts content from public Qiskit documentation, transcribes images via a multimodal model, generates question-answer pairs, and validates code through automated unit tests; (2) the first public multimodal dataset for quantum computing, containing 8,366 samples (45% with images) across function completion, code generation, and question answering tasks; and (3) fine-tuning experiments comparing five LoRA variants on Qwen3-VL-8B, with rsLoRA (rank 32) achieving optimal performance. Evaluation on the Qiskit HumanEval benchmark demonstrated a gain of 11.26 percentage points over the base model (32.45% --> 43.71% Pass@1). On multimodal samples, the specialized model achieved 63.39% Pass@1, surpassing text-only samples (45.45%) by 17.94 points. The dataset, models, and code are available under the Apache 2.0 license.