Fine-Tuning Large Language Models for Quantum Reasoning

The study presents two fine-tuning pipelines for large language models (LLMs) aimed at enhancing quantum reasoning capabilities. The first pipeline, Supervised Fine-Tuning (SFT), achieves near-perfect accuracy in predicting measurement probability distributions from quantum circuit simulations, outperforming the base model and GPT-OSS-120B. The second approach, SFT combined with Group Relative Policy Optimisation (GRPO), improves generalization to larger qubit systems, indicating that targeted fine-tuning on explicit reasoning traces is a viable method for developing LLMs capable of sophisticated quantum reasoning, which is crucial for applications in quantum computing.